0adb096 Merge pull request #84 from bnema/next · vev

A cmd/vev-perf-harness/evidence.go +324 -0

@@ -0,0 +1,324 @@
	1	+	package main
	2	+
	3	+	import (
	4	+	"bufio"
	5	+	"encoding/json"
	6	+	"errors"
	7	+	"fmt"
	8	+	"os"
	9	+	"sort"
	10	+	"strings"
	11	+	)
	12	+
	13	+	func measuredEventSamples(marks []harnessMark, interval measuredInterval) ([]eventSample, error) {
	14	+	if interval.End < interval.Start {
	15	+	return nil, errors.New("negative measured interval")
	16	+	}
	17	+	starts := map[uint64]int64{}
	18	+	events := make([]eventSample, 0, len(marks)/2)
	19	+	for _, m := range marks {
	20	+	if !m.Valid {
	21	+	return nil, errors.New("invalid harness boundary")
	22	+	}
	23	+	switch m.Kind {
	24	+	case "input_injected":
	25	+	if _, exists := starts[m.Sequence]; exists {
	26	+	return nil, errors.New("duplicate input boundary")
	27	+	}
	28	+	starts[m.Sequence] = m.Tick
	29	+	case "terminal_flushed":
	30	+	start, ok := starts[m.Sequence]
	31	+	if !ok {
	32	+	return nil, errors.New("terminal flush without input pair")
	33	+	}
	34	+	if m.Tick < start {
	35	+	return nil, errors.New("negative harness duration")
	36	+	}
	37	+	delete(starts, m.Sequence)
	38	+	if start >= interval.Start && m.Tick < interval.End {
	39	+	events = append(events, eventSample{Sequence: m.Sequence, Injected: start, Flushed: m.Tick, Latency: m.Tick - start})
	40	+	}
	41	+	default:
	42	+	return nil, errors.New("unknown harness mark")
	43	+	}
	44	+	}
	45	+	if len(starts) != 0 {
	46	+	return nil, errors.New("missing terminal flush pair")
	47	+	}
	48	+	sort.Slice(events, func(i, j int) bool {
	49	+	if events[i].Injected == events[j].Injected {
	50	+	return events[i].Sequence < events[j].Sequence
	51	+	}
	52	+	return events[i].Injected < events[j].Injected
	53	+	})
	54	+	return events, nil
	55	+	}
	56	+
	57	+	func requireMinimumEventSamples(events []eventSample) error {
	58	+	if len(events) < minimumInIntervalEventSamples {
	59	+	return fmt.Errorf("insufficient in-interval event samples: got %d, need at least %d", len(events), minimumInIntervalEventSamples)
	60	+	}
	61	+	return nil
	62	+	}
	63	+
	64	+	func eventLatencies(events []eventSample) []int64 {
	65	+	out := make([]int64, len(events))
	66	+	for i, event := range events {
	67	+	out[i] = event.Latency
	68	+	}
	69	+	return out
	70	+	}
	71	+
	72	+	func eventCadenceSamples(events []eventSample) ([]int64, int64) {
	73	+	if len(events) < 2 {
	74	+	return nil, 0
	75	+	}
	76	+	gaps := make([]int64, 0, len(events)-1)
	77	+	var maxGap int64
	78	+	for i := 1; i < len(events); i++ {
	79	+	gap := events[i].Injected - events[i-1].Injected
	80	+	gaps = append(gaps, gap)
	81	+	if gap > maxGap {
	82	+	maxGap = gap
	83	+	}
	84	+	}
	85	+	return gaps, maxGap
	86	+	}
	87	+
	88	+	func eventCadence(events []eventSample) (distribution, int64) {
	89	+	gaps, maxGap := eventCadenceSamples(events)
	90	+	if len(gaps) == 0 {
	91	+	return distribution{}, maxGap
	92	+	}
	93	+	return percentiles(gaps), maxGap
	94	+	}
	95	+
	96	+	func pairedSamples(marks []harnessMark) ([]int64, error) {
	97	+	starts := map[uint64]int64{}
	98	+	var out []int64
	99	+	for _, m := range marks {
	100	+	if !m.Valid {
	101	+	return nil, errors.New("invalid harness boundary")
	102	+	}
	103	+	switch m.Kind {
	104	+	case "input_injected":
	105	+	if _, ok := starts[m.Sequence]; ok {
	106	+	return nil, errors.New("duplicate input boundary")
	107	+	}
	108	+	starts[m.Sequence] = m.Tick
	109	+	case "terminal_flushed":
	110	+	start, ok := starts[m.Sequence]
	111	+	if !ok {
	112	+	return nil, errors.New("terminal flush without input pair")
	113	+	}
	114	+	if m.Tick < start {
	115	+	return nil, errors.New("negative harness duration")
	116	+	}
	117	+	out = append(out, m.Tick-start)
	118	+	delete(starts, m.Sequence)
	119	+	default:
	120	+	return nil, errors.New("unknown harness mark")
	121	+	}
	122	+	}
	123	+	if len(starts) != 0 {
	124	+	return nil, errors.New("missing terminal flush pair")
	125	+	}
	126	+	if len(out) == 0 {
	127	+	return nil, errors.New("insufficient harness samples")
	128	+	}
	129	+	return out, nil
	130	+	}
	131	+
	132	+	// traceRecord is intentionally local to the command. It mirrors JSONL fields
	133	+	// needed for correlation but never exposes a clock to any production component.
	134	+	type traceRecord struct {
	135	+	ProcessID string `json:"process_id"`
	136	+	Component string `json:"component"`
	137	+	Scenario string `json:"scenario"`
	138	+	Run uint64 `json:"run"`
	139	+	Sequence uint64 `json:"sequence"`
	140	+	RequestID uint64 `json:"request_id"`
	141	+	Epoch uint64 `json:"epoch"`
	142	+	Kind string `json:"kind"`
	143	+	Tick int64 `json:"tick"`
	144	+	Valid bool `json:"valid"`
	145	+	}
	146	+
	147	+	type spanPair struct{ start, end, name string }
	148	+
	149	+	// mergeProcessTraces permits only records from one manifest process in a span.
	150	+	// In particular, the correlation key includes process_id before ticks are read.
	151	+	func mergeProcessTraces(mappings []processMapping) ([]span, error) {
	152	+	known := map[string]processMapping{}
	153	+	for _, m := range mappings {
	154	+	if m.ProcessID == "" \|\| m.ClockDomain == "" \|\| m.TracePath == "" \|\| known[m.ProcessID].ProcessID != "" {
	155	+	return nil, errors.New("duplicate or incomplete manifest process mapping")
	156	+	}
	157	+	known[m.ProcessID] = m
	158	+	}
	159	+	out := []span{}
	160	+	for _, m := range mappings {
	161	+	f, err := os.Open(m.TracePath)
	162	+	if err != nil {
	163	+	return nil, fmt.Errorf("open manifest trace for %s role: %w", m.Role, err)
	164	+	}
	165	+	var records []traceRecord
	166	+	scan := bufio.NewScanner(f)
	167	+	for scan.Scan() {
	168	+	var r traceRecord
	169	+	if err := json.Unmarshal(scan.Bytes(), &r); err != nil {
	170	+	return nil, closeFile(f, err)
	171	+	}
	172	+	if r.ProcessID != m.ProcessID \|\| r.Scenario != m.Scenario \|\| r.Run != uint64(m.Run) {
	173	+	return nil, closeFile(f, errors.New("trace record identity does not match manifest"))
	174	+	}
	175	+	if r.Scenario == "" \|\| r.Run == 0 \|\| r.Component == "" \|\| r.Sequence == 0 \|\| r.RequestID == 0 \|\| r.Epoch == 0 {
	176	+	return nil, closeFile(f, errors.New("trace record has invalid correlation fields"))
	177	+	}
	178	+	if _, ok := known[r.ProcessID]; !ok {
	179	+	return nil, closeFile(f, errors.New("unknown trace process_id"))
	180	+	}
	181	+	records = append(records, r)
	182	+	}
	183	+	if err := scan.Err(); err != nil {
	184	+	return nil, closeFile(f, err)
	185	+	}
	186	+	if err := f.Close(); err != nil {
	187	+	return nil, err
	188	+	}
	189	+	// A process shares one observer across concurrent components. Correlation
	190	+	// IDs are allocated before a goroutine reaches its first mark, so a later
	191	+	// ID can be serialized before an earlier one. Their first appearance is
	192	+	// consequently not an ordering contract. Only marks in the same component
	193	+	// and exact correlation domain may pair, and each pair's ticks must retain
	194	+	// its own start-before-end ordering.
	195	+	starts := map[string]traceRecord{}
	196	+	for _, r := range records {
	197	+	for _, pair := range spanPairs {
	198	+	key := fmt.Sprintf("%s/%s/%s/%d/%d/%d/%d", r.ProcessID, r.Component, m.Scenario, m.Run, r.Sequence, r.RequestID, r.Epoch)
	199	+	if r.Kind == pair.start {
	200	+	k := pair.name + "/" + key
	201	+	if _, exists := starts[k]; exists {
	202	+	return nil, errors.New("duplicate process-local span start")
	203	+	}
	204	+	starts[k] = r
	205	+	}
	206	+	if r.Kind == pair.end {
	207	+	k := pair.name + "/" + key
	208	+	start, ok := starts[k]
	209	+	if !ok {
	210	+	return nil, errors.New("span end without same-process start")
	211	+	}
	212	+	if r.Tick < start.Tick {
	213	+	return nil, errors.New("negative same-process span")
	214	+	}
	215	+	// Validity affects only measurement eligibility, never structural
	216	+	// pairing: every start/end must still match in its exact process,
	217	+	// component, and correlation domain. A failed start, failed end, or
	218	+	// both is a paired diagnostic fact but contributes no latency sample.
	219	+	if start.Valid && r.Valid {
	220	+	out = append(out, span{Component: r.Component, Name: pair.name, Samples: []int64{r.Tick - start.Tick}})
	221	+	}
	222	+	delete(starts, k)
	223	+	}
	224	+	}
	225	+	}
	226	+	if len(starts) != 0 {
	227	+	keys := make([]string, 0, len(starts))
	228	+	for key := range starts {
	229	+	keys = append(keys, key)
	230	+	}
	231	+	sort.Strings(keys)
	232	+	return nil, fmt.Errorf("missing process-local span pair for %s: %s", m.ProcessID, strings.Join(keys, ", "))
	233	+	}
	234	+	}
	235	+	return out, nil
	236	+	}
	237	+
	238	+	func summarizeSpans(all []span) []spanSummary {
	239	+	grouped := map[string]*span{}
	240	+	for _, s := range all {
	241	+	k := s.Component + "\x00" + s.Name
	242	+	if grouped[k] == nil {
	243	+	grouped[k] = &span{Component: s.Component, Name: s.Name}
	244	+	}
	245	+	grouped[k].Samples = append(grouped[k].Samples, s.Samples...)
	246	+	}
	247	+	keys := make([]string, 0, len(grouped))
	248	+	for k := range grouped {
	249	+	keys = append(keys, k)
	250	+	}
	251	+	sort.Strings(keys)
	252	+	out := make([]spanSummary, 0, len(keys))
	253	+	for _, k := range keys {
	254	+	s := grouped[k]
	255	+	if len(s.Samples) < minimumRepetitions {
	256	+	out = append(out, spanSummary{Component: s.Component, Name: s.Name, Distribution: distribution{Count: len(s.Samples)}})
	257	+	continue
	258	+	}
	259	+	out = append(out, spanSummary{Component: s.Component, Name: s.Name, Distribution: percentiles(s.Samples)})
	260	+	}
	261	+	return out
	262	+	}
	263	+
	264	+	func percentiles(samples []int64) distribution {
	265	+	v := append([]int64(nil), samples...)
	266	+	sort.Slice(v, func(i, j int) bool { return v[i] < v[j] })
	267	+	at := func(p int) int64 { return v[(len(v)-1)*p/100] }
	268	+	return distribution{len(v), at(50), at(95), at(99), v[len(v)-1]}
	269	+	}
	270	+
	271	+	func closeFile(f *os.File, err error) error {
	272	+	return errors.Join(err, f.Close())
	273	+	}
	274	+
	275	+	func writeJSON(path string, v any) error {
	276	+	f, err := os.OpenFile(path, os.O_WRONLY\|os.O_CREATE\|os.O_EXCL, 0o600)
	277	+	if err != nil {
	278	+	return err
	279	+	}
	280	+	if err := json.NewEncoder(f).Encode(v); err != nil {
	281	+	return closeFile(f, err)
	282	+	}
	283	+	return f.Close()
	284	+	}
	285	+
	286	+	func mustJSON(v any) string {
	287	+	b, err := json.Marshal(v)
	288	+	if err != nil {
	289	+	panic(fmt.Sprintf("marshal JSON: %v", err))
	290	+	}
	291	+	return string(b)
	292	+	}
	293	+
	294	+	func safeName(s string) string {
	295	+	return strings.Map(func(r rune) rune {
	296	+	if r >= 'a' && r <= 'z' \|\| r >= 'A' && r <= 'Z' \|\| r >= '0' && r <= '9' \|\| r == '-' \|\| r == '_' {
	297	+	return r
	298	+	}
	299	+	return '-'
	300	+	}, s)
	301	+	}
	302	+
	303	+	func readJSONL(path string) ([]map[string]json.RawMessage, error) {
	304	+	f, err := os.Open(path)
	305	+	if err != nil {
	306	+	return nil, err
	307	+	}
	308	+	var r []map[string]json.RawMessage
	309	+	scan := bufio.NewScanner(f)
	310	+	for scan.Scan() {
	311	+	var v map[string]json.RawMessage
	312	+	if err := json.Unmarshal(scan.Bytes(), &v); err != nil {
	313	+	return nil, closeFile(f, err)
	314	+	}
	315	+	r = append(r, v)
	316	+	}
	317	+	if err := scan.Err(); err != nil {
	318	+	return nil, closeFile(f, err)
	319	+	}
	320	+	if err := f.Close(); err != nil {
	321	+	return nil, err
	322	+	}
	323	+	return r, nil
	324	+	}

A cmd/vev-perf-harness/evidence_test.go +387 -0

@@ -0,0 +1,387 @@
	1	+	package main
	2	+
	3	+	import (
	4	+	"bytes"
	5	+	"encoding/json"
	6	+	"errors"
	7	+	"os"
	8	+	"path/filepath"
	9	+	"strings"
	10	+	"testing"
	11	+	"time"
	12	+	)
	13	+
	14	+	func TestCLITransportSeamOwnsExclusivePeerTraceAndCleanup(t *testing.T) {
	15	+	fixtures := []struct {
	16	+	name string
	17	+	tr transport
	18	+	rtt time.Duration
	19	+	loss int
	20	+	}{
	21	+	{"baseline", transport{ID: "udp_baseline", Kind: "udp"}, 0, 0},
	22	+	{"25ms", transport{ID: "udp_25ms", Kind: "udp", RTTMS: 25}, 25 * time.Millisecond, 0},
	23	+	{"100ms", transport{ID: "udp_100ms", Kind: "udp", RTTMS: 100}, 100 * time.Millisecond, 0},
	24	+	{"loss0", transport{ID: "udp_loss_0pct", Kind: "udp"}, 0, 0},
	25	+	{"loss1", transport{ID: "udp_loss_1pct", Kind: "udp", LossPercent: 1}, 0, 1},
	26	+	}
	27	+	for _, tc := range fixtures {
	28	+	t.Run(tc.name, func(t *testing.T) {
	29	+	dir := t.TempDir()
	30	+	if err := os.Chmod(dir, 0o700); err != nil {
	31	+	t.Fatal(err)
	32	+	}
	33	+	var configs []udpNetemConfig
	34	+	netem := &fakeUDPNetem{port: 45678}
	35	+	l := &cliLauncher{bin: "/bin/true", netemFactory: func(c udpNetemConfig) (udpNetem, error) {
	36	+	configs = append(configs, c)
	37	+	return netem, nil
	38	+	}}
	39	+	m := processMapping{ProcessID: "udp-peer", TracePath: filepath.Join(dir, "udp.jsonl"), Role: "udp_peer"}
	40	+	if err := os.WriteFile(m.TracePath, nil, 0o600); err != nil {
	41	+	t.Fatal(err)
	42	+	}
	43	+	p, err := l.preparePeer(m, roleCommand{Args: []string{"_udp-proxy", "work"}, Transport: tc.tr})
	44	+	if err != nil {
	45	+	t.Fatal(err)
	46	+	}
	47	+	if len(configs) != 1 \|\| configs[0].RTT != tc.rtt \|\| configs[0].LossPercent != tc.loss \|\| configs[0].TargetPath != filepath.Join(dir, "udp-peer.target") {
	48	+	t.Fatalf("fixture did not reach netem seam: %+v", configs)
	49	+	}
	50	+	shim, err := os.ReadFile(filepath.Join(dir, "ssh"))
	51	+	if err != nil {
	52	+	t.Fatal(err)
	53	+	}
	54	+	text := string(shim)
	55	+	for _, want := range []string{"_udp-proxy", m.TracePath, m.ProcessID, "udp-peer.target", "VEV-UDP %s %s\\n' 45678", "udp-peer.pid"} {
	56	+	if !strings.Contains(text, want) {
	57	+	t.Errorf("seam does not retain %q:\n%s", want, text)
	58	+	}
	59	+	}
	60	+	if strings.Contains(text, "VEV_PERF_UDP_") {
	61	+	t.Fatalf("ignored vev UDP environment was used instead of netem: %s", text)
	62	+	}
	63	+	// A nonexistent pid is already-cleaned-up; Close must still close the
	64	+	// harness-owned emulator.
	65	+	if err := p.Close(); err != nil {
	66	+	t.Fatal(err)
	67	+	}
	68	+	if !netem.closed {
	69	+	t.Fatal("harness netem was not cleaned up")
	70	+	}
	71	+	})
	72	+	}
	73	+	}
	74	+
	75	+	func TestHarnessRejectsInvalidBoundaryPairs(t *testing.T) {
	76	+	for _, marks := range [][]harnessMark{
	77	+	{{Sequence: 1, Kind: "terminal_flushed", Tick: 1, Valid: true}},
	78	+	{{Sequence: 1, Kind: "input_injected", Tick: 2, Valid: true}, {Sequence: 1, Kind: "terminal_flushed", Tick: 1, Valid: true}},
	79	+	{{Sequence: 1, Kind: "input_injected", Tick: 1, Valid: true}},
	80	+	{{Sequence: 1, Kind: "input_injected", Tick: 1, Valid: true}, {Sequence: 1, Kind: "input_injected", Tick: 2, Valid: true}},
	81	+	} {
	82	+	if _, err := pairedSamples(marks); err == nil {
	83	+	t.Fatalf("invalid marks accepted: %+v", marks)
	84	+	}
	85	+	}
	86	+	got, err := pairedSamples([]harnessMark{{Sequence: 9, Kind: "input_injected", Tick: 1, Valid: true}, {Sequence: 9, Kind: "terminal_flushed", Tick: 5, Valid: true}})
	87	+	if err != nil \|\| len(got) != 1 \|\| got[0] != 4 {
	88	+	t.Fatalf("got %v %v", got, err)
	89	+	}
	90	+	}
	91	+
	92	+	func TestHarnessCollectsOnlyCompleteMeasuredIntervalEvents(t *testing.T) {
	93	+	interval := measuredInterval{Start: 100, End: 200}
	94	+	marks := []harnessMark{
	95	+	{Sequence: 1, Kind: "input_injected", Tick: 10, Valid: true},
	96	+	{Sequence: 1, Kind: "terminal_flushed", Tick: 20, Valid: true}, // warmup
	97	+	{Sequence: 2, Kind: "input_injected", Tick: 100, Valid: true},
	98	+	{Sequence: 2, Kind: "terminal_flushed", Tick: 120, Valid: true},
	99	+	{Sequence: 3, Kind: "input_injected", Tick: 150, Valid: true},
	100	+	{Sequence: 3, Kind: "terminal_flushed", Tick: 201, Valid: true}, // exits interval
	101	+	{Sequence: 4, Kind: "input_injected", Tick: 99, Valid: true},
	102	+	{Sequence: 4, Kind: "terminal_flushed", Tick: 110, Valid: true}, // enters interval
	103	+	{Sequence: 5, Kind: "input_injected", Tick: 180, Valid: true},
	104	+	{Sequence: 5, Kind: "terminal_flushed", Tick: 200, Valid: true}, // end is excluded
	105	+	}
	106	+	events, err := measuredEventSamples(marks, interval)
	107	+	if err != nil {
	108	+	t.Fatal(err)
	109	+	}
	110	+	if len(events) != 1 \|\| events[0].Sequence != 2 \|\| events[0].Latency != 20 {
	111	+	t.Fatalf("measured events=%+v, want only complete sequence 2", events)
	112	+	}
	113	+	}
	114	+
	115	+	func TestHarnessRejectsInsufficientMeasuredEventSamples(t *testing.T) {
	116	+	marks := make([]harnessMark, 0, 2*(minimumInIntervalEventSamples-1))
	117	+	for i := 0; i < minimumInIntervalEventSamples-1; i++ {
	118	+	marks = append(marks,
	119	+	harnessMark{Sequence: uint64(i + 1), Kind: "input_injected", Tick: int64(i * 100), Valid: true},
	120	+	harnessMark{Sequence: uint64(i + 1), Kind: "terminal_flushed", Tick: int64(i*100 + 10), Valid: true},
	121	+	)
	122	+	}
	123	+	events, err := measuredEventSamples(marks, measuredInterval{Start: 0, End: int64(minimumInIntervalEventSamples * 100)})
	124	+	if err != nil {
	125	+	t.Fatal(err)
	126	+	}
	127	+	if err := requireMinimumEventSamples(events); err == nil {
	128	+	t.Fatal("insufficient measured event samples accepted")
	129	+	}
	130	+	}
	131	+
	132	+	func TestHarnessRunRejectsStraddlingEventsAsInsufficient(t *testing.T) {
	133	+	dir := t.TempDir()
	134	+	raw, err := os.OpenFile(filepath.Join(dir, "raw.jsonl"), os.O_CREATE\|os.O_EXCL\|os.O_WRONLY, 0o600)
	135	+	if err != nil {
	136	+	t.Fatal(err)
	137	+	}
	138	+	t.Cleanup(func() {
	139	+	if err := raw.Close(); err != nil && !errors.Is(err, os.ErrClosed) {
	140	+	t.Error(err)
	141	+	}
	142	+	})
	143	+	clock := &fakeClock{}
	144	+	h := defaultHarness()
	145	+	h.clock, h.launcher = clock, insufficientEventLauncher{clock: clock}
	146	+	_, err = h.runOne(options{out: dir, warmup: time.Second, duration: minimumDuration}, manifest{}, scenario{ID: "late", Roles: []string{"daemon", "client"}}, 1, raw)
	147	+	if err == nil \|\| !strings.Contains(err.Error(), "insufficient in-interval event samples") {
	148	+	t.Fatalf("straddling run error=%v, want insufficient in-interval sample rejection", err)
	149	+	}
	150	+	if err := raw.Close(); err != nil {
	151	+	t.Fatal(err)
	152	+	}
	153	+	marks := readHarnessMarks(t, filepath.Join(dir, "raw.jsonl"))
	154	+	if got, want := len(marks), 4; got != want {
	155	+	t.Fatalf("raw marks=%d, want warmup and straddling pair=%d: %+v", got, want, marks)
	156	+	}
	157	+	if straddling := marks[len(marks)-1]; straddling.Sequence != 2 \|\| straddling.Kind != "terminal_flushed" \|\| !straddling.Valid {
	158	+	t.Fatalf("straddling boundary=%+v", straddling)
	159	+	}
	160	+	}
	161	+
	162	+	func TestHarnessCalculatesEventCadenceAndMaxGapFromRawPairs(t *testing.T) {
	163	+	events := []eventSample{
	164	+	{Sequence: 1, Injected: 100, Flushed: 110, Latency: 10},
	165	+	{Sequence: 2, Injected: 130, Flushed: 140, Latency: 10},
	166	+	{Sequence: 3, Injected: 180, Flushed: 190, Latency: 10},
	167	+	}
	168	+	cadence, maxGap := eventCadence(events)
	169	+	if cadence.Count != 2 \|\| cadence.P50 != 30 \|\| cadence.P95 != 30 \|\| cadence.P99 != 30 \|\| maxGap != 50 {
	170	+	t.Fatalf("cadence=%+v maxGap=%d", cadence, maxGap)
	171	+	}
	172	+	}
	173	+
	174	+	func TestHarnessIncludesAllRequiredProcessLocalSpans(t *testing.T) {
	175	+	got := make([]string, 0, len(spanPairs))
	176	+	for _, pair := range spanPairs {
	177	+	got = append(got, pair.name)
	178	+	}
	179	+	want := []string{"capture_duration", "compose_duration", "diff_duration", "queue_wait", "ack_blocked_interval", "emit_duration", "adapter_send_duration", "adapter_receive_duration"}
	180	+	if !equalStrings(got, want) {
	181	+	t.Fatalf("span pairs = %v, want %v", got, want)
	182	+	}
	183	+	}
	184	+
	185	+	func TestHarnessRawRecordsAreDeterministicJSONL(t *testing.T) {
	186	+	var b bytes.Buffer
	187	+	for _, m := range []harnessMark{{Scenario: "s", Run: 1, Sequence: 1, Kind: "input_injected", Tick: 1, Valid: true}, {Scenario: "s", Run: 1, Sequence: 1, Kind: "terminal_flushed", Tick: 2, Valid: true}} {
	188	+	b.WriteString(mustJSON(m))
	189	+	b.WriteByte('\n')
	190	+	}
	191	+	path := filepath.Join(t.TempDir(), "raw.jsonl")
	192	+	if err := os.WriteFile(path, b.Bytes(), 0o600); err != nil {
	193	+	t.Fatal(err)
	194	+	}
	195	+	if got, err := readJSONL(path); err != nil \|\| len(got) != 2 {
	196	+	t.Fatalf("records=%d err=%v", len(got), err)
	197	+	}
	198	+	}
	199	+
	200	+	func TestHarnessClosesRolesBeforeMergingReceiveSpans(t *testing.T) {
	201	+	dir := t.TempDir()
	202	+	raw, err := os.OpenFile(filepath.Join(dir, "raw.jsonl"), os.O_CREATE\|os.O_EXCL\|os.O_WRONLY, 0o600)
	203	+	if err != nil {
	204	+	t.Fatal(err)
	205	+	}
	206	+	t.Cleanup(func() {
	207	+	if err := raw.Close(); err != nil && !errors.Is(err, os.ErrClosed) {
	208	+	t.Error(err)
	209	+	}
	210	+	})
	211	+	launcher := &closeTraceLauncher{}
	212	+	h := defaultHarness()
	213	+	h.clock, h.launcher = &fakeClock{}, launcher
	214	+	result, err := h.runOne(options{out: dir, warmup: time.Second, duration: minimumDuration, repetitions: minimumRepetitions}, manifest{}, scenario{ID: "receive-cleanup", Roles: []string{"daemon", "client"}}, 1, raw)
	215	+	if err != nil {
	216	+	t.Fatal(err)
	217	+	}
	218	+	if len(result.Spans) != 2 {
	219	+	t.Fatalf("post-cleanup receive spans=%+v", result.Spans)
	220	+	}
	221	+	for _, s := range result.Spans {
	222	+	if s.Name != "adapter_receive_duration" \|\| len(s.Samples) != 1 \|\| s.Samples[0] != 10 {
	223	+	t.Fatalf("unmatched or wrong receive span: %+v", s)
	224	+	}
	225	+	}
	226	+	for i, p := range launcher.process {
	227	+	if !p.closed {
	228	+	t.Fatalf("role %s was not closed before trace merge", launcher.mappings[i].Role)
	229	+	}
	230	+	}
	231	+	}
	232	+
	233	+	func TestHarnessWritesRequiredEvidenceWithSufficientSpans(t *testing.T) {
	234	+	base, err := readManifest(filepath.Join("..", "..", "testdata", "perf", "manifest.json"))
	235	+	if err != nil {
	236	+	t.Fatal(err)
	237	+	}
	238	+	for i := range base.Scenarios {
	239	+	if i > 0 {
	240	+	base.Scenarios[i].InapplicableReason = "fixture limits this test to one public topology"
	241	+	base.Scenarios[i].Roles = nil
	242	+	}
	243	+	}
	244	+	dir := t.TempDir()
	245	+	manifestPath := filepath.Join(dir, "manifest.json")
	246	+	if err := os.WriteFile(manifestPath, []byte(mustJSON(base)), 0o600); err != nil {
	247	+	t.Fatal(err)
	248	+	}
	249	+	l := &traceLauncher{}
	250	+	h := defaultHarness()
	251	+	h.clock = &fakeClock{}
	252	+	h.launcher = l
	253	+	h.gitSHA = func() string { return "test-sha" }
	254	+	if err := run([]string{"--vev-bin", "ignored", "--manifest", manifestPath, "--out", filepath.Join(dir, "out"), "--warmup", "1s", "--duration", "30s", "--repetitions", "10"}, h); err != nil {
	255	+	t.Fatal(err)
	256	+	}
	257	+	for _, name := range []string{"raw-harness.jsonl", "runs.json", "summary.json"} {
	258	+	if _, err := os.Stat(filepath.Join(dir, "out", name)); err != nil {
	259	+	t.Errorf("%s: %v", name, err)
	260	+	}
	261	+	}
	262	+	var got summary
	263	+	b, err := os.ReadFile(filepath.Join(dir, "out", "summary.json"))
	264	+	if err != nil {
	265	+	t.Fatal(err)
	266	+	}
	267	+	if err := json.Unmarshal(b, &got); err != nil {
	268	+	t.Fatal(err)
	269	+	}
	270	+	if got.EndToEnd.Count < 10*minimumInIntervalEventSamples \|\| got.Repetitions != 10 \|\| got.RunP50Dispersion.Count != 10 \|\| got.Cadence.Count == 0 \|\| got.MaxGap == 0 \|\| len(got.Spans) == 0 {
	271	+	t.Fatalf("summary=%+v", got)
	272	+	}
	273	+	for _, s := range got.Spans {
	274	+	if s.Distribution.Count < 10 {
	275	+	t.Fatalf("insufficient span summary: %+v", s)
	276	+	}
	277	+	}
	278	+	}
	279	+
	280	+	func TestHarnessAcceptsConcurrentProductionTraceInterleaving(t *testing.T) {
	281	+	path := filepath.Join(t.TempDir(), "local-client.jsonl")
	282	+	if err := os.WriteFile(path, []byte(boundedLocalConcurrentTrace), 0o600); err != nil {
	283	+	t.Fatal(err)
	284	+	}
	285	+	spans, err := mergeProcessTraces([]processMapping{{
	286	+	ProcessID: "1x4-idle-local-r002-client-02", ClockDomain: "1x4-idle-local-r002-client-02", TracePath: path,
	287	+	Scenario: "1x4-idle-local", Run: 2,
	288	+	}})
	289	+	if err != nil {
	290	+	t.Fatalf("valid concurrent production trace rejected: %v", err)
	291	+	}
	292	+	if len(spans) != 2 \|\| spans[0].Component != "ipc" \|\| spans[0].Name != "adapter_send_duration" \|\| spans[0].Samples[0] != 21900 \|\| spans[1].Component != "ipc" \|\| spans[1].Name != "adapter_receive_duration" \|\| spans[1].Samples[0] != 178550 {
	293	+	t.Fatalf("spans=%+v", spans)
	294	+	}
	295	+	}
	296	+
	297	+	func TestHarnessRejectsCrossProcessAndBadTraceSpans(t *testing.T) {
	298	+	write := func(t *testing.T, records ...traceRecord) processMapping {
	299	+	t.Helper()
	300	+	path := filepath.Join(t.TempDir(), "trace.jsonl")
	301	+	f, err := os.Create(path)
	302	+	if err != nil {
	303	+	t.Fatal(err)
	304	+	}
	305	+	for _, r := range records {
	306	+	b, err := json.Marshal(r)
	307	+	if err != nil {
	308	+	t.Fatal(err)
	309	+	}
	310	+	if _, err := f.Write(append(b, '\n')); err != nil {
	311	+	t.Fatal(err)
	312	+	}
	313	+	}
	314	+	if err := f.Close(); err != nil {
	315	+	t.Fatal(err)
	316	+	}
	317	+	return processMapping{ProcessID: "one", ClockDomain: "one", TracePath: path, Scenario: "s", Run: 1}
	318	+	}
	319	+	base := func(kind string, tick int64) traceRecord {
	320	+	return traceRecord{ProcessID: "one", Component: "daemon", Scenario: "s", Run: 1, Sequence: 1, RequestID: 1, Epoch: 1, Kind: kind, Tick: tick, Valid: true}
	321	+	}
	322	+	for _, records := range [][]traceRecord{
	323	+	{base("diff_end", 1)},
	324	+	{base("diff_start", 2), base("diff_end", 1)},
	325	+	{base("diff_start", 1)},
	326	+	{base("diff_start", 1), base("diff_start", 2)},
	327	+	{base("diff_start", 1), func() traceRecord { r := base("diff_end", 2); r.Component = "ipc"; return r }()},
	328	+	{func() traceRecord { r := base("diff_start", 1); r.ProcessID = "other"; return r }()},
	329	+	{func() traceRecord { r := base("diff_start", 1); r.Scenario = "other-scenario"; return r }()},
	330	+	{func() traceRecord { r := base("diff_start", 1); r.Run = 2; return r }()},
	331	+	} {
	332	+	if _, err := mergeProcessTraces([]processMapping{write(t, records...)}); err == nil {
	333	+	t.Fatalf("invalid trace accepted: %+v", records)
	334	+	}
	335	+	}
	336	+	m := write(t, base("diff_start", 1), base("diff_end", 3))
	337	+	spans, err := mergeProcessTraces([]processMapping{m})
	338	+	if err != nil \|\| len(spans) != 1 \|\| spans[0].Samples[0] != 2 {
	339	+	t.Fatalf("spans=%+v err=%v", spans, err)
	340	+	}
	341	+	}
	342	+
	343	+	func TestHarnessExcludesFailedSpanDurationsWhileValidatingPairing(t *testing.T) {
	344	+	base := func(kind string, tick int64, valid bool) traceRecord {
	345	+	return traceRecord{ProcessID: "one", Component: "adapter", Scenario: "s", Run: 1, Sequence: 1, RequestID: 1, Epoch: 1, Kind: kind, Tick: tick, Valid: valid}
	346	+	}
	347	+	for _, tc := range []struct {
	348	+	name string
	349	+	records []traceRecord
	350	+	wantSpans int
	351	+	wantErr bool
	352	+	}{
	353	+	{"successful adapter send is sampled", []traceRecord{base("adapter_send_start", 10, true), base("adapter_send_end", 20, true)}, 1, false},
	354	+	{"failed adapter send end is excluded", []traceRecord{base("adapter_send_start", 10, true), base("adapter_send_end", 20, false)}, 0, false},
	355	+	{"failed adapter receive end is excluded", []traceRecord{base("adapter_receive_start", 10, true), base("adapter_receive_end", 20, false)}, 0, false},
	356	+	{"failed adapter send start is excluded", []traceRecord{base("adapter_send_start", 10, false), base("adapter_send_end", 20, true)}, 0, false},
	357	+	{"failed adapter receive boundaries are excluded", []traceRecord{base("adapter_receive_start", 10, false), base("adapter_receive_end", 20, false)}, 0, false},
	358	+	{"failed end without start is rejected", []traceRecord{base("adapter_send_end", 20, false)}, 0, true},
	359	+	{"duplicate failed start is rejected", []traceRecord{base("adapter_send_start", 10, false), base("adapter_send_start", 15, false)}, 0, true},
	360	+	{"negative failed span is rejected", []traceRecord{base("adapter_receive_start", 20, false), base("adapter_receive_end", 10, false)}, 0, true},
	361	+	{"cross-component failed end is rejected", []traceRecord{base("adapter_send_start", 10, true), func() traceRecord { r := base("adapter_send_end", 20, false); r.Component = "other"; return r }()}, 0, true},
	362	+	} {
	363	+	t.Run(tc.name, func(t *testing.T) {
	364	+	path := filepath.Join(t.TempDir(), "trace.jsonl")
	365	+	var lines []string
	366	+	for _, record := range tc.records {
	367	+	lines = append(lines, mustJSON(record))
	368	+	}
	369	+	if err := os.WriteFile(path, []byte(strings.Join(lines, "\n")+"\n"), 0o600); err != nil {
	370	+	t.Fatal(err)
	371	+	}
	372	+	spans, err := mergeProcessTraces([]processMapping{{ProcessID: "one", ClockDomain: "one", TracePath: path, Scenario: "s", Run: 1}})
	373	+	if tc.wantErr {
	374	+	if err == nil {
	375	+	t.Fatalf("invalid trace accepted: %+v", tc.records)
	376	+	}
	377	+	return
	378	+	}
	379	+	if err != nil \|\| len(spans) != tc.wantSpans {
	380	+	t.Fatalf("spans=%+v err=%v, want %d spans", spans, err, tc.wantSpans)
	381	+	}
	382	+	if tc.wantSpans == 1 && spans[0].Samples[0] != 10 {
	383	+	t.Fatalf("successful span=%+v, want 10ns", spans[0])
	384	+	}
	385	+	})
	386	+	}
	387	+	}

A cmd/vev-perf-harness/launcher.go +758 -0

@@ -0,0 +1,758 @@
	1	+	package main
	2	+
	3	+	import (
	4	+	"bytes"
	5	+	"context"
	6	+	"errors"
	7	+	"fmt"
	8	+	"io"
	9	+	"os"
	10	+	"os/exec"
	11	+	"path/filepath"
	12	+	"strconv"
	13	+	"strings"
	14	+	"sync"
	15	+	"syscall"
	16	+	"time"
	17	+
	18	+	"github.com/bnema/vev/pkg/safedir"
	19	+	"golang.org/x/sys/unix"
	20	+	)
	21	+
	22	+	type launchedProcess interface {
	23	+	Warmup([]byte) error
	24	+	Measure([]byte, func() error, func() error) error
	25	+	Close() error
	26	+	}
	27	+
	28	+	// roleCommand contains only argv accepted by app.parseArgs. Transport routing
	29	+	// metadata configures the harness's public-command seam; it is never passed as
	30	+	// an unrecognised vev flag.
	31	+	type roleCommand struct {
	32	+	Args []string
	33	+	Transport transport
	34	+	}
	35	+
	36	+	type launcher interface {
	37	+	Launch(processMapping, roleCommand) (launchedProcess, error)
	38	+	}
	39	+
	40	+	type launchedRole struct {
	41	+	mapping processMapping
	42	+	process launchedProcess
	43	+	}
	44	+
	45	+	// closeLaunchedRoles closes every role in reverse launch order. errors.Join
	46	+	// retains every cleanup failure in that stable order.
	47	+	func closeLaunchedRoles(processes []launchedRole) error {
	48	+	var failures []error
	49	+	for i := len(processes) - 1; i >= 0; i-- {
	50	+	if err := processes[i].process.Close(); err != nil {
	51	+	failures = append(failures, fmt.Errorf("close %s role: %w", processes[i].mapping.Role, err))
	52	+	}
	53	+	}
	54	+	return errors.Join(failures...)
	55	+	}
	56	+
	57	+	// cliLauncher is deliberately limited to executing the supplied public vev
	58	+	// binary. Scenario orchestration never reaches into an in-process daemon.
	59	+	type cliLauncher struct {
	60	+	bin string
	61	+	netemFactory udpNetemFactory
	62	+	mu sync.Mutex
	63	+	peers map[string]peerRoute
	64	+	runtimes map[string]string
	65	+	}
	66	+
	67	+	type peerRoute struct {
	68	+	mapping processMapping
	69	+	command roleCommand
	70	+	pidPath string
	71	+	netem udpNetem
	72	+	}
	73	+
	74	+	type preparedPeer struct {
	75	+	route peerRoute
	76	+	cleanupRuntime func() error
	77	+	}
	78	+
	79	+	func (p preparedPeer) Warmup([]byte) error { return nil }
	80	+
	81	+	func (p preparedPeer) Measure([]byte, func() error, func() error) error { return nil }
	82	+
	83	+	func (p preparedPeer) Close() error {
	84	+	var errs []error
	85	+	if p.route.pidPath != "" {
	86	+	b, err := os.ReadFile(p.route.pidPath)
	87	+	if err == nil {
	88	+	pid, parseErr := strconv.Atoi(strings.TrimSpace(string(b)))
	89	+	if parseErr != nil \|\| pid <= 0 {
	90	+	errs = append(errs, fmt.Errorf("invalid owned peer pid: %q", b))
	91	+	} else if killErr := syscall.Kill(pid, syscall.SIGTERM); killErr != nil && !errors.Is(killErr, syscall.ESRCH) {
	92	+	errs = append(errs, killErr)
	93	+	}
	94	+	} else if !errors.Is(err, os.ErrNotExist) {
	95	+	errs = append(errs, err)
	96	+	}
	97	+	}
	98	+	if p.route.netem != nil {
	99	+	if err := p.route.netem.Close(); err != nil {
	100	+	errs = append(errs, err)
	101	+	}
	102	+	}
	103	+	if p.cleanupRuntime != nil {
	104	+	if err := p.cleanupRuntime(); err != nil {
	105	+	errs = append(errs, err)
	106	+	}
	107	+	}
	108	+	return errors.Join(errs...)
	109	+	}
	110	+
	111	+	type terminalOutput interface {
	112	+	io.Writer
	113	+	Sync() error
	114	+	Close() error
	115	+	}
	116	+
	117	+	type cliProcess struct {
	118	+	cmd *exec.Cmd
	119	+	pty io.ReadWriteCloser
	120	+	output terminalOutput
	121	+	chunks chan []byte
	122	+	done chan struct{}
	123	+	closed sync.Once
	124	+	waitErr chan error
	125	+	waitTimeout func() <-chan time.Time
	126	+	forceCleanup func()
	127	+	forceKill func()
	128	+	shutdown func() error
	129	+	cleanupRuntime func() error
	130	+	terminalReady bool
	131	+	readyPath string
	132	+	}
	133	+
	134	+	func (l *cliLauncher) Launch(m processMapping, role roleCommand) (launchedProcess, error) {
	135	+	if len(role.Args) == 0 {
	136	+	return nil, errors.New("public CLI role has no arguments")
	137	+	}
	138	+	if info, err := os.Stat(l.bin); err != nil {
	139	+	return nil, err
	140	+	} else if info.IsDir() {
	141	+	return nil, fmt.Errorf("public CLI binary %q is a directory", l.bin)
	142	+	}
	143	+	if m.Role == "ssh_stdio_peer" \|\| m.Role == "udp_peer" {
	144	+	return l.preparePeer(m, role)
	145	+	}
	146	+	cmd := exec.Command(l.bin, role.Args...)
	147	+	runDir := filepath.Dir(m.TracePath)
	148	+	runtimeDir, err := l.runtimeDir(runDir)
	149	+	if err != nil {
	150	+	return nil, err
	151	+	}
	152	+	// All role inputs are absolute paths resolved at harness startup. Do not set
	153	+	// a role working directory: daemon-owned subprocess cleanup may otherwise
	154	+	// treat the evidence directory as its working tree and remove a preallocated
	155	+	// trace while a later repetition is being merged.
	156	+	cmd.Env = append(withoutEnv(os.Environ(), "VEV", "VEV_PERF_TRACE", "VEV_PERF_PROCESS_ID", "VEV_PERF_SCENARIO", "VEV_PERF_RUN", "VEV_REMOTE_TRANSPORT", "XDG_RUNTIME_DIR", "XDG_STATE_HOME"), traceEnvironment(m)...)
	157	+	cmd.Env = append(cmd.Env, "XDG_RUNTIME_DIR="+runtimeDir,
	158	+	"XDG_STATE_HOME="+filepath.Join(runDir, "state"), "TERM=xterm-256color")
	159	+	if m.Role == "client" {
	160	+	l.mu.Lock()
	161	+	peer, routed := l.peers[runDir]
	162	+	l.mu.Unlock()
	163	+	if routed {
	164	+	mode, err := remoteMode(peer.command.Transport)
	165	+	if err != nil {
	166	+	return nil, err
	167	+	}
	168	+	cmd.Env = append(cmd.Env, "PATH="+runDir+":"+os.Getenv("PATH"), "VEV_REMOTE_TRANSPORT="+mode)
	169	+	}
	170	+	}
	171	+	if err := safedir.EnsurePrivate(filepath.Join(runDir, "state")); err != nil {
	172	+	return nil, err
	173	+	}
	174	+	p := &cliProcess{cmd: cmd, waitErr: make(chan error, 1), readyPath: filepath.Join(runtimeDir, "vev", "daemon.sock")}
	175	+	if m.Role == "daemon" {
	176	+	p.cleanupRuntime = func() error { return l.releaseRuntime(runDir) }
	177	+	}
	178	+	if m.Role == "client" {
	179	+	master, slave, err := openPTY()
	180	+	if err != nil {
	181	+	return nil, err
	182	+	}
	183	+	output, err := os.OpenFile(filepath.Join(filepath.Dir(m.TracePath), "terminal-output"), os.O_WRONLY\|os.O_CREATE\|os.O_APPEND, 0o600)
	184	+	if err != nil {
	185	+	_ = master.Close()
	186	+	_ = slave.Close()
	187	+	return nil, err
	188	+	}
	189	+	cmd.Stdin, cmd.Stdout, cmd.Stderr = slave, slave, slave
	190	+	// Ctty is the child descriptor number after Start remaps slave to stdin.
	191	+	cmd.SysProcAttr = &syscall.SysProcAttr{Setsid: true, Setctty: true, Ctty: 0}
	192	+	if err := cmd.Start(); err != nil {
	193	+	_ = output.Close()
	194	+	_ = master.Close()
	195	+	_ = slave.Close()
	196	+	if p.cleanupRuntime != nil {
	197	+	_ = p.cleanupRuntime()
	198	+	}
	199	+	return nil, err
	200	+	}
	201	+	_ = slave.Close()
	202	+	p.pty, p.output, p.chunks, p.done = master, output, make(chan []byte, 32), make(chan struct{})
	203	+	go p.copyTerminal()
	204	+	} else {
	205	+	// Every launched role owns a dedicated process group, so the bounded
	206	+	// cleanup path never signals the harness/test runner's group.
	207	+	cmd.SysProcAttr = &syscall.SysProcAttr{Setpgid: true}
	208	+	devNull, err := os.OpenFile(os.DevNull, os.O_RDWR, 0)
	209	+	if err != nil {
	210	+	return nil, err
	211	+	}
	212	+	cmd.Stdin, cmd.Stdout, cmd.Stderr = devNull, devNull, devNull
	213	+	if err := cmd.Start(); err != nil {
	214	+	_ = devNull.Close()
	215	+	if p.cleanupRuntime != nil {
	216	+	_ = p.cleanupRuntime()
	217	+	}
	218	+	return nil, err
	219	+	}
	220	+	_ = devNull.Close()
	221	+	}
	222	+	if m.Role == "daemon" {
	223	+	// Stop the foreground daemon through its public CLI before its bounded
	224	+	// process-group fallback. This lets blocked carriage operations close and
	225	+	// serialize their failed end marks instead of being cut off by SIGKILL.
	226	+	p.shutdown = func() error {
	227	+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Second)
	228	+	defer cancel()
	229	+	shutdown := exec.CommandContext(ctx, l.bin, "kill", "--daemon")
	230	+	shutdown.Env = withoutEnv(cmd.Env, "VEV_PERF_TRACE", "VEV_PERF_PROCESS_ID", "VEV_PERF_SCENARIO", "VEV_PERF_RUN")
	231	+	return shutdown.Run()
	232	+	}
	233	+	}
	234	+	go func() { p.waitErr <- cmd.Wait() }()
	235	+	return p, nil
	236	+	}
	237	+
	238	+	// runtimeDir returns one short, private runtime directory shared by the roles
	239	+	// of one run. Keeping the Unix socket outside deeply nested evidence output
	240	+	// avoids the platform socket-path limit without weakening run isolation.
	241	+	func (l *cliLauncher) runtimeDir(runDir string) (string, error) {
	242	+	l.mu.Lock()
	243	+	defer l.mu.Unlock()
	244	+	if path := l.runtimes[runDir]; path != "" {
	245	+	return path, nil
	246	+	}
	247	+	path, err := os.MkdirTemp("", "vev-harness-runtime-")
	248	+	if err != nil {
	249	+	return "", err
	250	+	}
	251	+	if err := safedir.EnsurePrivate(path); err != nil {
	252	+	_ = os.RemoveAll(path)
	253	+	return "", err
	254	+	}
	255	+	if l.runtimes == nil {
	256	+	l.runtimes = make(map[string]string)
	257	+	}
	258	+	l.runtimes[runDir] = path
	259	+	return path, nil
	260	+	}
	261	+
	262	+	func (l *cliLauncher) releaseRuntime(runDir string) error {
	263	+	l.mu.Lock()
	264	+	path := l.runtimes[runDir]
	265	+	delete(l.runtimes, runDir)
	266	+	l.mu.Unlock()
	267	+	if path == "" {
	268	+	return nil
	269	+	}
	270	+	return os.RemoveAll(path)
	271	+	}
	272	+
	273	+	// preparePeer installs a per-run ssh command seam. The ordinary public client
	274	+	// still executes its documented remote attach command; its ssh child invokes
	275	+	// only the parsed public _stdio or _udp-proxy command. This makes the peer
	276	+	// carrying client traffic the declared, exclusively traced role rather than a
	277	+	// separately launched but unused process.
	278	+	func (l *cliLauncher) preparePeer(m processMapping, role roleCommand) (launchedProcess, error) {
	279	+	runDir := filepath.Dir(m.TracePath)
	280	+	l.mu.Lock()
	281	+	runtimeDir := l.runtimes[runDir]
	282	+	l.mu.Unlock()
	283	+	peerOwnsRuntime := runtimeDir == ""
	284	+	fail := func(err error) error {
	285	+	if peerOwnsRuntime {
	286	+	return errors.Join(err, l.releaseRuntime(runDir))
	287	+	}
	288	+	return err
	289	+	}
	290	+	if peerOwnsRuntime {
	291	+	var err error
	292	+	runtimeDir, err = l.runtimeDir(runDir)
	293	+	if err != nil {
	294	+	return nil, fail(err)
	295	+	}
	296	+	}
	297	+	if err := safedir.EnsurePrivate(runDir); err != nil {
	298	+	return nil, fail(err)
	299	+	}
	300	+	if err := safedir.EnsurePrivate(runtimeDir); err != nil {
	301	+	return nil, fail(err)
	302	+	}
	303	+	if err := safedir.EnsurePrivate(filepath.Join(runDir, "state")); err != nil {
	304	+	return nil, fail(err)
	305	+	}
	306	+	if len(role.Args) != 2 \|\| (role.Args[0] != "_stdio" && role.Args[0] != "_udp-proxy") {
	307	+	return nil, fail(fmt.Errorf("unsupported public peer command %q", role.Args))
	308	+	}
	309	+	route := peerRoute{mapping: m, command: role}
	310	+	closeNetem := func(err error) error {
	311	+	if route.netem != nil {
	312	+	err = errors.Join(err, route.netem.Close())
	313	+	}
	314	+	return fail(err)
	315	+	}
	316	+	if m.Role == "udp_peer" {
	317	+	route.pidPath = filepath.Join(runDir, "udp-peer.pid")
	318	+	factory := l.netemFactory
	319	+	if factory == nil {
	320	+	factory = newUDPNetem
	321	+	}
	322	+	netem, err := factory(udpNetemConfig{
	323	+	RTT: time.Duration(role.Transport.RTTMS) * time.Millisecond,
	324	+	LossPercent: role.Transport.LossPercent,
	325	+	TargetPath: filepath.Join(runDir, "udp-peer.target"),
	326	+	})
	327	+	if err != nil {
	328	+	return nil, fail(fmt.Errorf("start harness UDP netem: %w", err))
	329	+	}
	330	+	route.netem = netem
	331	+	}
	332	+	shim := filepath.Join(runDir, "ssh")
	333	+	ready := filepath.Join(runDir, "udp-peer.ready")
	334	+	target := filepath.Join(runDir, "udp-peer.target")
	335	+	netemPort := 0
	336	+	if route.netem != nil {
	337	+	netemPort = route.netem.Port()
	338	+	}
	339	+	stderr := filepath.Join(runDir, "udp-peer.stderr")
	340	+	// The bootstrap protocol requires the SSH command to exit after forwarding
	341	+	// the readiness line. The long-lived proxy is owned by pidPath and receives
	342	+	// the role's unique trace identity, never the bootstrap/client identity.
	343	+	body := fmt.Sprintf(`#!/bin/sh
	344	+	set -eu
	345	+	case "$*" in
	346	+	"_stdio")
	347	+	exec env VEV_PERF_TRACE=%[1]q VEV_PERF_PROCESS_ID=%[2]q VEV_PERF_SCENARIO=%[12]q VEV_PERF_RUN=%[13]d XDG_RUNTIME_DIR=%[3]q XDG_STATE_HOME=%[4]q TERM=xterm-256color %[5]q _stdio %[6]q
	348	+	;;
	349	+	"_udp-bootstrap")
	350	+	rm -f %[7]q %[8]q %[9]q
	351	+	env VEV_PERF_TRACE=%[1]q VEV_PERF_PROCESS_ID=%[2]q VEV_PERF_SCENARIO=%[12]q VEV_PERF_RUN=%[13]d XDG_RUNTIME_DIR=%[3]q XDG_STATE_HOME=%[4]q TERM=xterm-256color %[5]q _udp-proxy %[6]q >%[7]q 2>%[8]q &
	352	+	peer=$!
	353	+	printf '%%s\n' "$peer" > %[11]q
	354	+	i=0
	355	+	while [ ! -s %[7]q ] && [ "$i" -lt 1000 ]; do sleep 0.01; i=$((i+1)); done
	356	+	test -s %[7]q
	357	+	line=$(head -n 1 %[7]q)
	358	+	printf '%%s\n' "$line" > %[9]q
	359	+	set -- $line
	360	+	test "$1" = VEV-UDP
	361	+	test -n "$3"
	362	+	printf 'VEV-UDP %%s %%s\n' %[10]d "$3"
	363	+	exit 0
	364	+	;;
	365	+	*) echo 'vev harness ssh seam rejected non-vev command' >&2; exit 64 ;;
	366	+	esac
	367	+	`, m.TracePath, m.ProcessID, runtimeDir, filepath.Join(runDir, "state"), l.bin, role.Args[1], ready, stderr, target, netemPort, route.pidPath, m.Scenario, m.Run)
	368	+	if err := os.WriteFile(shim, []byte(body), 0o700); err != nil {
	369	+	return nil, closeNetem(err)
	370	+	}
	371	+	l.mu.Lock()
	372	+	if l.peers == nil {
	373	+	l.peers = make(map[string]peerRoute)
	374	+	}
	375	+	if _, exists := l.peers[runDir]; exists {
	376	+	l.mu.Unlock()
	377	+	return nil, closeNetem(fmt.Errorf("duplicate transport peer for %s", runDir))
	378	+	}
	379	+	l.peers[runDir] = route
	380	+	l.mu.Unlock()
	381	+	peer := preparedPeer{route: route}
	382	+	if peerOwnsRuntime {
	383	+	peer.cleanupRuntime = func() error { return l.releaseRuntime(runDir) }
	384	+	}
	385	+	return peer, nil
	386	+	}
	387	+
	388	+	func remoteMode(t transport) (string, error) {
	389	+	switch t.Kind {
	390	+	case "ssh_stdio":
	391	+	return "stdio", nil
	392	+	case "udp":
	393	+	return "udp", nil
	394	+	default:
	395	+	return "", fmt.Errorf("transport %q cannot route through a remote peer", t.ID)
	396	+	}
	397	+	}
	398	+
	399	+	// traceEnvironment is the complete process-local identity supplied by the
	400	+	// persisted harness processMapping. App turns Scenario/Run into production
	401	+	// mark correlation; sequence/request/epoch remain process-local operation IDs.
	402	+	func traceEnvironment(m processMapping) []string {
	403	+	return []string{
	404	+	"VEV_PERF_TRACE=" + m.TracePath,
	405	+	"VEV_PERF_PROCESS_ID=" + m.ProcessID,
	406	+	"VEV_PERF_SCENARIO=" + m.Scenario,
	407	+	"VEV_PERF_RUN=" + strconv.Itoa(m.Run),
	408	+	}
	409	+	}
	410	+
	411	+	func withoutEnv(env []string, names ...string) []string {
	412	+	remove := make(map[string]bool, len(names))
	413	+	for _, name := range names {
	414	+	remove[name] = true
	415	+	}
	416	+	out := make([]string, 0, len(env))
	417	+	for _, entry := range env {
	418	+	name, _, _ := strings.Cut(entry, "=")
	419	+	if !remove[name] {
	420	+	out = append(out, entry)
	421	+	}
	422	+	}
	423	+	return out
	424	+	}
	425	+
	426	+	func (p *cliProcess) copyTerminal() {
	427	+	defer close(p.chunks)
	428	+	buf := make([]byte, 4096)
	429	+	for {
	430	+	n, err := p.pty.Read(buf)
	431	+	if n > 0 {
	432	+	chunk := append([]byte(nil), buf[:n]...)
	433	+	if _, writeErr := p.output.Write(chunk); writeErr != nil {
	434	+	return
	435	+	}
	436	+	select {
	437	+	case p.chunks <- chunk:
	438	+	case <-p.done:
	439	+	return
	440	+	}
	441	+	}
	442	+	if err != nil {
	443	+	return
	444	+	}
	445	+	}
	446	+	}
	447	+
	448	+	// WaitReady confirms the explicitly launched public daemon owns its unique
	449	+	// socket before a client is launched. Without this barrier, the normal client
	450	+	// auto-spawn behavior can create an unmanifested daemon process during startup.
	451	+	func (p *cliProcess) WaitReady() error {
	452	+	if p.readyPath == "" {
	453	+	return errors.New("daemon readiness path is empty")
	454	+	}
	455	+	ticker := time.NewTicker(10 * time.Millisecond)
	456	+	defer ticker.Stop()
	457	+	timer := time.NewTimer(10 * time.Second)
	458	+	defer timer.Stop()
	459	+	for {
	460	+	if info, err := os.Stat(p.readyPath); err == nil {
	461	+	if info.Mode()&os.ModeSocket == 0 {
	462	+	return fmt.Errorf("daemon readiness path %s is not a socket", p.readyPath)
	463	+	}
	464	+	return nil
	465	+	}
	466	+	select {
	467	+	case err := <-p.waitErr:
	468	+	p.waitErr <- err
	469	+	if err == nil {
	470	+	return errors.New("daemon exited before readiness")
	471	+	}
	472	+	return fmt.Errorf("daemon exited before readiness: %w", err)
	473	+	case <-ticker.C:
	474	+	case <-timer.C:
	475	+	return fmt.Errorf("timed out waiting for daemon socket %s", p.readyPath)
	476	+	}
	477	+	}
	478	+	}
	479	+
	480	+	func (p *cliProcess) Warmup(input []byte) error {
	481	+	if p.pty == nil \|\| len(input) == 0 {
	482	+	return nil
	483	+	}
	484	+	// A client cannot write application output until it has received Welcome and
	485	+	// entered raw mode. Require that observable output before submitting any
	486	+	// warmup input, so the PTY's line discipline cannot echo it locally.
	487	+	if !p.terminalReady {
	488	+	if err := p.waitForTerminalReady(); err != nil {
	489	+	return err
	490	+	}
	491	+	p.terminalReady = true
	492	+	}
	493	+	if _, err := p.pty.Write(input); err != nil {
	494	+	return err
	495	+	}
	496	+	// Do not begin measured input until the public client has rendered a real
	497	+	// shell response to warmup input. This is intentionally not a measurement.
	498	+	return p.waitForTerminalMarker(inputMarker(input), input)
	499	+	}
	500	+
	501	+	// Measure records a boundary only after the harness-owned terminal output file
	502	+	// has accepted application output and Sync has reported a successful flush.
	503	+	// The marker is injected through the real client PTY, never fabricated by the
	504	+	// harness. PTY local echo is discarded and cannot satisfy this boundary.
	505	+	func (p *cliProcess) Measure(input []byte, injected, flushed func() error) error {
	506	+	if p.pty == nil {
	507	+	return errors.New("measured process has no terminal PTY")
	508	+	}
	509	+	if _, err := p.pty.Write(input); err != nil {
	510	+	return err
	511	+	}
	512	+	// The injected callback runs only after the owned PTY accepts the bytes.
	513	+	if err := injected(); err != nil {
	514	+	return err
	515	+	}
	516	+	if err := p.waitForTerminalMarker(inputMarker(input), input); err != nil {
	517	+	return err
	518	+	}
	519	+	if err := p.output.Sync(); err != nil {
	520	+	return fmt.Errorf("flush terminal output: %w", err)
	521	+	}
	522	+	return flushed()
	523	+	}
	524	+
	525	+	// waitForTerminalReady waits for public client output produced before any
	526	+	// harness input. The client enters raw mode before its output loop can render
	527	+	// this initial prompt/state, so this is an observable readiness boundary.
	528	+	func (p *cliProcess) waitForTerminalReady() error {
	529	+	timer := time.NewTimer(10 * time.Second)
	530	+	defer timer.Stop()
	531	+	for {
	532	+	select {
	533	+	case chunk, ok := <-p.chunks:
	534	+	if !ok {
	535	+	return p.terminalClosedError("readiness output")
	536	+	}
	537	+	if len(chunk) == 0 {
	538	+	continue
	539	+	}
	540	+	if err := p.output.Sync(); err != nil {
	541	+	return fmt.Errorf("flush terminal readiness output: %w", err)
	542	+	}
	543	+	return nil
	544	+	case err := <-p.waitErr:
	545	+	// Keep the exit result available to Close, which owns reaping.
	546	+	p.waitErr <- err
	547	+	if err == nil {
	548	+	return errors.New("client exited before readiness output")
	549	+	}
	550	+	return fmt.Errorf("client exited before readiness output: %w", err)
	551	+	case <-timer.C:
	552	+	return errors.New("timed out waiting for client readiness output")
	553	+	}
	554	+	}
	555	+	}
	556	+
	557	+	func (p *cliProcess) terminalClosedError(boundary string) error {
	558	+	if p.waitErr != nil {
	559	+	select {
	560	+	case err := <-p.waitErr:
	561	+	// Close owns reaping and may need the result after a failed boundary.
	562	+	p.waitErr <- err
	563	+	if err != nil {
	564	+	return fmt.Errorf("client terminal closed before %s: %w", boundary, err)
	565	+	}
	566	+	case <-time.After(10 * time.Millisecond):
	567	+	}
	568	+	}
	569	+	return fmt.Errorf("client terminal closed before %s", boundary)
	570	+	}
	571	+
	572	+	func (p *cliProcess) waitForTerminalMarker(marker, input []byte) error {
	573	+	var received []byte
	574	+	echo := newPTYLocalEcho(input)
	575	+	timer := time.NewTimer(10 * time.Second)
	576	+	defer timer.Stop()
	577	+	for !bytes.Contains(received, marker) {
	578	+	select {
	579	+	case chunk, ok := <-p.chunks:
	580	+	if !ok {
	581	+	return p.terminalClosedError("application output")
	582	+	}
	583	+	received = append(received, echo.applicationOutput(chunk)...)
	584	+	case err := <-p.waitErr:
	585	+	// Keep the exit result available to Close, which owns reaping.
	586	+	p.waitErr <- err
	587	+	if err == nil {
	588	+	return errors.New("client exited before application output")
	589	+	}
	590	+	return fmt.Errorf("client exited before application output: %w", err)
	591	+	case <-timer.C:
	592	+	return errors.New("timed out waiting for application terminal output")
	593	+	}
	594	+	}
	595	+	return nil
	596	+	}
	597	+
	598	+	// ptyLocalEcho removes the one ordered input echo that a not-yet-raw slave PTY
	599	+	// can produce. It normalizes NL to the conventional ONLCR echo spelling; any
	600	+	// divergence is retained as application output. The line discipline emits the
	601	+	// echo before the child can read and respond to the same write.
	602	+	type ptyLocalEcho struct {
	603	+	expected []byte
	604	+	matched int
	605	+	done bool
	606	+	}
	607	+
	608	+	func (e *ptyLocalEcho) applicationOutput(chunk []byte) []byte {
	609	+	if e.done \|\| len(e.expected) == 0 {
	610	+	return chunk
	611	+	}
	612	+	for i, b := range chunk {
	613	+	if b != e.expected[e.matched] {
	614	+	e.done = true
	615	+	return append([]byte(nil), chunk[i:]...)
	616	+	}
	617	+	e.matched++
	618	+	if e.matched == len(e.expected) {
	619	+	e.done = true
	620	+	return append([]byte(nil), chunk[i+1:]...)
	621	+	}
	622	+	}
	623	+	return nil
	624	+	}
	625	+
	626	+	func (p *cliProcess) Close() error {
	627	+	var result error
	628	+	p.closed.Do(func() {
	629	+	if p.done != nil {
	630	+	close(p.done)
	631	+	}
	632	+	// A public terminal client owns a shell session. Ask that shell to exit
	633	+	// before closing the PTY so its transport can finish its receive span;
	634	+	// this avoids turning ordinary harness teardown into an aborted trace.
	635	+	exited := false
	636	+	if p.shutdown != nil && p.waitErr != nil {
	637	+	select {
	638	+	case <-p.waitErr:
	639	+	exited = true
	640	+	default:
	641	+	// The client may have just ended the last session and raced this
	642	+	// best-effort public shutdown. The normal bounded reaping path below
	643	+	// remains authoritative in that case.
	644	+	_ = p.shutdown()
	645	+	}
	646	+	}
	647	+	if p.pty != nil && p.waitErr != nil {
	648	+	_, _ = p.pty.Write([]byte("exit\n"))
	649	+	select {
	650	+	case <-p.waitErr:
	651	+	exited = true
	652	+	case <-p.closeDeadline():
	653	+	}
	654	+	}
	655	+	if p.pty != nil {
	656	+	_ = p.pty.Close()
	657	+	}
	658	+	if p.waitErr != nil && !exited {
	659	+	select {
	660	+	case <-p.waitErr:
	661	+	case <-p.closeDeadline():
	662	+	// Teardown is deliberately bounded: a stuck public CLI or child
	663	+	// must not hold a local harness run indefinitely.
	664	+	p.forceProcessGroupCleanup()
	665	+	select {
	666	+	case <-p.waitErr:
	667	+	case <-p.closeDeadline():
	668	+	p.forceProcessGroupKill()
	669	+	select {
	670	+	case <-p.waitErr:
	671	+	case <-p.closeDeadline():
	672	+	result = errors.New("timed out reaping public CLI process")
	673	+	}
	674	+	}
	675	+	}
	676	+	}
	677	+	if p.output != nil {
	678	+	if err := p.output.Close(); result == nil {
	679	+	result = err
	680	+	}
	681	+	}
	682	+	if p.cleanupRuntime != nil {
	683	+	if err := p.cleanupRuntime(); result == nil {
	684	+	result = err
	685	+	}
	686	+	}
	687	+	})
	688	+	return result
	689	+	}
	690	+
	691	+	func (p *cliProcess) closeDeadline() <-chan time.Time {
	692	+	if p.waitTimeout != nil {
	693	+	return p.waitTimeout()
	694	+	}
	695	+	if p.shutdown != nil {
	696	+	// A public kill asks the daemon to finish bounded detached notices and
	697	+	// close blocked carriage before it exits; do not turn that graceful
	698	+	// protocol into a SIGKILL at the two-second client deadline.
	699	+	return time.After(5 * time.Second)
	700	+	}
	701	+	return time.After(2 * time.Second)
	702	+	}
	703	+
	704	+	func (p *cliProcess) forceProcessGroupCleanup() {
	705	+	if p.forceCleanup != nil {
	706	+	p.forceCleanup()
	707	+	return
	708	+	}
	709	+	if p.cmd != nil && p.cmd.Process != nil {
	710	+	// The client is a session/process-group leader; forced cleanup must
	711	+	// include its ssh seam/_stdio descendants.
	712	+	_ = syscall.Kill(-p.cmd.Process.Pid, syscall.SIGTERM)
	713	+	}
	714	+	}
	715	+
	716	+	func (p *cliProcess) forceProcessGroupKill() {
	717	+	if p.forceKill != nil {
	718	+	p.forceKill()
	719	+	return
	720	+	}
	721	+	if p.cmd != nil && p.cmd.Process != nil {
	722	+	_ = syscall.Kill(-p.cmd.Process.Pid, syscall.SIGKILL)
	723	+	}
	724	+	}
	725	+
	726	+	// openPTY is the small Linux PTY boundary needed to drive the public terminal
	727	+	// client. No daemon or adapter implementation is imported by this command.
	728	+	func openPTY() (os.File, os.File, error) {
	729	+	masterFD, err := unix.Open("/dev/ptmx", unix.O_RDWR\|unix.O_NOCTTY, 0)
	730	+	if err != nil {
	731	+	return nil, nil, err
	732	+	}
	733	+	closeMaster := func(err error) (os.File, os.File, error) { _ = unix.Close(masterFD); return nil, nil, err }
	734	+	number, err := unix.IoctlGetInt(masterFD, unix.TIOCGPTN)
	735	+	if err != nil {
	736	+	return closeMaster(err)
	737	+	}
	738	+	if err := unix.IoctlSetPointerInt(masterFD, unix.TIOCSPTLCK, 0); err != nil {
	739	+	return closeMaster(err)
	740	+	}
	741	+	slaveFD, err := unix.Open(fmt.Sprintf("/dev/pts/%d", number), unix.O_RDWR\|unix.O_NOCTTY, 0)
	742	+	if err != nil {
	743	+	return closeMaster(err)
	744	+	}
	745	+	return os.NewFile(uintptr(masterFD), "vev-client-master"), os.NewFile(uintptr(slaveFD), "vev-client-slave"), nil
	746	+	}
	747	+
	748	+	func inputMarker(input []byte) []byte {
	749	+	start := bytes.Index(input, []byte("__VEV_HARNESS_"))
	750	+	if start < 0 {
	751	+	return input
	752	+	}
	753	+	end := bytes.Index(input[start:], []byte(`\n`))
	754	+	if end < 0 {
	755	+	return input[start:]
	756	+	}
	757	+	return input[start : start+end]
	758	+	}

A cmd/vev-perf-harness/launcher_test.go +326 -0

@@ -0,0 +1,326 @@
	1	+	package main
	2	+
	3	+	import (
	4	+	"bytes"
	5	+	"context"
	6	+	"fmt"
	7	+	"net"
	8	+	"os"
	9	+	"os/exec"
	10	+	"path/filepath"
	11	+	"strings"
	12	+	"testing"
	13	+	"time"
	14	+
	15	+	"github.com/bnema/vev/pkg/safedir"
	16	+	)
	17	+
	18	+	func TestCLIProcessCloseIsBoundedWhenPublicRoleDoesNotExit(t *testing.T) {
	19	+	deadline := make(chan time.Time)
	20	+	close(deadline)
	21	+	var term, kill int
	22	+	p := &cliProcess{
	23	+	waitErr: make(chan error),
	24	+	waitTimeout: func() <-chan time.Time { return deadline },
	25	+	forceCleanup: func() { term++ },
	26	+	forceKill: func() { kill++ },
	27	+	}
	28	+	if err := p.Close(); err == nil {
	29	+	t.Fatal("stuck public role close succeeded")
	30	+	}
	31	+	if term != 1 \|\| kill != 1 {
	32	+	t.Fatalf("TERM/KILL calls = %d/%d, want 1/1", term, kill)
	33	+	}
	34	+	}
	35	+
	36	+	func TestCLIProcessWaitReadyRequiresDaemonSocket(t *testing.T) {
	37	+	path := filepath.Join(t.TempDir(), "daemon.sock")
	38	+	var lc net.ListenConfig
	39	+	listener, err := lc.Listen(context.Background(), "unix", path)
	40	+	if err != nil {
	41	+	t.Fatal(err)
	42	+	}
	43	+	t.Cleanup(func() {
	44	+	if err := listener.Close(); err != nil {
	45	+	t.Error(err)
	46	+	}
	47	+	})
	48	+	if err := (&cliProcess{readyPath: path, waitErr: make(chan error, 1)}).WaitReady(); err != nil {
	49	+	t.Fatal(err)
	50	+	}
	51	+	}
	52	+
	53	+	func TestCLILauncherIsolatesRoleXDGEnvironmentAcrossRepetitions(t *testing.T) {
	54	+	captureDir := t.TempDir()
	55	+	t.Setenv("CAPTURE_DIR", captureDir)
	56	+	t.Setenv("XDG_RUNTIME_DIR", filepath.Join(t.TempDir(), "inherited-runtime"))
	57	+	t.Setenv("XDG_STATE_HOME", filepath.Join(t.TempDir(), "inherited-state"))
	58	+
	59	+	bin := filepath.Join(t.TempDir(), "capture-env")
	60	+	if err := os.WriteFile(bin, []byte("#!/bin/sh\nenv > \"$CAPTURE_DIR/$VEV_PERF_PROCESS_ID\"\n"), 0o700); err != nil {
	61	+	t.Fatal(err)
	62	+	}
	63	+	launcher := &cliLauncher{bin: bin}
	64	+	for run := 1; run <= 3; run++ {
	65	+	runDir := filepath.Join(t.TempDir(), fmt.Sprintf("run-%d", run))
	66	+	if err := safedir.EnsurePrivate(runDir); err != nil {
	67	+	t.Fatal(err)
	68	+	}
	69	+	roles := []string{"daemon", "client"}
	70	+	for _, role := range roles {
	71	+	id := fmt.Sprintf("run-%d-%s", run, role)
	72	+	p, err := launcher.Launch(processMapping{
	73	+	Role: role, TracePath: filepath.Join(runDir, role+".jsonl"), ProcessID: id,
	74	+	}, roleCommand{Args: []string{"test"}})
	75	+	if err != nil {
	76	+	t.Fatalf("run %d %s launch: %v", run, role, err)
	77	+	}
	78	+	runtimeDir := launcher.runtimes[runDir]
	79	+	if err := p.Close(); err != nil {
	80	+	t.Fatalf("run %d %s close: %v", run, role, err)
	81	+	}
	82	+
	83	+	env, err := os.ReadFile(filepath.Join(captureDir, id))
	84	+	if err != nil {
	85	+	t.Fatalf("run %d %s capture: %v", run, role, err)
	86	+	}
	87	+	for _, want := range []string{
	88	+	"XDG_RUNTIME_DIR=" + runtimeDir,
	89	+	"XDG_STATE_HOME=" + filepath.Join(runDir, "state"),
	90	+	} {
	91	+	name := strings.SplitN(want, "=", 2)[0]
	92	+	if count := bytes.Count(env, []byte(name+"=")); count != 1 {
	93	+	t.Fatalf("run %d %s %s entries = %d, want exactly one; env:\n%s", run, role, name, count, env)
	94	+	}
	95	+	if !bytes.Contains(env, []byte(want+"\n")) {
	96	+	t.Fatalf("run %d %s missing %q; env:\n%s", run, role, want, env)
	97	+	}
	98	+	}
	99	+	}
	100	+	if err := launcher.releaseRuntime(runDir); err != nil {
	101	+	t.Fatalf("run %d release runtime: %v", run, err)
	102	+	}
	103	+	}
	104	+	}
	105	+
	106	+	func TestPreparePeerReleasesOwnedRuntimeOnPreparationFailure(t *testing.T) {
	107	+	t.Setenv("TMPDIR", t.TempDir())
	108	+	runDir := filepath.Join(t.TempDir(), "run")
	109	+	if err := safedir.EnsurePrivate(runDir); err != nil {
	110	+	t.Fatal(err)
	111	+	}
	112	+	launcher := &cliLauncher{}
	113	+	_, err := launcher.preparePeer(processMapping{
	114	+	Role: "ssh_stdio_peer",
	115	+	TracePath: filepath.Join(runDir, "peer.jsonl"),
	116	+	}, roleCommand{Args: []string{"invalid"}})
	117	+	if err == nil \|\| !strings.Contains(err.Error(), "unsupported public peer command") {
	118	+	t.Fatalf("preparePeer error = %v", err)
	119	+	}
	120	+	if runtimeDir := launcher.runtimes[runDir]; runtimeDir != "" {
	121	+	t.Fatalf("failed peer preparation leaked runtime directory %q", runtimeDir)
	122	+	}
	123	+	if paths, globErr := filepath.Glob(filepath.Join(os.TempDir(), "vev-harness-runtime-*")); globErr != nil \|\| len(paths) != 0 {
	124	+	t.Fatalf("failed peer preparation left runtime directories %q: %v", paths, globErr)
	125	+	}
	126	+	}
	127	+
	128	+	func TestCLIProcessCloseUsesPublicDaemonShutdown(t *testing.T) {
	129	+	wait := make(chan error, 1)
	130	+	wait <- nil
	131	+	var shutdown int
	132	+	p := &cliProcess{waitErr: wait, shutdown: func() error { shutdown++; return nil }}
	133	+	if err := p.Close(); err != nil {
	134	+	t.Fatal(err)
	135	+	}
	136	+	if shutdown != 0 {
	137	+	t.Fatalf("shutdown calls = %d, want 0 for an exited daemon", shutdown)
	138	+	}
	139	+
	140	+	wait = make(chan error, 1)
	141	+	p = &cliProcess{waitErr: wait, shutdown: func() error { shutdown++; wait <- nil; return nil }}
	142	+	if err := p.Close(); err != nil {
	143	+	t.Fatal(err)
	144	+	}
	145	+	if shutdown != 1 {
	146	+	t.Fatalf("shutdown calls = %d, want 1", shutdown)
	147	+	}
	148	+	}
	149	+
	150	+	func TestHarnessCanonicalLocalSmokeRealRoles(t *testing.T) {
	151	+	if testing.Short() {
	152	+	t.Skip("public CLI smoke")
	153	+	}
	154	+	root, err := filepath.Abs(filepath.Join("..", ".."))
	155	+	if err != nil {
	156	+	t.Fatal(err)
	157	+	}
	158	+	bin := filepath.Join(t.TempDir(), "vev")
	159	+	build := exec.Command("go", "build", "-o", bin, "./")
	160	+	build.Dir = root
	161	+	if output, err := build.CombinedOutput(); err != nil {
	162	+	t.Fatalf("build public CLI: %v\n%s", err, output)
	163	+	}
	164	+	m, err := readManifest(filepath.Join(root, "testdata", "perf", "manifest.json"))
	165	+	if err != nil {
	166	+	t.Fatal(err)
	167	+	}
	168	+	if err := validateManifest(m); err != nil {
	169	+	t.Fatal(err)
	170	+	}
	171	+	var local scenario
	172	+	for _, s := range m.Scenarios {
	173	+	if s.Topology == "1x4" && s.Workload == "idle" && s.Transport == "local" {
	174	+	local = s
	175	+	break
	176	+	}
	177	+	}
	178	+	if local.ID == "" {
	179	+	t.Fatal("canonical local scenario missing")
	180	+	}
	181	+	dir := t.TempDir()
	182	+	raw, err := os.OpenFile(filepath.Join(dir, "raw.jsonl"), os.O_CREATE\|os.O_EXCL\|os.O_WRONLY, 0o600)
	183	+	if err != nil {
	184	+	t.Fatal(err)
	185	+	}
	186	+	closeTestFile(t, raw)
	187	+	h := defaultHarness()
	188	+	h.clock = &fakeClock{} // injected duration keeps the public 30s contract bounded in test
	189	+	h.launcher = &cliLauncher{bin: bin}
	190	+	result, err := h.runOne(options{vevBin: bin, out: dir, duration: minimumDuration}, m, local, 1, raw)
	191	+	if err != nil {
	192	+	t.Fatal(err)
	193	+	}
	194	+	if result.Samples < minimumInIntervalEventSamples \|\| len(result.Processes) != 2 {
	195	+	t.Fatalf("canonical local smoke result=%+v", result)
	196	+	}
	197	+	}
	198	+
	199	+	func TestTraceEnvironmentMatchesProcessMapping(t *testing.T) {
	200	+	mapping := processMapping{TracePath: "/tmp/trace.jsonl", ProcessID: "p-01", Scenario: "1x4-idle-local", Run: 3}
	201	+	got := strings.Join(traceEnvironment(mapping), "\n")
	202	+	for _, want := range []string{"VEV_PERF_TRACE=/tmp/trace.jsonl", "VEV_PERF_PROCESS_ID=p-01", "VEV_PERF_SCENARIO=1x4-idle-local", "VEV_PERF_RUN=3"} {
	203	+	if !strings.Contains(got, want) {
	204	+	t.Errorf("trace environment missing %q: %s", want, got)
	205	+	}
	206	+	}
	207	+	}
	208	+
	209	+	func TestCLIProcessCloseGracefulDetachCompletesSpanBeforeForcedCleanup(t *testing.T) {
	210	+	var order []string
	211	+	waitErr := make(chan error, 1)
	212	+	timeout := make(chan time.Time)
	213	+	p := &cliProcess{
	214	+	pty: &orderedPTY{
	215	+	write: func(input []byte) {
	216	+	if string(input) != "exit\n" {
	217	+	t.Fatalf("graceful exit input=%q", input)
	218	+	}
	219	+	order = append(order, "exit_written")
	220	+	// The graceful client exit has closed its transport and emitted
	221	+	// its adapter receive-end mark before cmd.Wait completes.
	222	+	order = append(order, "adapter_receive_end")
	223	+	waitErr <- nil
	224	+	},
	225	+	close: func() { order = append(order, "pty_closed") },
	226	+	},
	227	+	output: &orderedOutput{close: func() { order = append(order, "output_closed") }},
	228	+	waitErr: waitErr,
	229	+	waitTimeout: func() <-chan time.Time { return timeout },
	230	+	forceCleanup: func() {
	231	+	order = append(order, "forced_process_group_cleanup")
	232	+	},
	233	+	}
	234	+
	235	+	if err := p.Close(); err != nil {
	236	+	t.Fatal(err)
	237	+	}
	238	+	if !equalStrings(order, []string{"exit_written", "adapter_receive_end", "pty_closed", "output_closed"}) {
	239	+	t.Fatalf("cleanup order=%q", order)
	240	+	}
	241	+	}
	242	+
	243	+	func TestCLIProcessWarmupWaitsForDelayedTerminalReadiness(t *testing.T) {
	244	+	allowWrite := make(chan struct{})
	245	+	pty, output := &stagedPTY{allow: allowWrite, writes: make(chan []byte, 1)}, &fakeOutput{}
	246	+	p := &cliProcess{pty: pty, output: output, chunks: make(chan []byte)}
	247	+	input := workloadInput(scenario{ID: "s"}, 1, "warmup")
	248	+	done := make(chan error, 1)
	249	+	go func() { done <- p.Warmup(input) }()
	250	+
	251	+	// This is public client output from the initial application prompt/state,
	252	+	// emitted only after the client has entered raw mode. A premature PTY write
	253	+	// blocks on allowWrite and therefore cannot consume this delayed readiness.
	254	+	go func() {
	255	+	p.chunks <- []byte("shell prompt$ ")
	256	+	close(allowWrite)
	257	+	}()
	258	+	select {
	259	+	case got := <-pty.writes:
	260	+	if !bytes.Equal(got, input) {
	261	+	t.Fatalf("warmup write=%q want %q", got, input)
	262	+	}
	263	+	case <-time.After(time.Second):
	264	+	t.Fatal("warmup did not wait for and then proceed from client readiness")
	265	+	}
	266	+	go func() { p.chunks <- append([]byte("application "), inputMarker(input)...); close(p.chunks) }()
	267	+	if err := <-done; err != nil {
	268	+	t.Fatal(err)
	269	+	}
	270	+	if output.syncs != 1 {
	271	+	t.Fatalf("readiness output was not flushed: syncs=%d", output.syncs)
	272	+	}
	273	+	}
	274	+
	275	+	func TestCLIProcessRejectsPTYLocalEchoAsFlushEvidence(t *testing.T) {
	276	+	pty, output := &fakePTY{}, &fakeOutput{}
	277	+	p := &cliProcess{pty: pty, output: output, chunks: make(chan []byte, 1)}
	278	+	input := workloadInput(scenario{ID: "s"}, 1, "measured-1")
	279	+	p.chunks <- newPTYLocalEcho(input).expected
	280	+	close(p.chunks)
	281	+	flushed := false
	282	+	if err := p.Measure(input, func() error { return nil }, func() error { flushed = true; return nil }); err == nil {
	283	+	t.Fatal("local PTY echo satisfied terminal flush boundary")
	284	+	}
	285	+	if flushed \|\| output.syncs != 0 {
	286	+	t.Fatalf("local echo stamped a flush: flushed=%t syncs=%d", flushed, output.syncs)
	287	+	}
	288	+	}
	289	+
	290	+	func TestCLIProcessPairsApplicationOutputWithSuccessfulFlush(t *testing.T) {
	291	+	pty, output := &fakePTY{}, &fakeOutput{}
	292	+	p := &cliProcess{pty: pty, output: output, chunks: make(chan []byte, 1)}
	293	+	input := workloadInput(scenario{ID: "s"}, 1, "measured-1")
	294	+	echo := newPTYLocalEcho(input).expected
	295	+	p.chunks <- append(append([]byte(nil), echo...), append([]byte("application "), inputMarker(input)...)...)
	296	+	order := []string{}
	297	+	if err := p.Measure(input, func() error { order = append(order, "injected"); return nil }, func() error { order = append(order, "flushed"); return nil }); err != nil {
	298	+	t.Fatal(err)
	299	+	}
	300	+	if len(pty.writes) != 1 \|\| output.syncs != 1 \|\| !equalStrings(order, []string{"injected", "flushed"}) {
	301	+	t.Fatalf("writes=%q syncs=%d boundary order=%q", pty.writes, output.syncs, order)
	302	+	}
	303	+	}
	304	+
	305	+	func TestHarnessUsesPublicRoleCommandsAndPTYWorkloads(t *testing.T) {
	306	+	for _, tc := range []struct {
	307	+	role string
	308	+	want []string
	309	+	}{
	310	+	{"daemon", []string{"--daemon"}},
	311	+	{"client", []string{"new", "perf-s-001"}},
	312	+	{"ssh_stdio_peer", []string{"_stdio", "perf-s-001"}},
	313	+	{"udp_peer", []string{"_udp-proxy", "perf-s-001"}},
	314	+	} {
	315	+	t.Run(tc.role, func(t *testing.T) {
	316	+	got := roleArgs(scenario{ID: "s", Transport: "local"}, processMapping{Role: tc.role, Run: 1})
	317	+	if !equalStrings(got.Args, tc.want) {
	318	+	t.Fatalf("args=%q want %q", got.Args, tc.want)
	319	+	}
	320	+	})
	321	+	}
	322	+	input := string(workloadInput(scenario{ID: "s", Workload: "interactive_flood"}, 1, "measured-1"))
	323	+	if string(inputMarker([]byte(input))) != "__VEV_HARNESS_s_r1_measured-1__" \|\| !bytes.Contains([]byte(input), []byte("while [ $i -lt 128 ]")) \|\| !strings.HasSuffix(input, "printf '__VEV_HARNESS_s_r1_measured-1__\\n'\n") {
	324	+	t.Fatalf("workload is not real PTY shell input with an observable marker: %q", input)
	325	+	}
	326	+	}

A cmd/vev-perf-harness/main.go +20 -0

@@ -0,0 +1,20 @@
	1	+	package main
	2	+
	3	+	import (
	4	+	"fmt"
	5	+	"os"
	6	+	"time"
	7	+
	8	+	"github.com/bnema/vev/pkg/safedir"
	9	+	)
	10	+
	11	+	func main() {
	12	+	if err := run(os.Args[1:], defaultHarness()); err != nil {
	13	+	fmt.Fprintln(os.Stderr, "vev-perf-harness:", err)
	14	+	os.Exit(2)
	15	+	}
	16	+	}
	17	+
	18	+	func defaultHarness() *harness {
	19	+	return &harness{clock: systemClock{time.Now()}, mkdir: safedir.EnsurePrivate, createRunDir: createExclusiveRunDir, create: func(p string) (*os.File, error) { return os.OpenFile(p, os.O_WRONLY\|os.O_CREATE\|os.O_EXCL, 0o600) }, removeAll: os.RemoveAll, gitSHA: recordedGitSHA}
	20	+	}

A cmd/vev-perf-harness/model.go +132 -0

@@ -0,0 +1,132 @@
	1	+	package main
	2	+
	3	+	import "time"
	4	+
	5	+	const (
	6	+	minimumDuration = 30 * time.Second
	7	+	minimumRepetitions = 10
	8	+	minimumInIntervalEventSamples = 10
	9	+	measuredEventCadence = time.Second
	10	+	)
	11	+
	12	+	type options struct {
	13	+	vevBin, manifest, out string
	14	+	scenario string
	15	+	warmup, duration time.Duration
	16	+	repetitions int
	17	+	}
	18	+
	19	+	type manifest struct {
	20	+	Schema uint16 `json:"schema"`
	21	+	Topologies []topology `json:"topologies"`
	22	+	Workloads []string `json:"workloads"`
	23	+	Transports []transport `json:"transports"`
	24	+	Scenarios []scenario `json:"scenarios"`
	25	+	}
	26	+
	27	+	type topology struct {
	28	+	ID string `json:"id"`
	29	+	Geometry string `json:"geometry"`
	30	+	RowsPerPane int `json:"rows_per_pane"`
	31	+	}
	32	+
	33	+	type transport struct {
	34	+	ID string `json:"id"`
	35	+	Kind string `json:"kind"`
	36	+	RTTMS int `json:"rtt_ms,omitempty"`
	37	+	LossPercent int `json:"loss_percent,omitempty"`
	38	+	}
	39	+
	40	+	type scenario struct {
	41	+	ID string `json:"id"`
	42	+	Topology string `json:"topology"`
	43	+	Workload string `json:"workload"`
	44	+	Transport string `json:"transport"`
	45	+	Roles []string `json:"roles"`
	46	+	InapplicableReason string `json:"inapplicable_reason,omitempty"`
	47	+	}
	48	+
	49	+	type processMapping struct {
	50	+	ProcessID string `json:"process_id"`
	51	+	ClockDomain string `json:"clock_domain"`
	52	+	TracePath string `json:"trace_path"`
	53	+	Role string `json:"role"`
	54	+	Scenario string `json:"scenario"`
	55	+	Run int `json:"run"`
	56	+	Identity string `json:"identity"`
	57	+	}
	58	+
	59	+	type runManifest struct {
	60	+	Scenario string `json:"scenario"`
	61	+	Run int `json:"run"`
	62	+	Processes []processMapping `json:"processes"`
	63	+	}
	64	+
	65	+	// measuredInterval belongs solely to the harness clock domain. A paired event
	66	+	// is eligible only when both owned boundaries fall inside this interval.
	67	+	type measuredInterval struct {
	68	+	Start int64
	69	+	End int64
	70	+	}
	71	+
	72	+	type eventSample struct {
	73	+	Sequence uint64 `json:"sequence"`
	74	+	Injected int64 `json:"injected_tick"`
	75	+	Flushed int64 `json:"flushed_tick"`
	76	+	Latency int64 `json:"latency_nanos"`
	77	+	}
	78	+
	79	+	type span struct {
	80	+	Component, Name string
	81	+	Samples []int64
	82	+	}
	83	+
	84	+	type runResult struct {
	85	+	Spans []span `json:"process_local_spans"`
	86	+	Scenario string `json:"scenario"`
	87	+	Run int `json:"run"`
	88	+	Samples int `json:"samples"`
	89	+	EndToEnd []int64 `json:"end_to_end_nanos"`
	90	+	Event distribution `json:"event_end_to_end"`
	91	+	Cadence distribution `json:"event_cadence_nanos"`
	92	+	CadenceSamples []int64 `json:"-"`
	93	+	MaxGap int64 `json:"event_max_gap_nanos"`
	94	+	Processes []processMapping `json:"processes"`
	95	+	}
	96	+
	97	+	type summary struct {
	98	+	Schema uint16 `json:"schema"`
	99	+	GitSHA string `json:"git_sha"`
	100	+	Warmup string `json:"warmup"`
	101	+	Duration string `json:"duration"`
	102	+	Repetitions int `json:"repetitions"`
	103	+	EndToEnd distribution `json:"harness_end_to_end"`
	104	+	Cadence distribution `json:"harness_event_cadence_nanos"`
	105	+	MaxGap int64 `json:"harness_event_max_gap_nanos"`
	106	+	RunP50Dispersion distribution `json:"run_p50_dispersion"`
	107	+	Spans []spanSummary `json:"process_local_spans"`
	108	+	Runs int `json:"runs"`
	109	+	}
	110	+
	111	+	type distribution struct {
	112	+	Count int `json:"count"`
	113	+	P50 int64 `json:"p50"`
	114	+	P95 int64 `json:"p95"`
	115	+	P99 int64 `json:"p99"`
	116	+	Max int64 `json:"max"`
	117	+	}
	118	+
	119	+	type spanSummary struct {
	120	+	Component string `json:"component"`
	121	+	Name string `json:"name"`
	122	+	Distribution distribution `json:"distribution"`
	123	+	}
	124	+
	125	+	type clock interface {
	126	+	Now() int64
	127	+	Sleep(time.Duration)
	128	+	}
	129	+
	130	+	type systemClock struct{ start time.Time }
	131	+
	132	+	var spanPairs = []spanPair{{"capture_start", "capture_end", "capture_duration"}, {"compose_start", "compose_end", "compose_duration"}, {"diff_start", "diff_end", "diff_duration"}, {"queue_enqueued", "queue_dequeued", "queue_wait"}, {"ack_blocked_start", "ack_blocked_end", "ack_blocked_interval"}, {"emit_start", "emit_end", "emit_duration"}, {"adapter_send_start", "adapter_send_end", "adapter_send_duration"}, {"adapter_receive_start", "adapter_receive_end", "adapter_receive_duration"}}

A cmd/vev-perf-harness/options_manifest.go +254 -0

@@ -0,0 +1,254 @@
	1	+	package main
	2	+
	3	+	import (
	4	+	"encoding/json"
	5	+	"errors"
	6	+	"flag"
	7	+	"fmt"
	8	+	"io"
	9	+	"os"
	10	+	"path/filepath"
	11	+	)
	12	+
	13	+	var (
	14	+	canonicalTopologyIDs = []string{"1x4", "4x1", "4x4", "8x1"}
	15	+	canonicalTopologySet = stringSet(canonicalTopologyIDs)
	16	+	canonicalWorkloadIDs = []string{
	17	+	"idle", "active_output", "all_output", "inactive_output", "interactive_flood",
	18	+	"copy_search", "resize_sweep", "snapshot_output_resize", "attach_restore_tab_switch",
	19	+	}
	20	+	canonicalWorkloadSet = stringSet(canonicalWorkloadIDs)
	21	+	)
	22	+
	23	+	func stringSet(ids []string) map[string]bool {
	24	+	set := make(map[string]bool, len(ids))
	25	+	for _, id := range ids {
	26	+	set[id] = true
	27	+	}
	28	+	return set
	29	+	}
	30	+
	31	+	func parseOptions(args []string) (options, error) {
	32	+	var o options
	33	+	fs := flag.NewFlagSet("vev-perf-harness", flag.ContinueOnError)
	34	+	fs.SetOutput(io.Discard)
	35	+	fs.StringVar(&o.vevBin, "vev-bin", "", "public vev binary")
	36	+	fs.StringVar(&o.manifest, "manifest", "", "manifest")
	37	+	fs.StringVar(&o.out, "out", "", "output directory")
	38	+	fs.StringVar(&o.scenario, "scenario", "", "one canonical scenario ID (after full manifest validation)")
	39	+	fs.DurationVar(&o.warmup, "warmup", 0, "warmup")
	40	+	fs.DurationVar(&o.duration, "duration", 0, "measurement duration")
	41	+	fs.IntVar(&o.repetitions, "repetitions", 0, "repetitions")
	42	+	if err := fs.Parse(args); err != nil {
	43	+	return o, err
	44	+	}
	45	+	seen := map[string]bool{}
	46	+	fs.Visit(func(f *flag.Flag) { seen[f.Name] = true })
	47	+	for _, name := range []string{"vev-bin", "manifest", "out", "warmup", "duration", "repetitions"} {
	48	+	if !seen[name] {
	49	+	return o, fmt.Errorf("--%s is required", name)
	50	+	}
	51	+	}
	52	+	if o.vevBin == "" \|\| o.manifest == "" \|\| o.out == "" {
	53	+	return o, errors.New("--vev-bin, --manifest, and --out are required")
	54	+	}
	55	+	if o.warmup < 0 {
	56	+	return o, errors.New("--warmup must not be negative")
	57	+	}
	58	+	if o.duration < minimumDuration {
	59	+	return o, fmt.Errorf("--duration must be at least %s", minimumDuration)
	60	+	}
	61	+	if o.repetitions < minimumRepetitions {
	62	+	return o, fmt.Errorf("--repetitions must be at least %d", minimumRepetitions)
	63	+	}
	64	+	return o, nil
	65	+	}
	66	+
	67	+	// resolvePathOptions captures the harness cwd once, before role launchers
	68	+	// assign their isolated working directories. All downstream filesystem access,
	69	+	// manifests, and process environments consequently refer to the same paths.
	70	+	func resolvePathOptions(o options) (options, error) {
	71	+	cwd, err := os.Getwd()
	72	+	if err != nil {
	73	+	return o, fmt.Errorf("resolve harness working directory: %w", err)
	74	+	}
	75	+	absolute := func(path string) string {
	76	+	if filepath.IsAbs(path) {
	77	+	return filepath.Clean(path)
	78	+	}
	79	+	return filepath.Clean(filepath.Join(cwd, path))
	80	+	}
	81	+	o.vevBin = absolute(o.vevBin)
	82	+	o.manifest = absolute(o.manifest)
	83	+	o.out = absolute(o.out)
	84	+	return o, nil
	85	+	}
	86	+
	87	+	func readManifest(path string) (manifest, error) {
	88	+	f, err := os.Open(path)
	89	+	if err != nil {
	90	+	return manifest{}, err
	91	+	}
	92	+	var m manifest
	93	+	if err := json.NewDecoder(f).Decode(&m); err != nil {
	94	+	return manifest{}, closeFile(f, err)
	95	+	}
	96	+	if err := f.Close(); err != nil {
	97	+	return manifest{}, err
	98	+	}
	99	+	return m, nil
	100	+	}
	101	+
	102	+	func validateManifest(m manifest) error {
	103	+	if m.Schema != 1 {
	104	+	return errors.New("unsupported manifest schema")
	105	+	}
	106	+	tops := map[string]bool{}
	107	+	for _, t := range m.Topologies {
	108	+	if t.ID == "" {
	109	+	return errors.New("empty topology id")
	110	+	}
	111	+	if tops[t.ID] {
	112	+	return fmt.Errorf("duplicate topology id %q", t.ID)
	113	+	}
	114	+	if !canonicalTopologySet[t.ID] {
	115	+	return fmt.Errorf("unknown topology id %q", t.ID)
	116	+	}
	117	+	if t.Geometry != "120x40" \|\| t.RowsPerPane != 10000 {
	118	+	return fmt.Errorf("invalid topology %q", t.ID)
	119	+	}
	120	+	tops[t.ID] = true
	121	+	}
	122	+	for _, v := range canonicalTopologyIDs {
	123	+	if !tops[v] {
	124	+	return fmt.Errorf("missing canonical topology %s", v)
	125	+	}
	126	+	}
	127	+	works := map[string]bool{}
	128	+	for _, workload := range m.Workloads {
	129	+	if workload == "" {
	130	+	return errors.New("empty workload id")
	131	+	}
	132	+	if works[workload] {
	133	+	return fmt.Errorf("duplicate workload id %q", workload)
	134	+	}
	135	+	if !canonicalWorkloadSet[workload] {
	136	+	return fmt.Errorf("unknown workload id %q", workload)
	137	+	}
	138	+	works[workload] = true
	139	+	}
	140	+	for _, v := range canonicalWorkloadIDs {
	141	+	if !works[v] {
	142	+	return fmt.Errorf("missing canonical workload %s", v)
	143	+	}
	144	+	}
	145	+	trans := map[string]bool{}
	146	+	for _, t := range m.Transports {
	147	+	if t.ID == "" {
	148	+	return errors.New("empty transport id")
	149	+	}
	150	+	if trans[t.ID] {
	151	+	return fmt.Errorf("duplicate transport id %q", t.ID)
	152	+	}
	153	+	if err := validateTransportFixture(t); err != nil {
	154	+	return err
	155	+	}
	156	+	trans[t.ID] = true
	157	+	}
	158	+	for _, v := range []string{"local", "ssh_stdio", "udp_baseline", "udp_25ms", "udp_100ms", "udp_loss_0pct", "udp_loss_1pct"} {
	159	+	if !trans[v] {
	160	+	return fmt.Errorf("missing canonical transport %s", v)
	161	+	}
	162	+	}
	163	+	seen := map[string]bool{}
	164	+	covered := map[string]bool{}
	165	+	for _, s := range m.Scenarios {
	166	+	if s.ID == "" \|\| seen[s.ID] {
	167	+	return fmt.Errorf("duplicate or empty scenario id %q", s.ID)
	168	+	}
	169	+	seen[s.ID] = true
	170	+	if !tops[s.Topology] \|\| !works[s.Workload] \|\| !trans[s.Transport] {
	171	+	return fmt.Errorf("scenario %s references unknown matrix entry", s.ID)
	172	+	}
	173	+	if s.InapplicableReason == "" && len(s.Roles) == 0 {
	174	+	return fmt.Errorf("scenario %s has no process roles", s.ID)
	175	+	}
	176	+	if s.InapplicableReason != "" && len(s.Roles) != 0 {
	177	+	return fmt.Errorf("inapplicable scenario %s must not launch roles", s.ID)
	178	+	}
	179	+	if s.InapplicableReason == "" && !equalRoleSet(s.Roles, requiredRoles(s.Transport)) {
	180	+	return fmt.Errorf("scenario %s roles do not route transport %s", s.ID, s.Transport)
	181	+	}
	182	+	covered[s.Topology+"\x00"+s.Workload+"\x00"+s.Transport] = true
	183	+	}
	184	+	for topologyID := range tops {
	185	+	for workload := range works {
	186	+	for transportID := range trans {
	187	+	key := topologyID + "\x00" + workload + "\x00" + transportID
	188	+	if !covered[key] {
	189	+	return fmt.Errorf("missing scenario or inapplicable reason for %s/%s/%s", topologyID, workload, transportID)
	190	+	}
	191	+	}
	192	+	}
	193	+	}
	194	+	return nil
	195	+	}
	196	+
	197	+	// selectScenario is intentionally called only after validateManifest. A bounded
	198	+	// local evidence run therefore cannot mask an incomplete canonical matrix.
	199	+	func selectScenario(m manifest, id string) ([]scenario, error) {
	200	+	for _, s := range m.Scenarios {
	201	+	if s.ID != id {
	202	+	continue
	203	+	}
	204	+	if s.InapplicableReason != "" {
	205	+	return nil, fmt.Errorf("scenario %s is inapplicable: %s", id, s.InapplicableReason)
	206	+	}
	207	+	return []scenario{s}, nil
	208	+	}
	209	+	return nil, fmt.Errorf("unknown canonical scenario %s", id)
	210	+	}
	211	+
	212	+	func requiredRoles(transportID string) []string {
	213	+	switch transportID {
	214	+	case "local":
	215	+	return []string{"daemon", "client"}
	216	+	case "ssh_stdio":
	217	+	return []string{"daemon", "client", "ssh_stdio_peer"}
	218	+	default:
	219	+	return []string{"daemon", "client", "udp_peer"}
	220	+	}
	221	+	}
	222	+
	223	+	func equalRoleSet(got, want []string) bool {
	224	+	if len(got) != len(want) {
	225	+	return false
	226	+	}
	227	+	seen := make(map[string]bool, len(got))
	228	+	for _, role := range got {
	229	+	seen[role] = true
	230	+	}
	231	+	for _, role := range want {
	232	+	if !seen[role] {
	233	+	return false
	234	+	}
	235	+	}
	236	+	return true
	237	+	}
	238	+
	239	+	func validateTransportFixture(t transport) error {
	240	+	want := map[string]transport{
	241	+	"local": {ID: "local", Kind: "local"},
	242	+	"ssh_stdio": {ID: "ssh_stdio", Kind: "ssh_stdio"},
	243	+	"udp_baseline": {ID: "udp_baseline", Kind: "udp"},
	244	+	"udp_25ms": {ID: "udp_25ms", Kind: "udp", RTTMS: 25},
	245	+	"udp_100ms": {ID: "udp_100ms", Kind: "udp", RTTMS: 100},
	246	+	"udp_loss_0pct": {ID: "udp_loss_0pct", Kind: "udp", LossPercent: 0},
	247	+	"udp_loss_1pct": {ID: "udp_loss_1pct", Kind: "udp", LossPercent: 1},
	248	+	}
	249	+	w, ok := want[t.ID]
	250	+	if !ok \|\| t.Kind != w.Kind \|\| t.RTTMS != w.RTTMS \|\| t.LossPercent != w.LossPercent {
	251	+	return fmt.Errorf("invalid canonical transport fixture %+v", t)
	252	+	}
	253	+	return nil
	254	+	}

A cmd/vev-perf-harness/options_manifest_test.go +279 -0

@@ -0,0 +1,279 @@
	1	+	package main
	2	+
	3	+	import (
	4	+	"encoding/json"
	5	+	"errors"
	6	+	"os"
	7	+	"path/filepath"
	8	+	"strings"
	9	+	"testing"
	10	+	"time"
	11	+	)
	12	+
	13	+	func TestHarnessScenarioSelectionKeepsCanonicalManifestValidation(t *testing.T) {
	14	+	root, err := filepath.Abs(filepath.Join("..", ".."))
	15	+	if err != nil {
	16	+	t.Fatal(err)
	17	+	}
	18	+	manifestPath := filepath.Join(root, "testdata", "perf", "manifest.json")
	19	+	m, err := readManifest(manifestPath)
	20	+	if err != nil {
	21	+	t.Fatal(err)
	22	+	}
	23	+	selected := m.Scenarios[0]
	24	+	h := defaultHarness()
	25	+	h.clock, h.launcher = &fakeClock{}, &fakeLauncher{}
	26	+	out := filepath.Join(t.TempDir(), "out")
	27	+	if err := run([]string{"--vev-bin", "ignored", "--manifest", manifestPath, "--out", out, "--scenario", selected.ID, "--warmup", "0s", "--duration", "30s", "--repetitions", "10"}, h); err != nil {
	28	+	t.Fatal(err)
	29	+	}
	30	+	var runs []runResult
	31	+	b, err := os.ReadFile(filepath.Join(out, "runs.json"))
	32	+	if err != nil {
	33	+	t.Fatal(err)
	34	+	}
	35	+	if err := json.Unmarshal(b, &runs); err != nil {
	36	+	t.Fatal(err)
	37	+	}
	38	+	if len(runs) != 10 {
	39	+	t.Fatalf("selected canonical runs=%d, want 10", len(runs))
	40	+	}
	41	+	for _, result := range runs {
	42	+	if result.Scenario != selected.ID {
	43	+	t.Fatalf("selected scenario = %q, want %q", result.Scenario, selected.ID)
	44	+	}
	45	+	}
	46	+	}
	47	+
	48	+	func TestHarnessScenarioSelectionRejectsIncompleteCanonicalManifest(t *testing.T) {
	49	+	m, err := readManifest(filepath.Join("..", "..", "testdata", "perf", "manifest.json"))
	50	+	if err != nil {
	51	+	t.Fatal(err)
	52	+	}
	53	+	selected := m.Scenarios[0]
	54	+	m.Scenarios = m.Scenarios[:len(m.Scenarios)-1]
	55	+	dir := t.TempDir()
	56	+	manifestPath := filepath.Join(dir, "manifest.json")
	57	+	if err := os.WriteFile(manifestPath, []byte(mustJSON(m)), 0o600); err != nil {
	58	+	t.Fatal(err)
	59	+	}
	60	+	launcher := &fakeLauncher{}
	61	+	h := defaultHarness()
	62	+	h.clock, h.launcher = &fakeClock{}, launcher
	63	+	out := filepath.Join(dir, "out")
	64	+	err = run([]string{"--vev-bin", "ignored", "--manifest", manifestPath, "--out", out, "--scenario", selected.ID, "--warmup", "0s", "--duration", "30s", "--repetitions", "10"}, h)
	65	+	if err == nil \|\| !strings.Contains(err.Error(), "missing scenario or inapplicable reason") {
	66	+	t.Fatalf("incomplete manifest selection error = %v", err)
	67	+	}
	68	+	if len(launcher.mappings) != 0 {
	69	+	t.Fatalf("launched with incomplete canonical manifest: %+v", launcher.mappings)
	70	+	}
	71	+	if _, statErr := os.Stat(out); !errors.Is(statErr, os.ErrNotExist) {
	72	+	t.Fatalf("output created before manifest validation: %v", statErr)
	73	+	}
	74	+	}
	75	+
	76	+	func TestResolvePathOptionsMakesRelativeAndAbsoluteInputsIdentical(t *testing.T) {
	77	+	cwd, err := os.Getwd()
	78	+	if err != nil {
	79	+	t.Fatal(err)
	80	+	}
	81	+	absolute := options{
	82	+	vevBin: filepath.Join(cwd, "vev"),
	83	+	manifest: filepath.Join(cwd, "testdata", "perf", "manifest.json"),
	84	+	out: filepath.Join(cwd, "testdata", "perf", "results"),
	85	+	}
	86	+	relative := options{
	87	+	vevBin: "vev",
	88	+	manifest: "testdata/perf/manifest.json",
	89	+	out: "testdata/perf/results",
	90	+	}
	91	+	gotRelative, err := resolvePathOptions(relative)
	92	+	if err != nil {
	93	+	t.Fatal(err)
	94	+	}
	95	+	gotAbsolute, err := resolvePathOptions(absolute)
	96	+	if err != nil {
	97	+	t.Fatal(err)
	98	+	}
	99	+	if gotRelative != gotAbsolute {
	100	+	t.Fatalf("relative paths resolved to %+v, absolute paths to %+v", gotRelative, gotAbsolute)
	101	+	}
	102	+	}
	103	+
	104	+	func TestHarnessFlagsRejectShortMeasurements(t *testing.T) {
	105	+	for _, args := range [][]string{
	106	+	{"--vev-bin", "vev", "--manifest", "m", "--out", "o", "--duration", "29s", "--repetitions", "10"},
	107	+	{"--vev-bin", "vev", "--manifest", "m", "--out", "o", "--duration", "30s", "--repetitions", "9"},
	108	+	{"--manifest", "m", "--out", "o", "--duration", "30s", "--repetitions", "10"},
	109	+	} {
	110	+	if _, err := parseOptions(args); err == nil {
	111	+	t.Fatalf("parseOptions(%v) succeeded", args)
	112	+	}
	113	+	}
	114	+	if _, err := parseOptions([]string{"--vev-bin", "vev", "--manifest", "m", "--out", "o", "--duration", "30s", "--repetitions", "10"}); err == nil {
	115	+	t.Fatal("missing --warmup accepted")
	116	+	}
	117	+	if _, err := parseOptions([]string{"--vev-bin", "vev", "--manifest", "m", "--out", "o", "--warmup", "1s", "--duration", "30s", "--repetitions", "10"}); err != nil {
	118	+	t.Fatal(err)
	119	+	}
	120	+	}
	121	+
	122	+	func TestValidateManifestRejectsEmptyAndDuplicateMatrixIDs(t *testing.T) {
	123	+	canonical, err := readManifest(filepath.Join("..", "..", "testdata", "perf", "manifest.json"))
	124	+	if err != nil {
	125	+	t.Fatal(err)
	126	+	}
	127	+	for _, tc := range []struct {
	128	+	name string
	129	+	mutate func(*manifest)
	130	+	want string
	131	+	}{
	132	+	{"empty topology", func(m *manifest) { m.Topologies[0].ID = "" }, "empty topology id"},
	133	+	{"duplicate topology", func(m *manifest) { m.Topologies[1].ID = m.Topologies[0].ID }, "duplicate topology id"},
	134	+	{"empty workload", func(m *manifest) { m.Workloads[0] = "" }, "empty workload id"},
	135	+	{"duplicate workload", func(m *manifest) { m.Workloads[1] = m.Workloads[0] }, "duplicate workload id"},
	136	+	{"empty transport", func(m *manifest) { m.Transports[0].ID = "" }, "empty transport id"},
	137	+	{"duplicate transport", func(m *manifest) { m.Transports[1].ID = m.Transports[0].ID }, "duplicate transport id"},
	138	+	} {
	139	+	t.Run(tc.name, func(t *testing.T) {
	140	+	m := canonical
	141	+	m.Topologies = append([]topology(nil), canonical.Topologies...)
	142	+	m.Workloads = append([]string(nil), canonical.Workloads...)
	143	+	m.Transports = append([]transport(nil), canonical.Transports...)
	144	+	tc.mutate(&m)
	145	+	if err := validateManifest(m); err == nil \|\| !strings.Contains(err.Error(), tc.want) {
	146	+	t.Fatalf("validateManifest() error = %v, want %q", err, tc.want)
	147	+	}
	148	+	})
	149	+	}
	150	+	}
	151	+
	152	+	func TestValidateManifestRejectsFullyCoveredUnknownMatrixIDs(t *testing.T) {
	153	+	canonical, err := readManifest(filepath.Join("..", "..", "testdata", "perf", "manifest.json"))
	154	+	if err != nil {
	155	+	t.Fatal(err)
	156	+	}
	157	+	for _, tc := range []struct {
	158	+	name string
	159	+	mutate func(*manifest)
	160	+	want string
	161	+	}{
	162	+	{
	163	+	name: "topology",
	164	+	mutate: func(m *manifest) {
	165	+	m.Topologies = append(m.Topologies, topology{ID: "9x9", Geometry: "120x40", RowsPerPane: 10000})
	166	+	for _, s := range canonical.Scenarios {
	167	+	s.ID = "9x9-" + s.ID
	168	+	s.Topology = "9x9"
	169	+	m.Scenarios = append(m.Scenarios, s)
	170	+	}
	171	+	},
	172	+	want: "unknown topology id \"9x9\"",
	173	+	},
	174	+	{
	175	+	name: "workload",
	176	+	mutate: func(m *manifest) {
	177	+	m.Workloads = append(m.Workloads, "unknown_workload")
	178	+	for _, s := range canonical.Scenarios {
	179	+	s.ID = "unknown-workload-" + s.ID
	180	+	s.Workload = "unknown_workload"
	181	+	m.Scenarios = append(m.Scenarios, s)
	182	+	}
	183	+	},
	184	+	want: "unknown workload id \"unknown_workload\"",
	185	+	},
	186	+	} {
	187	+	t.Run(tc.name, func(t *testing.T) {
	188	+	m := canonical
	189	+	m.Topologies = append([]topology(nil), canonical.Topologies...)
	190	+	m.Workloads = append([]string(nil), canonical.Workloads...)
	191	+	m.Scenarios = append([]scenario(nil), canonical.Scenarios...)
	192	+	tc.mutate(&m)
	193	+	if err := validateManifest(m); err == nil \|\| !strings.Contains(err.Error(), tc.want) {
	194	+	t.Fatalf("validateManifest() error = %v, want %q", err, tc.want)
	195	+	}
	196	+	})
	197	+	}
	198	+	}
	199	+
	200	+	func TestHarnessManifestCoversCanonicalMatrix(t *testing.T) {
	201	+	m, err := readManifest(filepath.Join("..", "..", "testdata", "perf", "manifest.json"))
	202	+	if err != nil {
	203	+	t.Fatal(err)
	204	+	}
	205	+	if err := validateManifest(m); err != nil {
	206	+	t.Fatal(err)
	207	+	}
	208	+	if len(m.Scenarios) != 497 {
	209	+	t.Fatalf("scenarios=%d", len(m.Scenarios))
	210	+	}
	211	+	m.Scenarios = m.Scenarios[:len(m.Scenarios)-1]
	212	+	if err := validateManifest(m); err == nil {
	213	+	t.Fatal("missing combination accepted")
	214	+	}
	215	+	}
	216	+
	217	+	func TestHarnessCreatesExclusiveTraceManifestAndEvidence(t *testing.T) {
	218	+	dir := t.TempDir()
	219	+	raw, err := os.OpenFile(filepath.Join(dir, "raw-harness.jsonl"), os.O_CREATE\|os.O_EXCL\|os.O_WRONLY, 0o600)
	220	+	if err != nil {
	221	+	t.Fatal(err)
	222	+	}
	223	+	closeTestFile(t, raw)
	224	+	launcher := &fakeLauncher{}
	225	+	h := defaultHarness()
	226	+	h.clock = &fakeClock{}
	227	+	h.launcher = launcher
	228	+	o := options{vevBin: "ignored", out: dir, warmup: time.Second, duration: minimumDuration, repetitions: minimumRepetitions}
	229	+	r, err := h.runOne(o, manifest{}, scenario{ID: "4x4-local", Roles: []string{"daemon", "client"}}, 1, raw)
	230	+	if err != nil {
	231	+	t.Fatal(err)
	232	+	}
	233	+	if !launcher.manifestPresent {
	234	+	t.Fatal("launched before complete run manifest")
	235	+	}
	236	+	if len(r.EndToEnd) < minimumInIntervalEventSamples \|\| r.Samples < minimumInIntervalEventSamples {
	237	+	t.Fatalf("samples=%+v", r)
	238	+	}
	239	+	seen := map[string]bool{}
	240	+	for _, m := range r.Processes {
	241	+	if m.ProcessID == "" \|\| m.ClockDomain != m.ProcessID \|\| m.TracePath == "" \|\| seen[m.TracePath] {
	242	+	t.Fatalf("bad mapping %+v", m)
	243	+	}
	244	+	seen[m.TracePath] = true
	245	+	if _, err := os.Stat(m.TracePath); err != nil {
	246	+	t.Fatal(err)
	247	+	}
	248	+	}
	249	+	b, err := os.ReadFile(filepath.Join(dir, "4x4-local-run-001", "manifest.json"))
	250	+	if err != nil {
	251	+	t.Fatal(err)
	252	+	}
	253	+	var got runManifest
	254	+	if err := json.Unmarshal(b, &got); err != nil {
	255	+	t.Fatal(err)
	256	+	}
	257	+	if len(got.Processes) != 2 {
	258	+	t.Fatalf("manifest=%+v", got)
	259	+	}
	260	+	if len(launcher.process) != 2 \|\| !launcher.process[0].closed \|\| len(launcher.process[0].warmups) != 0 \|\| len(launcher.process[1].warmups) != 1 \|\| len(launcher.process[1].measures) < minimumInIntervalEventSamples {
	261	+	t.Fatalf("process lifecycle not recorded: %+v", launcher.process)
	262	+	}
	263	+	}
	264	+
	265	+	func TestHarnessRejectsMissingManifestRoleTrace(t *testing.T) {
	266	+	dir := t.TempDir()
	267	+	missing := filepath.Join(dir, "daemon.jsonl")
	268	+	client := filepath.Join(dir, "client.jsonl")
	269	+	if err := os.WriteFile(client, nil, 0o600); err != nil {
	270	+	t.Fatal(err)
	271	+	}
	272	+	_, err := mergeProcessTraces([]processMapping{
	273	+	{ProcessID: "daemon", ClockDomain: "daemon", TracePath: missing, Role: "daemon", Scenario: "s", Run: 1},
	274	+	{ProcessID: "client", ClockDomain: "client", TracePath: client, Role: "client", Scenario: "s", Run: 1},
	275	+	})
	276	+	if !errors.Is(err, os.ErrNotExist) \|\| !strings.Contains(err.Error(), "daemon role") {
	277	+	t.Fatalf("missing daemon trace error = %v, want daemon role not-exist error", err)
	278	+	}
	279	+	}

A cmd/vev-perf-harness/runner.go +439 -0

@@ -0,0 +1,439 @@
	1	+	package main
	2	+
	3	+	import (
	4	+	"context"
	5	+	"encoding/json"
	6	+	"errors"
	7	+	"fmt"
	8	+	"io"
	9	+	"os"
	10	+	"os/exec"
	11	+	"path/filepath"
	12	+	"runtime/debug"
	13	+	"sort"
	14	+	"strings"
	15	+	"time"
	16	+
	17	+	"github.com/bnema/vev/pkg/safedir"
	18	+	)
	19	+
	20	+	type harnessMark struct {
	21	+	Scenario string `json:"scenario"`
	22	+	Run int `json:"run"`
	23	+	Sequence uint64 `json:"sequence"`
	24	+	Kind string `json:"kind"`
	25	+	Tick int64 `json:"tick"`
	26	+	Valid bool `json:"valid"`
	27	+	}
	28	+
	29	+	func (c systemClock) Now() int64 { return time.Since(c.start).Nanoseconds() }
	30	+
	31	+	func (systemClock) Sleep(d time.Duration) { time.Sleep(d) }
	32	+
	33	+	func newPTYLocalEcho(input []byte) *ptyLocalEcho {
	34	+	expected := make([]byte, 0, len(input))
	35	+	for _, b := range input {
	36	+	if b == '\n' {
	37	+	expected = append(expected, '\r')
	38	+	}
	39	+	expected = append(expected, b)
	40	+	}
	41	+	return &ptyLocalEcho{expected: expected}
	42	+	}
	43	+
	44	+	type harness struct {
	45	+	clock clock
	46	+	launcher launcher
	47	+	mkdir func(string) error
	48	+	createRunDir func(string) error
	49	+	create func(string) (*os.File, error)
	50	+	removeAll func(string) error
	51	+	gitSHA func() string
	52	+	selectScenarios func(manifest) []scenario // test-only bounded fixture seam
	53	+	}
	54	+
	55	+	// createExclusiveRunDir prevents stale runtime, socket, state, or trace files
	56	+	// from an earlier invocation being mistaken for the next repetition's evidence.
	57	+	func createExclusiveRunDir(path string) error {
	58	+	if err := os.Mkdir(path, 0o700); err != nil {
	59	+	return err
	60	+	}
	61	+	if err := safedir.EnsurePrivate(path); err != nil {
	62	+	return errors.Join(err, os.Remove(path))
	63	+	}
	64	+	return nil
	65	+	}
	66	+
	67	+	// recordedGitSHA prefers build provenance so an installed public harness does
	68	+	// not depend on its launch directory being a checkout. The git fallback keeps
	69	+	// go run evidence attributable as well.
	70	+	const gitSHACommandTimeout = time.Second
	71	+
	72	+	func recordedGitSHA() string {
	73	+	if info, ok := debug.ReadBuildInfo(); ok {
	74	+	for _, setting := range info.Settings {
	75	+	if setting.Key == "vcs.revision" && setting.Value != "" {
	76	+	return setting.Value
	77	+	}
	78	+	}
	79	+	}
	80	+	ctx, cancel := context.WithTimeout(context.Background(), gitSHACommandTimeout)
	81	+	defer cancel()
	82	+	out, err := exec.CommandContext(ctx, "git", "rev-parse", "HEAD").Output()
	83	+	if err == nil && strings.TrimSpace(string(out)) != "" {
	84	+	return strings.TrimSpace(string(out))
	85	+	}
	86	+	return "unknown"
	87	+	}
	88	+
	89	+	func run(args []string, h *harness) (err error) {
	90	+	opt, err := parseOptions(args)
	91	+	if err != nil {
	92	+	return err
	93	+	}
	94	+	opt, err = resolvePathOptions(opt)
	95	+	if err != nil {
	96	+	return err
	97	+	}
	98	+	m, err := readManifest(opt.manifest)
	99	+	if err != nil {
	100	+	return err
	101	+	}
	102	+	if err := validateManifest(m); err != nil {
	103	+	return err
	104	+	}
	105	+	if err = h.mkdir(opt.out); err != nil {
	106	+	return err
	107	+	}
	108	+	raw, err := os.OpenFile(filepath.Join(opt.out, "raw-harness.jsonl"), os.O_WRONLY\|os.O_CREATE\|os.O_EXCL, 0o600)
	109	+	if err != nil {
	110	+	return err
	111	+	}
	112	+	defer func() {
	113	+	if closeErr := raw.Close(); closeErr != nil {
	114	+	err = errors.Join(err, closeErr)
	115	+	}
	116	+	}()
	117	+	results := make([]runResult, 0, len(m.Scenarios)*opt.repetitions)
	118	+	var localSpans []span
	119	+	scenarios := m.Scenarios
	120	+	if opt.scenario != "" {
	121	+	scenarios, err = selectScenario(m, opt.scenario)
	122	+	if err != nil {
	123	+	return err
	124	+	}
	125	+	}
	126	+	if h.selectScenarios != nil {
	127	+	scenarios = h.selectScenarios(m)
	128	+	}
	129	+	for _, s := range scenarios {
	130	+	if s.InapplicableReason != "" {
	131	+	continue
	132	+	}
	133	+	for r := 1; r <= opt.repetitions; r++ {
	134	+	rr, err := h.runOne(opt, m, s, r, raw)
	135	+	if err != nil {
	136	+	return fmt.Errorf("scenario %s run %d: %w", s.ID, r, err)
	137	+	}
	138	+	results = append(results, rr)
	139	+	localSpans = append(localSpans, rr.Spans...)
	140	+	}
	141	+	}
	142	+	all, cadence, maxGap, runP50s := aggregateEventResults(results)
	143	+	if len(all) == 0 {
	144	+	return errors.New("no measured harness input/flush pairs")
	145	+	}
	146	+	if err := writeJSON(filepath.Join(opt.out, "runs.json"), results); err != nil {
	147	+	return err
	148	+	}
	149	+	return writeJSON(filepath.Join(opt.out, "summary.json"), summary{Schema: 1, GitSHA: h.gitSHA(), Warmup: opt.warmup.String(), Duration: opt.duration.String(), Repetitions: opt.repetitions, EndToEnd: percentiles(all), Cadence: percentiles(cadence), MaxGap: maxGap, RunP50Dispersion: percentiles(runP50s), Spans: summarizeSpans(localSpans), Runs: len(results)})
	150	+	}
	151	+
	152	+	// aggregateEventResults deliberately preserves raw event latencies for
	153	+	// end-to-end percentiles. Per-run p50s are a separate dispersion series.
	154	+	func aggregateEventResults(results []runResult) (all, cadence []int64, maxGap int64, runP50s []int64) {
	155	+	for _, result := range results {
	156	+	all = append(all, result.EndToEnd...)
	157	+	cadence = append(cadence, result.CadenceSamples...)
	158	+	runP50s = append(runP50s, result.Event.P50)
	159	+	if result.MaxGap > maxGap {
	160	+	maxGap = result.MaxGap
	161	+	}
	162	+	}
	163	+	return all, cadence, maxGap, runP50s
	164	+	}
	165	+
	166	+	func (h *harness) runOne(o options, mat manifest, s scenario, run int, raw io.Writer) (res runResult, err error) {
	167	+	dir := filepath.Join(o.out, fmt.Sprintf("%s-run-%03d", safeName(s.ID), run))
	168	+	createRunDir := h.createRunDir
	169	+	if createRunDir == nil {
	170	+	createRunDir = createExclusiveRunDir
	171	+	}
	172	+	if err = createRunDir(dir); err != nil {
	173	+	return res, fmt.Errorf("create isolated run directory: %w", err)
	174	+	}
	175	+	success := false
	176	+	defer func() {
	177	+	if !success {
	178	+	if cleanupErr := h.removeAll(dir); cleanupErr != nil {
	179	+	err = errors.Join(err, fmt.Errorf("remove failed run directory: %w", cleanupErr))
	180	+	}
	181	+	}
	182	+	}()
	183	+	maps := make([]processMapping, 0, len(s.Roles))
	184	+	for i, role := range s.Roles {
	185	+	pid := fmt.Sprintf("%s-r%03d-%s-%02d", safeName(s.ID), run, safeName(role), i+1)
	186	+	path := filepath.Join(dir, pid+".jsonl")
	187	+	f, e := h.create(path)
	188	+	if e != nil {
	189	+	return res, fmt.Errorf("allocate exclusive trace %s: %w", role, e)
	190	+	}
	191	+	if e = f.Close(); e != nil {
	192	+	return res, e
	193	+	}
	194	+	maps = append(maps, processMapping{ProcessID: pid, ClockDomain: pid, TracePath: path, Role: role, Scenario: s.ID, Run: run, Identity: "harness-assigned:" + pid})
	195	+	}
	196	+	// Persist the complete mapping before any process can be launched. This is
	197	+	// both crash evidence and the only legal source for later ID correlation.
	198	+	if err = writeJSON(filepath.Join(dir, "manifest.json"), runManifest{Scenario: s.ID, Run: run, Processes: maps}); err != nil {
	199	+	return res, err
	200	+	}
	201	+	if h.launcher == nil {
	202	+	h.launcher = &cliLauncher{bin: o.vevBin}
	203	+	}
	204	+	processes := make([]launchedRole, 0, len(maps))
	205	+	closed := false
	206	+	closeRoles := func() error {
	207	+	if closed {
	208	+	return nil
	209	+	}
	210	+	closed = true
	211	+	// Clients own SSH-stdio descendants; close them before the deferred UDP
	212	+	// peer and daemon so no role survives a failed or completed run.
	213	+	return closeLaunchedRoles(processes)
	214	+	}
	215	+	defer func() {
	216	+	if cleanupErr := closeRoles(); cleanupErr != nil {
	217	+	if err == nil {
	218	+	err = cleanupErr
	219	+	} else {
	220	+	// Keep the primary workload failure first while retaining every
	221	+	// cleanup failure in reverse launch order.
	222	+	err = errors.Join(err, cleanupErr)
	223	+	}
	224	+	}
	225	+	}()
	226	+	selectedTransport, e := manifestTransport(mat, s.Transport)
	227	+	if e != nil {
	228	+	return res, e
	229	+	}
	230	+	for _, pm := range launchOrder(maps) {
	231	+	p, e := h.launcher.Launch(pm, routeRoleArgs(s, pm, selectedTransport))
	232	+	if e != nil {
	233	+	return res, e
	234	+	}
	235	+	processes = append(processes, launchedRole{mapping: pm, process: p})
	236	+	if pm.Role == "daemon" {
	237	+	if ready, ok := p.(interface{ WaitReady() error }); ok {
	238	+	if err = ready.WaitReady(); err != nil {
	239	+	return res, err
	240	+	}
	241	+	}
	242	+	}
	243	+	}
	244	+	marks := []harnessMark{}
	245	+	recordMark := func(mark harnessMark) error {
	246	+	marks = append(marks, mark)
	247	+	if err := json.NewEncoder(raw).Encode(mark); err != nil {
	248	+	return fmt.Errorf("write raw harness mark: %w", err)
	249	+	}
	250	+	return nil
	251	+	}
	252	+	seq := uint64(0)
	253	+	for _, p := range processes {
	254	+	if p.mapping.Role != "client" {
	255	+	continue
	256	+	}
	257	+	seq++
	258	+	sequence := seq
	259	+	if err = recordMark(harnessMark{Scenario: s.ID, Run: run, Sequence: sequence, Kind: "input_injected", Tick: h.clock.Now(), Valid: true}); err != nil {
	260	+	return res, err
	261	+	}
	262	+	warmupErr := p.process.Warmup(workloadInput(s, run, "warmup"))
	263	+	if markErr := recordMark(harnessMark{Scenario: s.ID, Run: run, Sequence: sequence, Kind: "terminal_flushed", Tick: h.clock.Now(), Valid: warmupErr == nil}); markErr != nil {
	264	+	return res, errors.Join(warmupErr, markErr)
	265	+	}
	266	+	if warmupErr != nil {
	267	+	return res, warmupErr
	268	+	}
	269	+	}
	270	+	// The harness clock, rather than a process clock, owns the explicit warmup
	271	+	// and measured intervals. Raw evidence retains every boundary, while only
	272	+	// complete valid pairs within [Start, End) contribute measurement samples.
	273	+	h.clock.Sleep(o.warmup)
	274	+	interval := measuredInterval{Start: h.clock.Now()}
	275	+	interval.End = interval.Start + o.duration.Nanoseconds()
	276	+	nextInjection := interval.Start
	277	+	for {
	278	+	now := h.clock.Now()
	279	+	if now < nextInjection {
	280	+	h.clock.Sleep(time.Duration(nextInjection - now))
	281	+	now = h.clock.Now()
	282	+	}
	283	+	if now >= interval.End {
	284	+	break
	285	+	}
	286	+	for _, p := range processes {
	287	+	if p.mapping.Role != "client" {
	288	+	continue
	289	+	}
	290	+	seq++
	291	+	sequence := seq
	292	+	injectedRecorded := false
	293	+	flushedRecorded := false
	294	+	injected := func() error {
	295	+	if markErr := recordMark(harnessMark{Scenario: s.ID, Run: run, Sequence: sequence, Kind: "input_injected", Tick: h.clock.Now(), Valid: true}); markErr != nil {
	296	+	return markErr
	297	+	}
	298	+	injectedRecorded = true
	299	+	return nil
	300	+	}
	301	+	flushed := func() error {
	302	+	if markErr := recordMark(harnessMark{Scenario: s.ID, Run: run, Sequence: sequence, Kind: "terminal_flushed", Tick: h.clock.Now(), Valid: true}); markErr != nil {
	303	+	return markErr
	304	+	}
	305	+	flushedRecorded = true
	306	+	return nil
	307	+	}
	308	+	if measureErr := p.process.Measure(workloadInput(s, run, fmt.Sprintf("measured-%d", sequence)), injected, flushed); measureErr != nil {
	309	+	if injectedRecorded && !flushedRecorded {
	310	+	diagnosticErr := recordMark(harnessMark{Scenario: s.ID, Run: run, Sequence: sequence, Kind: "terminal_flushed", Tick: h.clock.Now(), Valid: false})
	311	+	return res, errors.Join(measureErr, diagnosticErr)
	312	+	}
	313	+	return res, measureErr
	314	+	}
	315	+	}
	316	+	// Cap injection rate rather than bursting events after a slow terminal
	317	+	// flush. Any resulting spacing is retained as raw cadence evidence.
	318	+	nextInjection = h.clock.Now() + measuredEventCadence.Nanoseconds()
	319	+	}
	320	+	if now := h.clock.Now(); now < interval.End {
	321	+	h.clock.Sleep(time.Duration(interval.End - now))
	322	+	}
	323	+	events, e := measuredEventSamples(marks, interval)
	324	+	if e != nil {
	325	+	return res, e
	326	+	}
	327	+	if err := requireMinimumEventSamples(events); err != nil {
	328	+	return res, err
	329	+	}
	330	+	samples := eventLatencies(events)
	331	+	cadenceSamples, maxGap := eventCadenceSamples(events)
	332	+	// Process teardown can unblock a receive and cause its failed end mark to
	333	+	// be serialized. Merge only the complete, post-cleanup per-process traces.
	334	+	if cleanupErr := closeRoles(); cleanupErr != nil {
	335	+	return res, cleanupErr
	336	+	}
	337	+	spans, e := mergeProcessTraces(maps)
	338	+	if e != nil {
	339	+	return res, e
	340	+	}
	341	+	success = true
	342	+	return runResult{Spans: spans, Scenario: s.ID, Run: run, Samples: len(samples), EndToEnd: samples, Event: percentiles(samples), Cadence: percentiles(cadenceSamples), CadenceSamples: cadenceSamples, MaxGap: maxGap, Processes: maps}, nil
	343	+	}
	344	+
	345	+	func roleArgs(s scenario, m processMapping) roleCommand {
	346	+	return routeRoleArgs(s, m, scenarioTransport(s))
	347	+	}
	348	+
	349	+	func routeRoleArgs(s scenario, m processMapping, selected transport) roleCommand {
	350	+	session := "perf-" + safeName(s.ID) + fmt.Sprintf("-%03d", m.Run)
	351	+	remote := []string{"attach", "harness@127.0.0.1:" + session}
	352	+	switch m.Role {
	353	+	case "daemon":
	354	+	return roleCommand{Args: []string{"--daemon"}}
	355	+	case "client":
	356	+	switch s.Transport {
	357	+	case "local":
	358	+	return roleCommand{Args: []string{"new", session}}
	359	+	case "ssh_stdio", "udp_baseline", "udp_25ms", "udp_100ms", "udp_loss_0pct", "udp_loss_1pct":
	360	+	return roleCommand{Args: remote}
	361	+	}
	362	+	case "ssh_stdio_peer":
	363	+	return roleCommand{Args: []string{"_stdio", session}, Transport: transport{ID: s.Transport, Kind: "ssh_stdio"}}
	364	+	case "udp_peer":
	365	+	return roleCommand{Args: []string{"_udp-proxy", session}, Transport: selected}
	366	+	}
	367	+	return roleCommand{}
	368	+	}
	369	+
	370	+	func manifestTransport(m manifest, id string) (transport, error) {
	371	+	for _, t := range m.Transports {
	372	+	if t.ID == id {
	373	+	return t, nil
	374	+	}
	375	+	}
	376	+	// Focused unit tests may call runOne with a minimal manifest; production
	377	+	// already validates every scenario reference before this point.
	378	+	if m.Schema == 0 {
	379	+	return scenarioTransport(scenario{Transport: id}), nil
	380	+	}
	381	+	return transport{}, fmt.Errorf("scenario references missing transport %q", id)
	382	+	}
	383	+
	384	+	func scenarioTransport(s scenario) transport {
	385	+	switch s.Transport {
	386	+	case "udp_baseline":
	387	+	return transport{ID: s.Transport, Kind: "udp"}
	388	+	case "udp_25ms":
	389	+	return transport{ID: s.Transport, Kind: "udp", RTTMS: 25}
	390	+	case "udp_100ms":
	391	+	return transport{ID: s.Transport, Kind: "udp", RTTMS: 100}
	392	+	case "udp_loss_0pct":
	393	+	return transport{ID: s.Transport, Kind: "udp", LossPercent: 0}
	394	+	case "udp_loss_1pct":
	395	+	return transport{ID: s.Transport, Kind: "udp", LossPercent: 1}
	396	+	default:
	397	+	return transport{ID: s.Transport}
	398	+	}
	399	+	}
	400	+
	401	+	func launchOrder(mappings []processMapping) []processMapping {
	402	+	out := append([]processMapping(nil), mappings...)
	403	+	priority := func(role string) int {
	404	+	switch role {
	405	+	case "daemon":
	406	+	return 0
	407	+	case "ssh_stdio_peer", "udp_peer":
	408	+	return 1
	409	+	case "client":
	410	+	return 2
	411	+	default:
	412	+	return 3
	413	+	}
	414	+	}
	415	+	sort.SliceStable(out, func(i, j int) bool { return priority(out[i].Role) < priority(out[j].Role) })
	416	+	return out
	417	+	}
	418	+
	419	+	// workloadInput is shell input delivered through the client PTY. It avoids any
	420	+	// private daemon API and makes every topology/workload/transport manifest row
	421	+	// execute a real terminal workload. The marker is the output boundary awaited
	422	+	// by cliProcess; it is not a synthetic timestamp.
	423	+	func workloadInput(s scenario, run int, phase string) []byte {
	424	+	marker := fmt.Sprintf("__VEV_HARNESS_%s_r%d_%s__", safeName(s.ID), run, safeName(phase))
	425	+	body := "printf 'vev perf %s\\n'"
	426	+	switch s.Workload {
	427	+	case "active_output", "all_output", "inactive_output", "interactive_flood":
	428	+	body = "i=0; while [ $i -lt 128 ]; do printf 'vev perf output %s\\n'; i=$((i+1)); done"
	429	+	case "resize_sweep":
	430	+	body = "printf 'vev perf resize-sweep %s\\n'"
	431	+	case "copy_search":
	432	+	body = "printf 'vev perf copy-search %s\\n'"
	433	+	case "snapshot_output_resize":
	434	+	body = "printf 'vev perf snapshot-output-resize %s\\n'"
	435	+	case "attach_restore_tab_switch":
	436	+	body = "printf 'vev perf attach-restore-tab-switch %s\\n'"
	437	+	}
	438	+	return []byte(fmt.Sprintf(body+"; printf '%s\\n'\n", s.ID, marker))
	439	+	}

A cmd/vev-perf-harness/runner_test.go +458 -0

@@ -0,0 +1,458 @@
	1	+	package main
	2	+
	3	+	import (
	4	+	"bytes"
	5	+	"encoding/json"
	6	+	"errors"
	7	+	"fmt"
	8	+	"os"
	9	+	"os/exec"
	10	+	"path/filepath"
	11	+	"strings"
	12	+	"testing"
	13	+	"time"
	14	+
	15	+	"github.com/bnema/vev/pkg/safedir"
	16	+	)
	17	+
	18	+	func TestHarnessCanonicalLocalRolesAreIsolatedAcrossRepetitions(t *testing.T) {
	19	+	if testing.Short() {
	20	+	t.Skip("public CLI repetition lifecycle")
	21	+	}
	22	+	root, err := filepath.Abs(filepath.Join("..", ".."))
	23	+	if err != nil {
	24	+	t.Fatal(err)
	25	+	}
	26	+	bin := filepath.Join(t.TempDir(), "vev")
	27	+	build := exec.Command("go", "build", "-o", bin, "./")
	28	+	build.Dir = root
	29	+	if output, err := build.CombinedOutput(); err != nil {
	30	+	t.Fatalf("build public CLI: %v\n%s", err, output)
	31	+	}
	32	+	m, err := readManifest(filepath.Join(root, "testdata", "perf", "manifest.json"))
	33	+	if err != nil {
	34	+	t.Fatal(err)
	35	+	}
	36	+	var local scenario
	37	+	for _, candidate := range m.Scenarios {
	38	+	if candidate.ID == "1x4-idle-local" {
	39	+	local = candidate
	40	+	break
	41	+	}
	42	+	}
	43	+	if local.ID == "" {
	44	+	t.Fatal("canonical local scenario missing")
	45	+	}
	46	+	out, err := os.MkdirTemp("", "vev-ob-")
	47	+	if err != nil {
	48	+	t.Fatal(err)
	49	+	}
	50	+	removeTestTree(t, out)
	51	+	raw, err := os.OpenFile(filepath.Join(out, "raw.jsonl"), os.O_CREATE\|os.O_EXCL\|os.O_WRONLY, 0o600)
	52	+	if err != nil {
	53	+	t.Fatal(err)
	54	+	}
	55	+	closeTestFile(t, raw)
	56	+	h := defaultHarness()
	57	+	h.clock = &fakeClock{}
	58	+	h.launcher = &cliLauncher{bin: bin}
	59	+	for run := 1; run <= 3; run++ {
	60	+	result, err := h.runOne(options{vevBin: bin, out: out, duration: minimumDuration}, m, local, run, raw)
	61	+	if err != nil {
	62	+	t.Fatalf("run %d: %v", run, err)
	63	+	}
	64	+	if len(result.Processes) != 2 {
	65	+	t.Fatalf("run %d process mappings = %+v", run, result.Processes)
	66	+	}
	67	+	runDir := filepath.Join(out, fmt.Sprintf("%s-run-%03d", safeName(local.ID), run))
	68	+	manifestBytes, err := os.ReadFile(filepath.Join(runDir, "manifest.json"))
	69	+	if err != nil {
	70	+	t.Fatalf("run %d manifest: %v", run, err)
	71	+	}
	72	+	var persisted runManifest
	73	+	if err := json.Unmarshal(manifestBytes, &persisted); err != nil {
	74	+	t.Fatalf("run %d manifest JSON: %v", run, err)
	75	+	}
	76	+	if persisted.Run != run \|\| len(persisted.Processes) != len(result.Processes) {
	77	+	t.Fatalf("run %d manifest = %+v", run, persisted)
	78	+	}
	79	+	for _, process := range result.Processes {
	80	+	if process.Role != "daemon" && process.Role != "client" {
	81	+	t.Fatalf("run %d unexpected role %q", run, process.Role)
	82	+	}
	83	+	if filepath.Dir(process.TracePath) != runDir {
	84	+	t.Fatalf("run %d %s trace outside its isolated run directory: %q", run, process.Role, process.TracePath)
	85	+	}
	86	+	trace, err := os.ReadFile(process.TracePath)
	87	+	if err != nil {
	88	+	t.Fatalf("run %d %s trace %q: %v", run, process.Role, process.TracePath, err)
	89	+	}
	90	+	if len(bytes.TrimSpace(trace)) == 0 {
	91	+	t.Fatalf("run %d %s did not emit its declared trace", run, process.Role)
	92	+	}
	93	+	for _, line := range bytes.Split(bytes.TrimSpace(trace), []byte{'\n'}) {
	94	+	var record traceRecord
	95	+	if err := json.Unmarshal(line, &record); err != nil {
	96	+	t.Fatalf("run %d %s trace JSON: %v", run, process.Role, err)
	97	+	}
	98	+	if record.ProcessID != process.ProcessID \|\| record.Scenario != local.ID \|\| record.Run != uint64(run) {
	99	+	t.Fatalf("run %d %s trace escaped its manifest identity: %+v", run, process.Role, record)
	100	+	}
	101	+	}
	102	+	}
	103	+	if _, err := os.Stat(filepath.Join(runDir, "runtime", "vev", "daemon.sock")); !errors.Is(err, os.ErrNotExist) {
	104	+	t.Fatalf("run %d left daemon socket for the next repetition: %v", run, err)
	105	+	}
	106	+	}
	107	+	}
	108	+
	109	+	func TestHarnessResolvesCanonicalRelativePathsBeforeRoleCWDChanges(t *testing.T) {
	110	+	if testing.Short() {
	111	+	t.Skip("public CLI relative-path lifecycle")
	112	+	}
	113	+	root, err := filepath.Abs(filepath.Join("..", ".."))
	114	+	if err != nil {
	115	+	t.Fatal(err)
	116	+	}
	117	+	// Keep the real Unix-socket path below its platform limit while testing the
	118	+	// exact documented relative output hierarchy.
	119	+	workspace, err := os.MkdirTemp("", "vev-rel-")
	120	+	if err != nil {
	121	+	t.Fatal(err)
	122	+	}
	123	+	removeTestTree(t, workspace)
	124	+	bin := filepath.Join(workspace, "vev")
	125	+	build := exec.Command("go", "build", "-o", bin, "./")
	126	+	build.Dir = root
	127	+	if output, err := build.CombinedOutput(); err != nil {
	128	+	t.Fatalf("build public CLI: %v\n%s", err, output)
	129	+	}
	130	+	manifestPath := filepath.Join(workspace, "testdata", "perf", "manifest.json")
	131	+	if err := safedir.EnsurePrivate(filepath.Dir(manifestPath)); err != nil {
	132	+	t.Fatal(err)
	133	+	}
	134	+	contents, err := os.ReadFile(filepath.Join(root, "testdata", "perf", "manifest.json"))
	135	+	if err != nil {
	136	+	t.Fatal(err)
	137	+	}
	138	+	if err := os.WriteFile(manifestPath, contents, 0o600); err != nil {
	139	+	t.Fatal(err)
	140	+	}
	141	+	originalWD, err := os.Getwd()
	142	+	if err != nil {
	143	+	t.Fatal(err)
	144	+	}
	145	+	if err := os.Chdir(workspace); err != nil {
	146	+	t.Fatal(err)
	147	+	}
	148	+	t.Cleanup(func() {
	149	+	if err := os.Chdir(originalWD); err != nil {
	150	+	t.Errorf("restore cwd: %v", err)
	151	+	}
	152	+	})
	153	+
	154	+	// These are the documented relative locations. cliLauncher changes every
	155	+	// role to its run directory, so a failure to resolve them before launch
	156	+	// would make VEV_PERF_TRACE point at a nonexistent nested directory.
	157	+	h := defaultHarness()
	158	+	h.clock = &fakeClock{}
	159	+	if err := run([]string{
	160	+	"--vev-bin", "./vev",
	161	+	"--manifest", "testdata/perf/manifest.json",
	162	+	"--out", "testdata/perf/results",
	163	+	"--scenario", "1x4-idle-local",
	164	+	"--warmup", "0s",
	165	+	"--duration", "30s",
	166	+	"--repetitions", "10",
	167	+	}, h); err != nil {
	168	+	t.Fatal(err)
	169	+	}
	170	+
	171	+	out := filepath.Join(workspace, "testdata", "perf", "results")
	172	+	for run := 1; run <= 10; run++ {
	173	+	runDir := filepath.Join(out, fmt.Sprintf("1x4-idle-local-run-%03d", run))
	174	+	manifestBytes, err := os.ReadFile(filepath.Join(runDir, "manifest.json"))
	175	+	if err != nil {
	176	+	t.Fatalf("run %d manifest: %v", run, err)
	177	+	}
	178	+	var persisted runManifest
	179	+	if err := json.Unmarshal(manifestBytes, &persisted); err != nil {
	180	+	t.Fatalf("run %d manifest JSON: %v", run, err)
	181	+	}
	182	+	if persisted.Run != run \|\| len(persisted.Processes) != 2 {
	183	+	t.Fatalf("run %d manifest=%+v", run, persisted)
	184	+	}
	185	+	for _, process := range persisted.Processes {
	186	+	if !filepath.IsAbs(process.TracePath) \|\| filepath.Dir(process.TracePath) != runDir {
	187	+	t.Fatalf("run %d %s trace path=%q, want absolute path in %q", run, process.Role, process.TracePath, runDir)
	188	+	}
	189	+	trace, err := os.ReadFile(process.TracePath)
	190	+	if err != nil \|\| len(bytes.TrimSpace(trace)) == 0 {
	191	+	t.Fatalf("run %d %s trace=%q err=%v", run, process.Role, process.TracePath, err)
	192	+	}
	193	+	}
	194	+	}
	195	+	}
	196	+
	197	+	func TestHarnessRejectsExistingRunDirectoryBeforeLaunchingRoles(t *testing.T) {
	198	+	out := t.TempDir()
	199	+	raw, err := os.OpenFile(filepath.Join(out, "raw.jsonl"), os.O_CREATE\|os.O_EXCL\|os.O_WRONLY, 0o600)
	200	+	if err != nil {
	201	+	t.Fatal(err)
	202	+	}
	203	+	closeTestFile(t, raw)
	204	+	s := scenario{ID: "isolated", Roles: []string{"daemon", "client"}}
	205	+	stale := filepath.Join(out, "isolated-run-001")
	206	+	if err := os.Mkdir(stale, 0o700); err != nil {
	207	+	t.Fatal(err)
	208	+	}
	209	+	if err := os.WriteFile(filepath.Join(stale, "preserve"), []byte("stale"), 0o600); err != nil {
	210	+	t.Fatal(err)
	211	+	}
	212	+	launcher := &fakeLauncher{}
	213	+	h := defaultHarness()
	214	+	h.clock, h.launcher = &fakeClock{}, launcher
	215	+	_, err = h.runOne(options{out: out, duration: minimumDuration}, manifest{}, s, 1, raw)
	216	+	if err == nil \|\| !strings.Contains(err.Error(), "create isolated run directory") {
	217	+	t.Fatalf("existing run directory error = %v", err)
	218	+	}
	219	+	if len(launcher.mappings) != 0 {
	220	+	t.Fatalf("roles launched against stale run directory: %+v", launcher.mappings)
	221	+	}
	222	+	if contents, readErr := os.ReadFile(filepath.Join(stale, "preserve")); readErr != nil \|\| string(contents) != "stale" {
	223	+	t.Fatalf("stale directory changed: contents=%q err=%v", contents, readErr)
	224	+	}
	225	+	}
	226	+
	227	+	func TestHarnessDoesNotInheritNestedSession(t *testing.T) {
	228	+	env := withoutEnv([]string{"VEV=session=old", "PATH=/bin", "VEV_PERF_TRACE=old"}, "VEV")
	229	+	if !equalStrings(env, []string{"PATH=/bin", "VEV_PERF_TRACE=old"}) {
	230	+	t.Fatalf("environment=%q", env)
	231	+	}
	232	+	}
	233	+
	234	+	func TestHarnessRoutesEveryRemoteFixtureThroughItsDeclaredPeer(t *testing.T) {
	235	+	cases := []struct {
	236	+	name string
	237	+	transport transport
	238	+	peer string
	239	+	wantRTT int
	240	+	wantLoss int
	241	+	}{
	242	+	{"ssh", transport{ID: "ssh_stdio", Kind: "ssh_stdio"}, "ssh_stdio_peer", 0, 0},
	243	+	{"udp baseline", transport{ID: "udp_baseline", Kind: "udp"}, "udp_peer", 0, 0},
	244	+	{"udp 25ms", transport{ID: "udp_25ms", Kind: "udp", RTTMS: 25}, "udp_peer", 25, 0},
	245	+	{"udp 100ms", transport{ID: "udp_100ms", Kind: "udp", RTTMS: 100}, "udp_peer", 100, 0},
	246	+	{"udp loss zero", transport{ID: "udp_loss_0pct", Kind: "udp", LossPercent: 0}, "udp_peer", 0, 0},
	247	+	{"udp loss", transport{ID: "udp_loss_1pct", Kind: "udp", LossPercent: 1}, "udp_peer", 0, 1},
	248	+	}
	249	+	for _, tc := range cases {
	250	+	t.Run(tc.name, func(t *testing.T) {
	251	+	s := scenario{ID: "route", Transport: tc.transport.ID}
	252	+	client := routeRoleArgs(s, processMapping{Role: "client", Run: 2}, tc.transport)
	253	+	peer := routeRoleArgs(s, processMapping{Role: tc.peer, Run: 2}, tc.transport)
	254	+	if !equalStrings(client.Args, []string{"attach", "harness@127.0.0.1:perf-route-002"}) {
	255	+	t.Fatalf("client did not use public remote attach: %q", client.Args)
	256	+	}
	257	+	if tc.peer == "ssh_stdio_peer" && !equalStrings(peer.Args, []string{"_stdio", "perf-route-002"}) {
	258	+	t.Fatalf("ssh peer command=%q", peer.Args)
	259	+	}
	260	+	if tc.peer == "udp_peer" && !equalStrings(peer.Args, []string{"_udp-proxy", "perf-route-002"}) {
	261	+	t.Fatalf("udp peer command=%q", peer.Args)
	262	+	}
	263	+	if peer.Transport.RTTMS != tc.wantRTT \|\| peer.Transport.LossPercent != tc.wantLoss {
	264	+	t.Fatalf("peer lost manifest network settings: %+v", peer.Transport)
	265	+	}
	266	+	})
	267	+	}
	268	+	}
	269	+
	270	+	func TestHarnessFakeRunnerRoutesClientToPeerAndCleansEveryRole(t *testing.T) {
	271	+	for _, tc := range []struct {
	272	+	name, transport, peer, peerCommand string
	273	+	}{
	274	+	{"ssh", "ssh_stdio", "ssh_stdio_peer", "_stdio"},
	275	+	{"udp baseline", "udp_baseline", "udp_peer", "_udp-proxy"},
	276	+	{"udp 25ms", "udp_25ms", "udp_peer", "_udp-proxy"},
	277	+	{"udp 100ms", "udp_100ms", "udp_peer", "_udp-proxy"},
	278	+	{"udp loss zero", "udp_loss_0pct", "udp_peer", "_udp-proxy"},
	279	+	{"udp loss", "udp_loss_1pct", "udp_peer", "_udp-proxy"},
	280	+	} {
	281	+	t.Run(tc.name, func(t *testing.T) {
	282	+	dir := t.TempDir()
	283	+	raw, err := os.OpenFile(filepath.Join(dir, "raw.jsonl"), os.O_CREATE\|os.O_EXCL\|os.O_WRONLY, 0o600)
	284	+	if err != nil {
	285	+	t.Fatal(err)
	286	+	}
	287	+	closeTestFile(t, raw)
	288	+	l := &fakeLauncher{}
	289	+	h := defaultHarness()
	290	+	h.clock, h.launcher = &fakeClock{}, l
	291	+	s := scenario{ID: "routed", Transport: tc.transport, Roles: []string{"daemon", "client", tc.peer}}
	292	+	if _, err := h.runOne(options{out: dir, warmup: time.Second, duration: minimumDuration, repetitions: minimumRepetitions}, manifest{}, s, 1, raw); err != nil {
	293	+	t.Fatal(err)
	294	+	}
	295	+	if len(l.mappings) != 3 \|\| l.mappings[1].Role != tc.peer \|\| l.mappings[2].Role != "client" {
	296	+	t.Fatalf("dependency launch order=%+v", l.mappings)
	297	+	}
	298	+	if l.commands[1].Args[0] != tc.peerCommand \|\| l.commands[2].Args[0] != "attach" {
	299	+	t.Fatalf("commands do not connect client through declared peer: %+v", l.commands)
	300	+	}
	301	+	for i, p := range l.process {
	302	+	if !p.closed {
	303	+	t.Fatalf("role %s was not cleaned up", l.mappings[i].Role)
	304	+	}
	305	+	}
	306	+	})
	307	+	}
	308	+	}
	309	+
	310	+	func TestHarnessAggregatesRawEventsSeparatelyFromRunDispersion(t *testing.T) {
	311	+	first := append([]int64(nil), make([]int64, minimumInIntervalEventSamples+5)...)
	312	+	for i := range first {
	313	+	first[i] = 1
	314	+	}
	315	+	second := []int64{100, 100, 100, 100, 100, 100, 100, 100, 100, 100, 100, 100, 100, 1_000, 1_000}
	316	+	all, _, _, runP50s := aggregateEventResults([]runResult{
	317	+	{EndToEnd: first, Event: distribution{P50: 1}},
	318	+	{EndToEnd: second, Event: distribution{P50: 100}},
	319	+	})
	320	+	if got, dispersion := percentiles(all), percentiles(runP50s); got.Count != 30 \|\| got.P99 != 1_000 \|\| dispersion.Count != 2 \|\| dispersion.P99 != 1 {
	321	+	t.Fatalf("event aggregation=%+v run dispersion=%+v", got, dispersion)
	322	+	}
	323	+	}
	324	+
	325	+	func TestHarnessRawRecordsContainEveryMarkWhileSamplesRemainFiltered(t *testing.T) {
	326	+	dir := t.TempDir()
	327	+	raw, err := os.OpenFile(filepath.Join(dir, "raw.jsonl"), os.O_CREATE\|os.O_EXCL\|os.O_WRONLY, 0o600)
	328	+	if err != nil {
	329	+	t.Fatal(err)
	330	+	}
	331	+	h := defaultHarness()
	332	+	h.clock, h.launcher = &fakeClock{}, &fakeLauncher{}
	333	+	result, err := h.runOne(options{out: dir, warmup: time.Second, duration: minimumDuration}, manifest{}, scenario{ID: "raw-complete", Roles: []string{"daemon", "client"}}, 1, raw)
	334	+	if err != nil {
	335	+	t.Fatal(err)
	336	+	}
	337	+	if err := raw.Close(); err != nil {
	338	+	t.Fatal(err)
	339	+	}
	340	+	marks := readHarnessMarks(t, filepath.Join(dir, "raw.jsonl"))
	341	+	if got, want := len(marks), 2*(result.Samples+1); got != want {
	342	+	t.Fatalf("raw marks=%d, want warmup plus every measured boundary=%d", got, want)
	343	+	}
	344	+	for i, mark := range marks[:2] {
	345	+	if mark.Sequence != 1 \|\| !mark.Valid \|\| []string{"input_injected", "terminal_flushed"}[i] != mark.Kind {
	346	+	t.Fatalf("warmup mark %d=%+v", i, mark)
	347	+	}
	348	+	}
	349	+	for sequence := uint64(2); sequence <= uint64(result.Samples+1); sequence++ {
	350	+	if marks[2(sequence-1)].Sequence != sequence \|\| marks[2(sequence-1)+1].Sequence != sequence {
	351	+	t.Fatalf("raw marks did not retain both boundaries for sequence %d: %+v", sequence, marks)
	352	+	}
	353	+	}
	354	+	}
	355	+
	356	+	func TestHarnessPersistsFailedFlushDiagnostics(t *testing.T) {
	357	+	dir := t.TempDir()
	358	+	raw, err := os.OpenFile(filepath.Join(dir, "raw.jsonl"), os.O_CREATE\|os.O_EXCL\|os.O_WRONLY, 0o600)
	359	+	if err != nil {
	360	+	t.Fatal(err)
	361	+	}
	362	+	failure := errors.New("flush failed")
	363	+	h := defaultHarness()
	364	+	h.clock, h.launcher = &fakeClock{}, failedFlushLauncher{err: failure}
	365	+	_, err = h.runOne(options{out: dir, warmup: time.Second, duration: minimumDuration}, manifest{}, scenario{ID: "failed-flush", Roles: []string{"daemon", "client"}}, 1, raw)
	366	+	if !errors.Is(err, failure) {
	367	+	t.Fatalf("run error=%v, want failed flush", err)
	368	+	}
	369	+	if err := raw.Close(); err != nil {
	370	+	t.Fatal(err)
	371	+	}
	372	+	marks := readHarnessMarks(t, filepath.Join(dir, "raw.jsonl"))
	373	+	if got, want := len(marks), 4; got != want {
	374	+	t.Fatalf("raw marks=%d, want warmup and failed measured pair=%d: %+v", got, want, marks)
	375	+	}
	376	+	failed := marks[len(marks)-1]
	377	+	if failed.Sequence != 2 \|\| failed.Kind != "terminal_flushed" \|\| failed.Valid {
	378	+	t.Fatalf("failed flush diagnostic=%+v", failed)
	379	+	}
	380	+	}
	381	+
	382	+	func TestHarnessSchedulesMeasuredEventsAcrossFullInterval(t *testing.T) {
	383	+	dir := t.TempDir()
	384	+	raw, err := os.OpenFile(filepath.Join(dir, "raw.jsonl"), os.O_CREATE\|os.O_EXCL\|os.O_WRONLY, 0o600)
	385	+	if err != nil {
	386	+	t.Fatal(err)
	387	+	}
	388	+	closeTestFile(t, raw)
	389	+	h := defaultHarness()
	390	+	h.clock, h.launcher = &fakeClock{}, &fakeLauncher{}
	391	+	result, err := h.runOne(options{out: dir, warmup: time.Second, duration: minimumDuration}, manifest{}, scenario{ID: "schedule", Roles: []string{"daemon", "client"}}, 1, raw)
	392	+	if err != nil {
	393	+	t.Fatal(err)
	394	+	}
	395	+	if result.Samples < minimumInIntervalEventSamples \|\| result.Cadence.Count != result.Samples-1 \|\| result.MaxGap <= 0 {
	396	+	t.Fatalf("result did not include repeated measured events: %+v", result)
	397	+	}
	398	+	if err := raw.Close(); err != nil {
	399	+	t.Fatal(err)
	400	+	}
	401	+	records, err := readJSONL(filepath.Join(dir, "raw.jsonl"))
	402	+	if err != nil {
	403	+	t.Fatal(err)
	404	+	}
	405	+	if want := result.Samples*2 + 2; len(records) != want {
	406	+	t.Fatalf("raw records=%d, want %d warmup and measured boundaries", len(records), want)
	407	+	}
	408	+	}
	409	+
	410	+	func TestHarnessPreservesWorkloadAndCleanupErrors(t *testing.T) {
	411	+	dir := t.TempDir()
	412	+	raw, err := os.OpenFile(filepath.Join(dir, "raw.jsonl"), os.O_CREATE\|os.O_EXCL\|os.O_WRONLY, 0o600)
	413	+	if err != nil {
	414	+	t.Fatal(err)
	415	+	}
	416	+	closeTestFile(t, raw)
	417	+	primary := errors.New("primary workload failure")
	418	+	cleanup := errors.New("daemon cleanup failure")
	419	+	launcher := &fakeLauncher{
	420	+	measureErr: map[string]error{"client": primary},
	421	+	closeErr: map[string]error{"daemon": cleanup},
	422	+	}
	423	+	h := defaultHarness()
	424	+	h.clock, h.launcher = &fakeClock{}, launcher
	425	+	_, err = h.runOne(options{out: dir, warmup: time.Second, duration: minimumDuration, repetitions: minimumRepetitions}, manifest{}, scenario{ID: "errors", Roles: []string{"daemon", "client"}}, 1, raw)
	426	+	if !errors.Is(err, primary) \|\| !errors.Is(err, cleanup) {
	427	+	t.Fatalf("workload and cleanup errors were not both preserved: %v", err)
	428	+	}
	429	+	if got, want := err.Error(), primary.Error()+"\nclose daemon role: "+cleanup.Error(); got != want {
	430	+	t.Fatalf("error ordering=%q want %q", got, want)
	431	+	}
	432	+	for i, p := range launcher.process {
	433	+	if !p.closed {
	434	+	t.Fatalf("role %s was not closed", launcher.mappings[i].Role)
	435	+	}
	436	+	}
	437	+	}
	438	+
	439	+	func TestHarnessCleansRunDirectoryOnProcessFailure(t *testing.T) {
	440	+	dir := t.TempDir()
	441	+	raw, err := os.OpenFile(filepath.Join(dir, "raw-harness.jsonl"), os.O_CREATE\|os.O_EXCL\|os.O_WRONLY, 0o600)
	442	+	if err != nil {
	443	+	t.Fatal(err)
	444	+	}
	445	+	closeTestFile(t, raw)
	446	+	launcher := &fakeLauncher{}
	447	+	h := defaultHarness()
	448	+	h.clock = &fakeClock{}
	449	+	// The client is the second role and fails after the daemon has started.
	450	+	h.launcher = &failingLauncher{fakeLauncher: launcher}
	451	+	o := options{vevBin: "ignored", out: dir, warmup: time.Second, duration: minimumDuration, repetitions: minimumRepetitions}
	452	+	if _, err := h.runOne(o, manifest{}, scenario{ID: "cleanup", Roles: []string{"daemon", "client"}}, 1, raw); err == nil {
	453	+	t.Fatal("failed process accepted")
	454	+	}
	455	+	if _, err := os.Stat(filepath.Join(dir, "cleanup-run-001")); !os.IsNotExist(err) {
	456	+	t.Fatalf("failed run directory remains: %v", err)
	457	+	}
	458	+	}

A cmd/vev-perf-harness/test_helpers_test.go +341 -0

@@ -0,0 +1,341 @@
	1	+	package main
	2	+
	3	+	import (
	4	+	"bytes"
	5	+	"encoding/json"
	6	+	"errors"
	7	+	"fmt"
	8	+	"os"
	9	+	"path/filepath"
	10	+	"testing"
	11	+	"time"
	12	+	)
	13	+
	14	+	type fakeClock struct{ tick int64 }
	15	+
	16	+	func (c *fakeClock) Now() int64 { c.tick += 10; return c.tick }
	17	+
	18	+	func (c *fakeClock) Sleep(d time.Duration) {
	19	+	c.tick += d.Nanoseconds()
	20	+	}
	21	+
	22	+	type fakeUDPNetem struct {
	23	+	port int
	24	+	closed bool
	25	+	}
	26	+
	27	+	func (n *fakeUDPNetem) Port() int { return n.port }
	28	+
	29	+	func (n *fakeUDPNetem) Close() error { n.closed = true; return nil }
	30	+
	31	+	type fakeProcess struct {
	32	+	fail bool
	33	+	measureErr error
	34	+	closeErr error
	35	+	warmups [][]byte
	36	+	measures [][]byte
	37	+	closed bool
	38	+	}
	39	+
	40	+	func (p *fakeProcess) Warmup(input []byte) error {
	41	+	p.warmups = append(p.warmups, append([]byte(nil), input...))
	42	+	return nil
	43	+	}
	44	+
	45	+	func (p *fakeProcess) Measure(input []byte, injected, flush func() error) error {
	46	+	p.measures = append(p.measures, append([]byte(nil), input...))
	47	+	if p.measureErr != nil {
	48	+	return p.measureErr
	49	+	}
	50	+	if p.fail {
	51	+	return errors.New("measure failure")
	52	+	}
	53	+	if err := injected(); err != nil {
	54	+	return err
	55	+	}
	56	+	return flush()
	57	+	}
	58	+
	59	+	func (p *fakeProcess) Close() error { p.closed = true; return p.closeErr }
	60	+
	61	+	type fakeLauncher struct {
	62	+	mappings []processMapping
	63	+	args [][]string
	64	+	commands []roleCommand
	65	+	process []*fakeProcess
	66	+	measureErr map[string]error
	67	+	closeErr map[string]error
	68	+	manifestPresent bool
	69	+	}
	70	+
	71	+	func (l *fakeLauncher) Launch(m processMapping, command roleCommand) (launchedProcess, error) {
	72	+	l.mappings = append(l.mappings, m)
	73	+	l.args = append(l.args, append([]string(nil), command.Args...))
	74	+	l.commands = append(l.commands, command)
	75	+	_, e := os.Stat(filepath.Join(filepath.Dir(m.TracePath), "manifest.json"))
	76	+	l.manifestPresent = e == nil
	77	+	p := &fakeProcess{measureErr: l.measureErr[m.Role], closeErr: l.closeErr[m.Role]}
	78	+	l.process = append(l.process, p)
	79	+	return p, nil
	80	+	}
	81	+
	82	+	type lateFlushProcess struct{ clock *fakeClock }
	83	+
	84	+	func (*lateFlushProcess) Warmup([]byte) error { return nil }
	85	+
	86	+	func (p *lateFlushProcess) Measure(_ []byte, injected, flushed func() error) error {
	87	+	if err := injected(); err != nil {
	88	+	return err
	89	+	}
	90	+	p.clock.tick += minimumDuration.Nanoseconds()
	91	+	return flushed()
	92	+	}
	93	+
	94	+	func (*lateFlushProcess) Close() error { return nil }
	95	+
	96	+	type insufficientEventLauncher struct{ clock *fakeClock }
	97	+
	98	+	func (l insufficientEventLauncher) Launch(m processMapping, _ roleCommand) (launchedProcess, error) {
	99	+	if m.Role == "client" {
	100	+	return &lateFlushProcess{clock: l.clock}, nil
	101	+	}
	102	+	return &fakeProcess{}, nil
	103	+	}
	104	+
	105	+	type failedFlushProcess struct{ err error }
	106	+
	107	+	func (*failedFlushProcess) Warmup([]byte) error { return nil }
	108	+
	109	+	func (p *failedFlushProcess) Measure(_ []byte, injected, _ func() error) error {
	110	+	if err := injected(); err != nil {
	111	+	return err
	112	+	}
	113	+	return p.err
	114	+	}
	115	+
	116	+	func (*failedFlushProcess) Close() error { return nil }
	117	+
	118	+	type failedFlushLauncher struct{ err error }
	119	+
	120	+	func (l failedFlushLauncher) Launch(m processMapping, _ roleCommand) (launchedProcess, error) {
	121	+	if m.Role == "client" {
	122	+	return &failedFlushProcess{err: l.err}, nil
	123	+	}
	124	+	return &fakeProcess{}, nil
	125	+	}
	126	+
	127	+	func closeTestFile(t testing.T, f os.File) {
	128	+	t.Helper()
	129	+	t.Cleanup(func() {
	130	+	if err := f.Close(); err != nil && !errors.Is(err, os.ErrClosed) {
	131	+	t.Error(err)
	132	+	}
	133	+	})
	134	+	}
	135	+
	136	+	func removeTestTree(t *testing.T, path string) {
	137	+	t.Helper()
	138	+	t.Cleanup(func() {
	139	+	if err := os.RemoveAll(path); err != nil {
	140	+	t.Error(err)
	141	+	}
	142	+	})
	143	+	}
	144	+
	145	+	func readHarnessMarks(t *testing.T, path string) []harnessMark {
	146	+	t.Helper()
	147	+	records, err := readJSONL(path)
	148	+	if err != nil {
	149	+	t.Fatal(err)
	150	+	}
	151	+	marks := make([]harnessMark, len(records))
	152	+	for i, record := range records {
	153	+	encoded, err := json.Marshal(record)
	154	+	if err != nil {
	155	+	t.Fatal(err)
	156	+	}
	157	+	if err := json.Unmarshal(encoded, &marks[i]); err != nil {
	158	+	t.Fatal(err)
	159	+	}
	160	+	}
	161	+	return marks
	162	+	}
	163	+
	164	+	type fakePTY struct{ writes [][]byte }
	165	+
	166	+	func (p *fakePTY) Read([]byte) (int, error) { return 0, os.ErrClosed }
	167	+
	168	+	func (p *fakePTY) Write(b []byte) (int, error) {
	169	+	p.writes = append(p.writes, append([]byte(nil), b...))
	170	+	return len(b), nil
	171	+	}
	172	+
	173	+	func (*fakePTY) Close() error { return nil }
	174	+
	175	+	type stagedPTY struct {
	176	+	allow <-chan struct{}
	177	+	writes chan []byte
	178	+	}
	179	+
	180	+	func (*stagedPTY) Read([]byte) (int, error) { return 0, os.ErrClosed }
	181	+
	182	+	func (p *stagedPTY) Write(b []byte) (int, error) {
	183	+	if p.allow != nil {
	184	+	<-p.allow
	185	+	}
	186	+	p.writes <- append([]byte(nil), b...)
	187	+	return len(b), nil
	188	+	}
	189	+
	190	+	func (*stagedPTY) Close() error { return nil }
	191	+
	192	+	type fakeOutput struct {
	193	+	bytes.Buffer
	194	+	syncs int
	195	+	}
	196	+
	197	+	func (o *fakeOutput) Sync() error { o.syncs++; return nil }
	198	+
	199	+	func (*fakeOutput) Close() error { return nil }
	200	+
	201	+	type orderedPTY struct {
	202	+	write func([]byte)
	203	+	close func()
	204	+	}
	205	+
	206	+	func (*orderedPTY) Read([]byte) (int, error) { return 0, os.ErrClosed }
	207	+
	208	+	func (p *orderedPTY) Write(b []byte) (int, error) {
	209	+	if p.write != nil {
	210	+	p.write(b)
	211	+	}
	212	+	return len(b), nil
	213	+	}
	214	+
	215	+	func (p *orderedPTY) Close() error {
	216	+	p.close()
	217	+	return nil
	218	+	}
	219	+
	220	+	type orderedOutput struct{ close func() }
	221	+
	222	+	func (*orderedOutput) Write(b []byte) (int, error) { return len(b), nil }
	223	+
	224	+	func (*orderedOutput) Sync() error { return nil }
	225	+
	226	+	func (o *orderedOutput) Close() error {
	227	+	o.close()
	228	+	return nil
	229	+	}
	230	+
	231	+	type failingLauncher struct{ *fakeLauncher }
	232	+
	233	+	func (l *failingLauncher) Launch(m processMapping, command roleCommand) (launchedProcess, error) {
	234	+	p, err := l.fakeLauncher.Launch(m, command)
	235	+	if err != nil {
	236	+	return nil, err
	237	+	}
	238	+	if m.Role == "client" {
	239	+	fake, ok := p.(*fakeProcess)
	240	+	if !ok {
	241	+	return nil, fmt.Errorf("fake launcher returned %T, want *fakeProcess", p)
	242	+	}
	243	+	fake.fail = true
	244	+	}
	245	+	return p, nil
	246	+	}
	247	+
	248	+	func equalStrings(a, b []string) bool {
	249	+	if len(a) != len(b) {
	250	+	return false
	251	+	}
	252	+	for i := range a {
	253	+	if a[i] != b[i] {
	254	+	return false
	255	+	}
	256	+	}
	257	+	return true
	258	+	}
	259	+
	260	+	type traceLauncher struct{ fakeLauncher }
	261	+
	262	+	func (l *traceLauncher) Launch(m processMapping, command roleCommand) (launchedProcess, error) {
	263	+	p, err := l.fakeLauncher.Launch(m, command)
	264	+	if err != nil {
	265	+	return nil, err
	266	+	}
	267	+	f, err := os.OpenFile(m.TracePath, os.O_WRONLY\|os.O_APPEND, 0o600)
	268	+	if err != nil {
	269	+	return nil, err
	270	+	}
	271	+	for i, pair := range spanPairs {
	272	+	for n, kind := range []string{pair.start, pair.end} {
	273	+	b, err := json.Marshal(traceRecord{ProcessID: m.ProcessID, Component: m.Role, Scenario: m.Scenario, Run: uint64(m.Run), Sequence: uint64(i + 1), RequestID: 1, Epoch: 1, Kind: kind, Tick: int64(i20 + n5), Valid: true})
	274	+	if err != nil {
	275	+	return nil, errors.Join(err, f.Close())
	276	+	}
	277	+	if _, err := f.Write(append(b, '\n')); err != nil {
	278	+	return nil, errors.Join(err, f.Close())
	279	+	}
	280	+	}
	281	+	}
	282	+	if err := f.Close(); err != nil {
	283	+	return nil, err
	284	+	}
	285	+	return p, nil
	286	+	}
	287	+
	288	+	type closeTraceProcess struct {
	289	+	*fakeProcess
	290	+	traceEnd func() error
	291	+	}
	292	+
	293	+	func (p *closeTraceProcess) Close() error {
	294	+	p.closed = true
	295	+	return errors.Join(p.closeErr, p.traceEnd())
	296	+	}
	297	+
	298	+	type closeTraceLauncher struct{ fakeLauncher }
	299	+
	300	+	func (l *closeTraceLauncher) Launch(m processMapping, command roleCommand) (launchedProcess, error) {
	301	+	p, err := l.fakeLauncher.Launch(m, command)
	302	+	if err != nil {
	303	+	return nil, err
	304	+	}
	305	+	start := traceRecord{ProcessID: m.ProcessID, Component: m.Role, Scenario: m.Scenario, Run: uint64(m.Run), Sequence: 1, RequestID: 1, Epoch: 1, Kind: "adapter_receive_start", Tick: 10, Valid: true}
	306	+	if err := appendTraceRecord(m.TracePath, start); err != nil {
	307	+	return nil, err
	308	+	}
	309	+	fake, ok := p.(*fakeProcess)
	310	+	if !ok {
	311	+	return nil, fmt.Errorf("fake launcher returned %T, want *fakeProcess", p)
	312	+	}
	313	+	return &closeTraceProcess{fakeProcess: fake, traceEnd: func() error {
	314	+	return appendTraceRecord(m.TracePath, traceRecord{ProcessID: m.ProcessID, Component: m.Role, Scenario: m.Scenario, Run: uint64(m.Run), Sequence: 1, RequestID: 1, Epoch: 1, Kind: "adapter_receive_end", Tick: 20, Valid: true})
	315	+	}}, nil
	316	+	}
	317	+
	318	+	func appendTraceRecord(path string, record traceRecord) error {
	319	+	f, err := os.OpenFile(path, os.O_WRONLY\|os.O_APPEND, 0o600)
	320	+	if err != nil {
	321	+	return err
	322	+	}
	323	+	b, err := json.Marshal(record)
	324	+	if err != nil {
	325	+	return errors.Join(err, f.Close())
	326	+	}
	327	+	if _, err := f.Write(append(b, '\n')); err != nil {
	328	+	return errors.Join(err, f.Close())
	329	+	}
	330	+	return f.Close()
	331	+	}
	332	+
	333	+	// boundedLocalConcurrentTrace is an excerpt from the second run of the
	334	+	// bounded local public-CLI smoke. Receive sequence 3 is in flight while the
	335	+	// same process serializes sequence 4; shared-process records therefore cannot
	336	+	// have a global sequence ordering requirement.
	337	+	const boundedLocalConcurrentTrace = `{"schema":1,"process_id":"1x4-idle-local-r002-client-02","component":"ipc","scenario":"1x4-idle-local","run":2,"sequence":3,"request_id":3,"epoch":3,"kind":"adapter_receive_start","tick":1783902621495284782,"bytes":0,"fragments":0,"retransmits":0,"pending":0,"ack_rtt_nanos":0,"valid":true}
	338	+	{"schema":1,"process_id":"1x4-idle-local-r002-client-02","component":"ipc","scenario":"1x4-idle-local","run":2,"sequence":4,"request_id":4,"epoch":4,"kind":"adapter_send_start","tick":1783902621495364552,"bytes":96,"fragments":0,"retransmits":0,"pending":0,"ack_rtt_nanos":0,"valid":true}
	339	+	{"schema":1,"process_id":"1x4-idle-local-r002-client-02","component":"ipc","scenario":"1x4-idle-local","run":2,"sequence":4,"request_id":4,"epoch":4,"kind":"adapter_send_end","tick":1783902621495386452,"bytes":96,"fragments":0,"retransmits":0,"pending":0,"ack_rtt_nanos":0,"valid":true}
	340	+	{"schema":1,"process_id":"1x4-idle-local-r002-client-02","component":"ipc","scenario":"1x4-idle-local","run":2,"sequence":3,"request_id":3,"epoch":3,"kind":"adapter_receive_end","tick":1783902621495463332,"bytes":0,"fragments":0,"retransmits":0,"pending":0,"ack_rtt_nanos":0,"valid":true}
	341	+	`

A cmd/vev-perf-harness/udp_netem.go +273 -0

@@ -0,0 +1,273 @@
	1	+	package main
	2	+
	3	+	import (
	4	+	"container/heap"
	5	+	"context"
	6	+	"errors"
	7	+	"fmt"
	8	+	"net"
	9	+	"os"
	10	+	"strconv"
	11	+	"strings"
	12	+	"sync"
	13	+	"sync/atomic"
	14	+	"time"
	15	+	)
	16	+
	17	+	// udpNetem is a harness-owned, user-space UDP relay. It sits between the
	18	+	// public dgram client and vev's public _udp-proxy readiness address, so fixture
	19	+	// impairment is real network carriage rather than an ignored vev option.
	20	+	type udpNetem interface {
	21	+	Port() int
	22	+	Close() error
	23	+	}
	24	+
	25	+	type udpNetemFactory func(udpNetemConfig) (udpNetem, error)
	26	+
	27	+	type udpNetemConfig struct {
	28	+	// RTT is split equally across both relay directions.
	29	+	RTT time.Duration
	30	+	LossPercent int
	31	+	TargetPath string
	32	+	}
	33	+
	34	+	// udpNetemQueueCapacity bounds both packets waiting to be scheduled and
	35	+	// packets waiting for their delivery deadline. When it is full, the relay
	36	+	// drops the arriving packet and records it in queueOverflowDrops.
	37	+	const udpNetemQueueCapacity = 256
	38	+
	39	+	type scheduledPacket struct {
	40	+	packet []byte
	41	+	to net.Addr
	42	+	due time.Time
	43	+	}
	44	+
	45	+	type scheduledPackets []scheduledPacket
	46	+
	47	+	func (p scheduledPackets) Len() int { return len(p) }
	48	+	func (p scheduledPackets) Less(i, j int) bool { return p[i].due.Before(p[j].due) }
	49	+	func (p scheduledPackets) Swap(i, j int) { p[i], p[j] = p[j], p[i] }
	50	+	func (p *scheduledPackets) Push(x any) {
	51	+	packet, ok := x.(scheduledPacket)
	52	+	if !ok {
	53	+	panic(fmt.Sprintf("scheduled packet type %T", x))
	54	+	}
	55	+	p = append(p, packet)
	56	+	}
	57	+	func (p *scheduledPackets) Pop() any {
	58	+	old := *p
	59	+	n := len(old)
	60	+	item := old[n-1]
	61	+	*p = old[:n-1]
	62	+	return item
	63	+	}
	64	+
	65	+	type udpNetemRelay struct {
	66	+	config udpNetemConfig
	67	+	conn net.PacketConn
	68	+	done chan struct{}
	69	+	wg sync.WaitGroup
	70	+	once sync.Once
	71	+	closeErr error
	72	+
	73	+	// slots bounds queued and deadline-pending packets together. The reader and
	74	+	// the sole scheduler are the relay's only goroutines.
	75	+	slots chan struct{}
	76	+	queue chan scheduledPacket
	77	+	queueOverflowDrops atomic.Uint64
	78	+	deliveryErrors atomic.Uint64
	79	+
	80	+	mu sync.Mutex
	81	+	client, dst net.Addr
	82	+	packet uint64
	83	+	}
	84	+
	85	+	func newUDPNetem(config udpNetemConfig) (udpNetem, error) {
	86	+	if config.RTT < 0 \|\| config.LossPercent < 0 \|\| config.LossPercent > 100 \|\| config.TargetPath == "" {
	87	+	return nil, fmt.Errorf("invalid UDP netem fixture: rtt=%s loss=%d target=%q", config.RTT, config.LossPercent, config.TargetPath)
	88	+	}
	89	+	var lc net.ListenConfig
	90	+	conn, err := lc.ListenPacket(context.Background(), "udp", "127.0.0.1:0")
	91	+	if err != nil {
	92	+	return nil, err
	93	+	}
	94	+	r := &udpNetemRelay{
	95	+	config: config,
	96	+	conn: conn,
	97	+	done: make(chan struct{}),
	98	+	slots: make(chan struct{}, udpNetemQueueCapacity),
	99	+	queue: make(chan scheduledPacket, udpNetemQueueCapacity),
	100	+	}
	101	+	r.wg.Add(2)
	102	+	go r.run()
	103	+	go r.schedule()
	104	+	return r, nil
	105	+	}
	106	+
	107	+	func (r *udpNetemRelay) Port() int {
	108	+	addr, ok := r.conn.LocalAddr().(*net.UDPAddr)
	109	+	if !ok {
	110	+	return 0
	111	+	}
	112	+	return addr.Port
	113	+	}
	114	+
	115	+	func (r *udpNetemRelay) Close() error {
	116	+	r.once.Do(func() {
	117	+	close(r.done)
	118	+	r.closeErr = r.conn.Close()
	119	+	r.wg.Wait()
	120	+	if drops := r.queueOverflowDrops.Load(); drops != 0 {
	121	+	r.closeErr = errors.Join(r.closeErr, fmt.Errorf("udp netem queue overflow drops: %d", drops))
	122	+	}
	123	+	if writes := r.deliveryErrors.Load(); writes != 0 {
	124	+	r.closeErr = errors.Join(r.closeErr, fmt.Errorf("udp netem delivery errors: %d", writes))
	125	+	}
	126	+	})
	127	+	return r.closeErr
	128	+	}
	129	+
	130	+	func (r *udpNetemRelay) run() {
	131	+	defer r.wg.Done()
	132	+	buf := make([]byte, 64*1024)
	133	+	for {
	134	+	n, from, err := r.conn.ReadFrom(buf)
	135	+	if err != nil {
	136	+	return
	137	+	}
	138	+	to, ok := r.destination(from)
	139	+	if !ok \|\| r.drop() {
	140	+	continue
	141	+	}
	142	+	// Copy because PacketConn reuses buf on its next read.
	143	+	r.enqueue(scheduledPacket{packet: append([]byte(nil), buf[:n]...), to: to, due: time.Now().Add(r.config.RTT / 2)})
	144	+	}
	145	+	}
	146	+
	147	+	// enqueue implements the explicit overflow policy: never block the reader;
	148	+	// drop a packet if the bounded scheduler capacity has been reached.
	149	+	func (r *udpNetemRelay) enqueue(packet scheduledPacket) bool {
	150	+	select {
	151	+	case r.slots <- struct{}{}:
	152	+	case <-r.done:
	153	+	return false
	154	+	default:
	155	+	r.queueOverflowDrops.Add(1)
	156	+	return false
	157	+	}
	158	+	select {
	159	+	case r.queue <- packet:
	160	+	return true
	161	+	case <-r.done:
	162	+	<-r.slots
	163	+	return false
	164	+	}
	165	+	}
	166	+
	167	+	func (r *udpNetemRelay) schedule() {
	168	+	defer r.wg.Done()
	169	+	pending := scheduledPackets{}
	170	+	heap.Init(&pending)
	171	+	timer := time.NewTimer(time.Hour)
	172	+	if !timer.Stop() {
	173	+	<-timer.C
	174	+	}
	175	+	defer timer.Stop()
	176	+	var timerC <-chan time.Time
	177	+	resetTimer := func() {
	178	+	if timerC != nil && !timer.Stop() {
	179	+	select {
	180	+	case <-timer.C:
	181	+	default:
	182	+	}
	183	+	}
	184	+	if pending.Len() == 0 {
	185	+	timerC = nil
	186	+	return
	187	+	}
	188	+	delay := time.Until(pending[0].due)
	189	+	if delay < 0 {
	190	+	delay = 0
	191	+	}
	192	+	timer.Reset(delay)
	193	+	timerC = timer.C
	194	+	}
	195	+	for {
	196	+	resetTimer()
	197	+	select {
	198	+	case <-r.done:
	199	+	return
	200	+	case packet := <-r.queue:
	201	+	heap.Push(&pending, packet)
	202	+	case <-timerC:
	203	+	now := time.Now()
	204	+	for pending.Len() > 0 && !pending[0].due.After(now) {
	205	+	packet, ok := heap.Pop(&pending).(scheduledPacket)
	206	+	if !ok {
	207	+	panic("scheduled packet heap contained unexpected type")
	208	+	}
	209	+	select {
	210	+	case <-r.done:
	211	+	return
	212	+	default:
	213	+	if _, err := r.conn.WriteTo(packet.packet, packet.to); err != nil {
	214	+	select {
	215	+	case <-r.done:
	216	+	default:
	217	+	r.deliveryErrors.Add(1)
	218	+	}
	219	+	}
	220	+	<-r.slots
	221	+	}
	222	+	}
	223	+	}
	224	+	}
	225	+	}
	226	+
	227	+	func (r *udpNetemRelay) destination(from net.Addr) (net.Addr, bool) {
	228	+	r.mu.Lock()
	229	+	defer r.mu.Unlock()
	230	+	if r.dst == nil {
	231	+	dst, err := udpNetemTarget(r.config.TargetPath)
	232	+	if err != nil {
	233	+	return nil, false
	234	+	}
	235	+	r.dst = dst
	236	+	}
	237	+	if sameUDPAddress(from, r.dst) {
	238	+	return r.client, r.client != nil
	239	+	}
	240	+	r.client = from
	241	+	return r.dst, true
	242	+	}
	243	+
	244	+	func (r *udpNetemRelay) drop() bool {
	245	+	r.mu.Lock()
	246	+	defer r.mu.Unlock()
	247	+	r.packet++
	248	+	// The deterministic cycle makes the configured percentage auditable: a 1%
	249	+	// fixture drops exactly one carriage packet in every 100 attempts.
	250	+	return r.config.LossPercent > 0 && int((r.packet-1)%100) < r.config.LossPercent
	251	+	}
	252	+
	253	+	func udpNetemTarget(path string) (net.Addr, error) {
	254	+	line, err := os.ReadFile(path)
	255	+	if err != nil {
	256	+	return nil, err
	257	+	}
	258	+	fields := strings.Fields(string(line))
	259	+	if len(fields) != 3 \|\| fields[0] != "VEV-UDP" {
	260	+	return nil, errors.New("invalid UDP proxy readiness")
	261	+	}
	262	+	port, err := strconv.Atoi(fields[1])
	263	+	if err != nil \|\| port < 1 \|\| port > 65535 \|\| fields[2] == "" {
	264	+	return nil, errors.New("invalid UDP proxy readiness")
	265	+	}
	266	+	return &net.UDPAddr{IP: net.ParseIP("127.0.0.1"), Port: port}, nil
	267	+	}
	268	+
	269	+	func sameUDPAddress(a, b net.Addr) bool {
	270	+	ua, aOK := a.(*net.UDPAddr)
	271	+	ub, bOK := b.(*net.UDPAddr)
	272	+	return aOK && bOK && ua.Port == ub.Port && ua.IP.Equal(ub.IP)
	273	+	}

A cmd/vev-perf-harness/udp_netem_test.go +149 -0

@@ -0,0 +1,149 @@
	1	+	package main
	2	+
	3	+	import (
	4	+	"context"
	5	+	"net"
	6	+	"os"
	7	+	"path/filepath"
	8	+	"strconv"
	9	+	"testing"
	10	+	"time"
	11	+	)
	12	+
	13	+	func closeUDPFixture(t *testing.T, c interface{ Close() error }) {
	14	+	t.Helper()
	15	+	t.Cleanup(func() {
	16	+	if err := c.Close(); err != nil {
	17	+	t.Error(err)
	18	+	}
	19	+	})
	20	+	}
	21	+
	22	+	func TestUDPNetemDropsWhenBoundedSchedulerIsFull(t *testing.T) {
	23	+	r := &udpNetemRelay{
	24	+	done: make(chan struct{}),
	25	+	slots: make(chan struct{}, udpNetemQueueCapacity),
	26	+	queue: make(chan scheduledPacket, udpNetemQueueCapacity),
	27	+	}
	28	+	packet := scheduledPacket{packet: []byte("bounded"), to: &net.UDPAddr{IP: net.ParseIP("127.0.0.1"), Port: 1}}
	29	+	for range udpNetemQueueCapacity {
	30	+	if !r.enqueue(packet) {
	31	+	t.Fatal("enqueue() dropped packet before scheduler capacity was full")
	32	+	}
	33	+	}
	34	+	if r.enqueue(packet) {
	35	+	t.Fatal("enqueue() accepted packet beyond scheduler capacity")
	36	+	}
	37	+	if got := r.queueOverflowDrops.Load(); got != 1 {
	38	+	t.Fatalf("queue overflow drops=%d, want 1", got)
	39	+	}
	40	+	}
	41	+
	42	+	func TestUDPNetemCloseRejectsQueueOverflow(t *testing.T) {
	43	+	netem, err := newUDPNetem(udpNetemConfig{TargetPath: filepath.Join(t.TempDir(), "target")})
	44	+	if err != nil {
	45	+	t.Fatal(err)
	46	+	}
	47	+	relay, ok := netem.(*udpNetemRelay)
	48	+	if !ok {
	49	+	t.Fatalf("netem type=%T, want *udpNetemRelay", netem)
	50	+	}
	51	+	relay.queueOverflowDrops.Store(3)
	52	+	if err := netem.Close(); err == nil \|\| err.Error() != "udp netem queue overflow drops: 3" {
	53	+	t.Fatalf("Close error=%v, want overflow rejection", err)
	54	+	}
	55	+	}
	56	+
	57	+	func TestUDPNetemExecutesRTTAndLossFixtures(t *testing.T) {
	58	+	fixtures := []struct {
	59	+	name string
	60	+	rtt time.Duration
	61	+	loss int
	62	+	sent int
	63	+	want int
	64	+	}{
	65	+	{"25ms RTT", 25 * time.Millisecond, 0, 1, 1},
	66	+	{"100ms RTT", 100 * time.Millisecond, 0, 1, 1},
	67	+	{"zero loss", 0, 0, 10, 10},
	68	+	{"one percent loss", 0, 1, 100, 99},
	69	+	}
	70	+	for _, tc := range fixtures {
	71	+	t.Run(tc.name, func(t *testing.T) {
	72	+	var lc net.ListenConfig
	73	+	target, err := lc.ListenPacket(context.Background(), "udp", "127.0.0.1:0")
	74	+	if err != nil {
	75	+	t.Fatal(err)
	76	+	}
	77	+	closeUDPFixture(t, target)
	78	+	targetAddr, ok := target.LocalAddr().(*net.UDPAddr)
	79	+	if !ok {
	80	+	t.Fatalf("target address type=%T, want *net.UDPAddr", target.LocalAddr())
	81	+	}
	82	+	path := filepath.Join(t.TempDir(), "target")
	83	+	if err := os.WriteFile(path, []byte("VEV-UDP "+strconv.Itoa(targetAddr.Port)+" key\\n"), 0o600); err != nil {
	84	+	t.Fatal(err)
	85	+	}
	86	+	netem, err := newUDPNetem(udpNetemConfig{RTT: tc.rtt, LossPercent: tc.loss, TargetPath: path})
	87	+	if err != nil {
	88	+	t.Fatal(err)
	89	+	}
	90	+	closeUDPFixture(t, netem)
	91	+	client, err := lc.ListenPacket(context.Background(), "udp", "127.0.0.1:0")
	92	+	if err != nil {
	93	+	t.Fatal(err)
	94	+	}
	95	+	closeUDPFixture(t, client)
	96	+	started := time.Now()
	97	+	for i := 0; i < tc.sent; i++ {
	98	+	if _, err := client.WriteTo([]byte{byte(i)}, &net.UDPAddr{IP: net.ParseIP("127.0.0.1"), Port: netem.Port()}); err != nil {
	99	+	t.Fatal(err)
	100	+	}
	101	+	}
	102	+	if err := target.SetReadDeadline(time.Now().Add(2 * time.Second)); err != nil {
	103	+	t.Fatal(err)
	104	+	}
	105	+	got := 0
	106	+	buf := make([]byte, 32)
	107	+	var relayAddr net.Addr
	108	+	for got < tc.want {
	109	+	if _, relayAddr, err = target.ReadFrom(buf); err != nil {
	110	+	t.Fatalf("received %d packets, want %d: %v", got, tc.want, err)
	111	+	}
	112	+	got++
	113	+	}
	114	+	if tc.rtt > 0 {
	115	+	if elapsed := time.Since(started); elapsed < tc.rtt/2 {
	116	+	t.Fatalf("client-to-target netem delay=%s, want at least %s", elapsed, tc.rtt/2)
	117	+	}
	118	+	}
	119	+	if tc.loss == 0 {
	120	+	// The relay learns the client from the first carriage packet. Replying
	121	+	// through that same relay address verifies the target-to-client half
	122	+	// of the configured RTT as well.
	123	+	started = time.Now()
	124	+	if _, err := target.WriteTo([]byte("reply"), relayAddr); err != nil {
	125	+	t.Fatal(err)
	126	+	}
	127	+	if err := client.SetReadDeadline(time.Now().Add(2 * time.Second)); err != nil {
	128	+	t.Fatal(err)
	129	+	}
	130	+	if _, _, err := client.ReadFrom(buf); err != nil {
	131	+	t.Fatalf("relay reply: %v", err)
	132	+	}
	133	+	if tc.rtt > 0 {
	134	+	if elapsed := time.Since(started); elapsed < tc.rtt/2 {
	135	+	t.Fatalf("target-to-client netem delay=%s, want at least %s", elapsed, tc.rtt/2)
	136	+	}
	137	+	}
	138	+	}
	139	+	// Keep the deadline short: extra carriage would prove the exact 1%%
	140	+	// loss cycle was not enforced.
	141	+	if err := target.SetReadDeadline(time.Now().Add(20 * time.Millisecond)); err != nil {
	142	+	t.Fatal(err)
	143	+	}
	144	+	if _, _, err := target.ReadFrom(buf); tc.want != tc.sent && err == nil {
	145	+	t.Fatalf("received more than %d packets with %d%% loss", tc.want, tc.loss)
	146	+	}
	147	+	})
	148	+	}
	149	+	}

M docs/performance.md +63 -87

@@ -1,103 +1,79 @@
1	-	# Performance methodology (M5)	1	+	# Performance measurement
2	-		2	+
3	-	This baseline measures in-process daemon work only. It deliberately does not	3	+	## S5a transport observability
4	-	measure coordinator behavior, chunk encoding, compression, GC tuning, or real	4	+
5	-	network latency. Output metrics count frames and encoded output data produced	5	+	S5a is measurement-only. It records process-local runtime marks and harness-owned
6	-	by the daemon; snapshot writes and serialized bytes are local snapshot-store	6	+	end-to-end boundaries without changing rendering, pacing, transport policy, wire
7	-	proxies. These metrics approximate local rendering, IPC payload, and persistence	7	+	bytes, or `ProtocolVersion`. The canonical matrix and workload/topology/
8	-	work, not a Unix socket, SSH, or UDP transport. Real transport impairment and	8	+	transport coverage are in [`testdata/perf/manifest.json`](../testdata/perf/manifest.json):
9	-	latency benchmarks are deferred.	9	+	`1x4`, `4x1`, `4x4`, and `8x1`, all at `120x40` with 10,000 rows per pane;
10	-		10	+	idle, active/all/inactive output, interactive flood, copy search, resize sweep,
11	-	## Large-history daemon baseline	11	+	snapshot/output/resize, and attach/restore/tab switch; local, SSH stdio, UDP
12	-		12	+	baseline, UDP 25/100 ms, and UDP 0%/1% loss. It has 252 scenarios (4×9×7),
13	-	Use a single iteration only as a smoke check; it is not suitable for drawing	13	+	with inapplicable cases required to be explicit. The seven transport entries are
14	-	performance conclusions:	14	+	`local`, `ssh_stdio`, `udp_baseline`, `udp_loss_0pct`, `udp_25ms`, `udp_100ms`,
15	-		15	+	and `udp_loss_1pct`.
16	-	```sh	16	+
17	-	go test ./internal/usecase/daemon -run '^$' -bench '^BenchmarkDaemonHistory' -benchtime=1x -benchmem	17	+	### Trace schema and clock domains
18	-	```	18	+
19	-		19	+	Every process JSONL record has exactly these fields:
20	-	For before/after comparison, collect repeated samples for the full matrix	20	+	`schema` (uint), `process_id` (string), `component` (string), `scenario`
21	-	(all workloads and topologies), pinning CPU/governor and recording the host.	21	+	(string), `run`, `sequence`, `request_id`, `epoch` (uint), `kind` (string),
22	-	The benchmark harness must be committed and present in both measured	22	+	`tick` (int64), `bytes`, `fragments`, `retransmits`, `pending` (uint),
23	-	revisions: collect the baseline after the harness is introduced but before the	23	+	`ack_rtt_nanos` (int64), and `valid` (bool). Schema 1 is represented by:
24	-	optimization, then collect the optimized revision with that same harness. The	24	+
25	-	fixed duration and count make the output suitable for `benchstat`:	25	+	```json
26	-		26	+	{"schema":1,"process_id":"p","component":"daemon","scenario":"s","run":1,"sequence":1,"request_id":1,"epoch":1,"kind":"diff_start","tick":0,"bytes":0,"fragments":0,"retransmits":0,"pending":0,"ack_rtt_nanos":0,"valid":true}
27	-	```sh
28	-	go test ./internal/usecase/daemon -run '^$' -bench '^BenchmarkDaemonHistory' -benchtime=1s -count=5 -benchmem > /tmp/vev-history-after.txt
29	-	benchstat /tmp/vev-history-before.txt /tmp/vev-history-after.txt
30		```	27		```
31			28
32	-	The matrix exercises live paint, snapshot capture, copy-mode entry, copy search,	29	+	`tick` is stamped only by the concrete JSONL observer using its process-local
33	-	and resize at `120x40` client geometry. Each pane is populated with 10,000	30	+	injected monotonic clock; producers supply no timestamp. The harness sets `VEV_PERF_TRACE`,
34	-	full-width deterministic history rows containing a stable `needle-<tab>-<row>`	31	+	`VEV_PERF_PROCESS_ID`, `VEV_PERF_SCENARIO`, and `VEV_PERF_RUN` for every
35	-	prefix and patterned text. The layouts are:	32	+	launched role, identifying its exclusive trace path and process/scenario/run
			33	+	mapping. Components in one OS
			34	+	process share that observer. The harness clock is a separate monotonic domain
			35	+	and is the sole clock for input-injection → terminal-bytes-successfully-flushed
			36	+	end-to-end samples. Cross-process ticks are never ordered or subtracted;
			37	+	records correlate only by scenario/run/sequence/request/epoch IDs.
36			38
37	-	- 1 tab x 1 pane (control)	39	+	The named process-local spans are capture, compose, diff, queue wait,
38	-	- 1 tab x 4 panes	40	+	ACK-blocked interval, emit, adapter send, and adapter receive. Their start/end
39	-	- 4 tabs x 1 pane	41	+	marks are paired within one process and summarized by component/adapter with
40	-	- 4 tabs x 4 panes	42	+	raw records, counts, p50/p95/p99, and max. Diagnostic fields (bytes, fragments,
41	-	- 8 tabs x 1 pane	43	+	retransmits, pending, ACK RTT) remain separate from duration calculations.
42			44
43	-	The fixture starts no PTY readers, schedulers, clocks, or transports. It primes	45	+	### Baseline procedure and validity
44	-	the renderer shadow before timing; live paint then alternates fixed writes. The
45	-	reported `outputframes/op` value counts output messages. `outputbytes/op`
46	-	counts their encoded terminal data and excludes the 24-byte output header;
47	-	`framepayloadbytes/op` counts the complete output payload including that header.
48	-	`snapshotwrites/op` and `snapshotbytes/op` are local snapshot-store writes and
49	-	serialized bytes. They must not be presented as network throughput or latency.
50			46
51	-	Capture a pre-change result outside the repository, including Go and kernel	47	+	Use the public CLI harness (not internal packages):
52	-	context:
53			48
54		```sh	49		```sh
55	-	{	50	+	go build -o /tmp/vev-perf ./
56	-	date -Is	51	+	rm -rf testdata/perf/results
57	-	go version	52	+	go run ./cmd/vev-perf-harness --vev-bin /tmp/vev-perf \
58	-	uname -a	53	+	--manifest testdata/perf/manifest.json --out testdata/perf/results \
59	-	git rev-parse HEAD	54	+	--warmup 10s --duration 30s --repetitions 10
60	-	# Smoke only; do not use this result for comparisons.
61	-	go test ./internal/usecase/daemon -run '^$' -bench '^BenchmarkDaemonHistory' -benchtime=1x -benchmem
62	-	# Repeated full matrix for every workload/topology used in conclusions.
63	-	go test ./internal/usecase/daemon -run '^$' -bench '^BenchmarkDaemonHistory' -benchtime=1s -count=5 -benchmem
64	-	} > /tmp/vev-history-before.txt 2>&1
65		```	55		```
66			56
67	-	## Immutable document allocation gate	57	+	The minimums are 30 seconds measured and 10 independent repetitions; warmup is
			58	+	excluded. During the measured interval the harness injects at most one event per
			59	+	second. At least 10 complete event pairs are required: both injection and
			60	+	successful terminal flush must fall inside the measured interval; straddling or
			61	+	failed events are retained only as raw diagnostics and are not samples. Each run must contain `manifest.json`, `raw-harness.jsonl`, per-process
			62	+	JSONL, `runs.json`, and `summary.json`. Reject results with missing input/flush
			63	+	boundaries, unmatched IDs, bad correlation, nonmonotonic same-process sequence,
			64	+	negative same-domain spans, fewer than 10 complete in-interval event pairs,
			65	+	invalid denominators,
			66	+	any cross-process tick arithmetic, or unmet duration/repetition minima. Raw
			67	+	JSONL and run manifests are the evidence files; no result authorizes a policy
			68	+	or optimization change.
68			69
69	-	The copy-mode document retains an immutable view of existing scrollback rows and	70	+	## Existing in-process smoke checks
70	-	clones only the visible frame. Visual-search models share that document; cloning
71	-	a model copies only its mutable input and match state. Check those two paths at
72	-	10,000 history rows with:
73			71
74	-	```sh	72	+	These commands measure local daemon work only and are not end-to-end transport
75	-	go test ./internal/usecase/copy ./internal/usecase/visualsearch -run '^$' -bench '^(BenchmarkNewSnapshot10KRows\|BenchmarkVisualSearchClone10KRows)$' -benchtime=1s -count=5 -benchmem	73	+	results:
76	-	```
77	-
78	-	Treat `allocs/op` and `B/op` as allocation gates: compare the complete repeated
79	-	output with the accepted baseline before merging, and investigate any increase.
80	-	The automated scaling gates also compare equal row/match counts at narrow and
81	-	wide row widths; `B/op` may grow by at most 2x, a conservative relative
82	-	allowance that catches full cell-row copies without depending on an absolute
83	-	Go-version-specific budget. This scope does not yet include an
84	-	incremental/full-text search index: queries still scan the immutable document
85	-	and produce a fresh match list. Coordinator
86	-	behavior, chunk encoding, compression, GC tuning, and real Unix-socket, SSH, or
87	-	UDP impairment/latency remain out of scope.
88	-
89	-	## Lower-level smoke baseline
90	-
91	-	Use this command to check renderer, VT, IPC, and daemon microbenchmarks without
92	-	claiming an end-to-end transport result:
93			74
94		```sh	75		```sh
			76	+	go test ./internal/usecase/daemon -run '^$' -bench '^BenchmarkDaemonHistory' -benchtime=1x -benchmem
95		go test ./pkg/renderer ./pkg/vt ./internal/adapters/ipc -run '^$' -bench=. -benchmem	77		go test ./pkg/renderer ./pkg/vt ./internal/adapters/ipc -run '^$' -bench=. -benchmem
96		go test ./internal/usecase/daemon -run '^$' -bench '^BenchmarkComposeFloatingFrameCached$' -benchmem	78		go test ./internal/usecase/daemon -run '^$' -bench '^BenchmarkComposeFloatingFrameCached$' -benchmem
97		```	79		```
98	-
99	-	Enable daemon pprof explicitly when collecting a daemon profile:
100	-
101	-	```sh
102	-	VEV_PPROF_ADDR=127.0.0.1:6060 vev --daemon
103	-	```

M internal/adapters/dgram/ssh_dialer.go +2 -1

 	BootstrapTimeout time.Duration
 	ProbeTimeout     time.Duration
 	Log              *slog.Logger
+	RuntimeObserver  ports.SerializedRuntimeObserver
 }
 func NewRemoteDialer(target, session string) RemoteDialer {
 		RebindPacketConn: func(net.PacketConn) (net.PacketConn, error) {
 			return listenUDPPacket(ctx)
 		},
-	})
+	}, WithRuntimeObserver(d.RuntimeObserver))
 	if err != nil {
 		_ = pc.Close()
 		return nil, udpUnavailable("create UDP transport", err, &stderr)

M internal/adapters/dgram/transport.go +96 -7


@@ -45,6 +45,14 @@ var (
45		errControlFull = errors.New("dgram: control queue full")	45		errControlFull = errors.New("dgram: control queue full")
46		)	46		)
47			47
			48	+	type RuntimeOption func(*Transport)
			49	+
			50	+	// WithRuntimeObserver enables process-local marks without changing dgram's
			51	+	// behavior clock or carriage policy.
			52	+	func WithRuntimeObserver(observer ports.SerializedRuntimeObserver) RuntimeOption {
			53	+	return func(t *Transport) { t.runtimeObserver = observer }
			54	+	}
			55	+
48		type Options struct {	56		type Options struct {
49		MTU int	57		MTU int
50		ResendAfter time.Duration	58		ResendAfter time.Duration

@@ -168,6 +176,11 @@ type Transport struct {
168		// mutation and before ordered event publication.	176		// mutation and before ordered event publication.
169		afterLinkStateCommit func(ports.LinkState)	177		afterLinkStateCommit func(ports.LinkState)
170		observe DiagnosticObserver	178		observe DiagnosticObserver
			179	+	runtimeObserver ports.SerializedRuntimeObserver
			180	+	operationMu sync.Mutex
			181	+	operationCount int
			182	+	closing bool
			183	+	operationsDone chan struct{}
171		diagnosticCh chan Diagnostic	184		diagnosticCh chan Diagnostic
172		heartbeat time.Duration	185		heartbeat time.Duration
173		resendAfter time.Duration	186		resendAfter time.Duration

@@ -311,11 +324,11 @@ func (s *writeDeadlineState) expire(now time.Time) {
311		s.mu.Unlock()	324		s.mu.Unlock()
312		}	325		}
313			326
314	-	func NewTransport(pc net.PacketConn, peer net.Addr, key []byte, sendDir, recvDir uint32) (*Transport, error) {	327	+	func NewTransport(pc net.PacketConn, peer net.Addr, key []byte, sendDir, recvDir uint32, runtimeOpts ...RuntimeOption) (*Transport, error) {
315	-	return NewTransportWithOptions(pc, peer, key, sendDir, recvDir, Options{})	328	+	return NewTransportWithOptions(pc, peer, key, sendDir, recvDir, Options{}, runtimeOpts...)
316		}	329		}
317			330
318	-	func NewTransportWithOptions(pc net.PacketConn, peer net.Addr, key []byte, sendDir, recvDir uint32, opts Options) (*Transport, error) {	331	+	func NewTransportWithOptions(pc net.PacketConn, peer net.Addr, key []byte, sendDir, recvDir uint32, opts Options, runtimeOpts ...RuntimeOption) (*Transport, error) {
319		c, err := pdgram.NewCodec(key)	332		c, err := pdgram.NewCodec(key)
320		if err != nil {	333		if err != nil {
321		return nil, err	334		return nil, err

@@ -360,6 +373,11 @@ func NewTransportWithOptions(pc net.PacketConn, peer net.Addr, key []byte, sendD
360		nextRecvSeq: 1,	373		nextRecvSeq: 1,
361		recvBuf: make(map[uint64]ports.Frame),	374		recvBuf: make(map[uint64]ports.Frame),
362		}	375		}
			376	+	for _, opt := range runtimeOpts {
			377	+	if opt != nil {
			378	+	opt(t)
			379	+	}
			380	+	}
363		t.sendWake = make(chan struct{})	381		t.sendWake = make(chan struct{})
364		t.control = make(chan controlRecord, controlQueueSize)	382		t.control = make(chan controlRecord, controlQueueSize)
365		t.ackWake = make(chan struct{}, 1)	383		t.ackWake = make(chan struct{}, 1)

@@ -387,9 +405,19 @@ func NewTransportWithOptions(pc net.PacketConn, peer net.Addr, key []byte, sendD
387		return t, nil	405		return t, nil
388		}	406		}
389			407
390	-	func (t *Transport) Send(f ports.Frame) error { return t.send(f, false) }	408	+	func (t *Transport) Send(f ports.Frame) error {
			409	+	end := t.beginRuntimeOperation(ports.RuntimeAdapterSendStart, uint64(len(f.Payload)))
			410	+	err := t.send(f, false)
			411	+	end(err == nil)
			412	+	return err
			413	+	}
391			414
392	-	func (t *Transport) SendAsync(f ports.Frame) error { return t.send(f, true) }	415	+	func (t *Transport) SendAsync(f ports.Frame) error {
			416	+	end := t.beginRuntimeOperation(ports.RuntimeAdapterSendStart, uint64(len(f.Payload)))
			417	+	err := t.send(f, true)
			418	+	end(err == nil)
			419	+	return err
			420	+	}
393			421
394		// SendSynchronous owns queue reservation, fragment pacing, per-write deadlines,	422		// SendSynchronous owns queue reservation, fragment pacing, per-write deadlines,
395		// and close cancellation. Callers must not layer a shorter preflight timer over it.	423		// and close cancellation. Callers must not layer a shorter preflight timer over it.

@@ -568,25 +596,35 @@ func (t *Transport) removePendingLocked(seq uint64) bool {
568		}	596		}
569			597
570		func (t *Transport) Recv() (ports.Frame, error) {	598		func (t *Transport) Recv() (ports.Frame, error) {
			599	+	end := t.beginRuntimeOperation(ports.RuntimeAdapterReceiveStart, 0)
571		select {	600		select {
572		case f := <-t.in:	601		case f := <-t.in:
			602	+	end(true)
573		return f, nil	603		return f, nil
574		case <-t.done:	604		case <-t.done:
575		t.mu.Lock()	605		t.mu.Lock()
576		err := t.closeErr	606		err := t.closeErr
577		t.mu.Unlock()	607		t.mu.Unlock()
578		if err != nil {	608		if err != nil {
			609	+	end(false)
579		return ports.Frame{}, err	610		return ports.Frame{}, err
580		}	611		}
581	-	return ports.Frame{}, errors.New("dgram: closed")	612	+	err = errors.New("dgram: closed")
			613	+	end(false)
			614	+	return ports.Frame{}, err
582		}	615		}
583		}	616		}
584		func (t *Transport) Close() error {	617		func (t *Transport) Close() error {
			618	+	done := t.beginShutdown()
585		t.closeWithError(errors.New("dgram: closed"))	619		t.closeWithError(errors.New("dgram: closed"))
586		t.mu.Lock()	620		t.mu.Lock()
587		pc := t.pc	621		pc := t.pc
588		t.mu.Unlock()	622		t.mu.Unlock()
589	-	return pc.Close()	623	+	err := pc.Close()
			624	+	if done != nil {
			625	+	<-done
			626	+	}
			627	+	return err
590		}	628		}
591			629
592		func (t *Transport) Peer() net.Addr { t.mu.Lock(); defer t.mu.Unlock(); return t.peer }	630		func (t *Transport) Peer() net.Addr { t.mu.Lock(); defer t.mu.Unlock(); return t.peer }

@@ -1585,4 +1623,55 @@ func decodeData(b []byte) (uint64, bool, ports.Frame, bool) {
1585		}	1623		}
1586			1624
1587		// DatagramTransport marks Transport as a UDP-style datagram transport.	1625		// DatagramTransport marks Transport as a UDP-style datagram transport.
			1626	+	func (t *Transport) beginRuntimeOperation(start ports.RuntimeMarkKind, bytes uint64) func(bool) {
			1627	+	if t.runtimeObserver == nil {
			1628	+	return func(bool) {}
			1629	+	}
			1630	+	if !t.beginObservedOperation() {
			1631	+	return func(bool) {}
			1632	+	}
			1633	+	correlation := ports.NewRuntimeCorrelation()
			1634	+	t.runtimeObserver.ObserveRuntime(ports.NewRuntimeMarkWithCorrelation("dgram", correlation, start, bytes, true))
			1635	+	end := ports.RuntimeAdapterSendEnd
			1636	+	if start == ports.RuntimeAdapterReceiveStart {
			1637	+	end = ports.RuntimeAdapterReceiveEnd
			1638	+	}
			1639	+	return func(valid bool) {
			1640	+	defer t.finishObservedOperation()
			1641	+	t.runtimeObserver.ObserveRuntime(ports.NewRuntimeMarkWithCorrelation("dgram", correlation, end, bytes, valid))
			1642	+	}
			1643	+	}
			1644	+
			1645	+	func (t *Transport) beginObservedOperation() bool {
			1646	+	t.operationMu.Lock()
			1647	+	defer t.operationMu.Unlock()
			1648	+	if t.closing {
			1649	+	return false
			1650	+	}
			1651	+	if t.operationCount == 0 {
			1652	+	t.operationsDone = make(chan struct{})
			1653	+	}
			1654	+	t.operationCount++
			1655	+	return true
			1656	+	}
			1657	+
			1658	+	func (t *Transport) finishObservedOperation() {
			1659	+	t.operationMu.Lock()
			1660	+	defer t.operationMu.Unlock()
			1661	+	t.operationCount--
			1662	+	if t.operationCount == 0 {
			1663	+	close(t.operationsDone)
			1664	+	}
			1665	+	}
			1666	+
			1667	+	func (t *Transport) beginShutdown() <-chan struct{} {
			1668	+	t.operationMu.Lock()
			1669	+	defer t.operationMu.Unlock()
			1670	+	t.closing = true
			1671	+	if t.operationCount == 0 {
			1672	+	return nil
			1673	+	}
			1674	+	return t.operationsDone
			1675	+	}
			1676	+
1588		func (*Transport) DatagramTransport() {}	1677		func (*Transport) DatagramTransport() {}

A internal/adapters/dgram/transport_observability_test.go +208 -0

@@ -0,0 +1,208 @@
	1	+	package dgram
	2	+
	3	+	import (
	4	+	"encoding/json"
	5	+	"errors"
	6	+	"net"
	7	+	"os"
	8	+	"reflect"
	9	+	"strings"
	10	+	"sync"
	11	+	"testing"
	12	+	"time"
	13	+
	14	+	"github.com/bnema/vev/internal/adapters/observability"
	15	+	"github.com/bnema/vev/internal/ports"
	16	+	)
	17	+
	18	+	func TestTransportObservabilityDgramFailedCloseEndsReceiveAndSend(t *testing.T) {
	19	+	behaviorClock := newManualClock(time.Unix(0, 417))
	20	+	link := newSimulatedLink(behaviorClock, packetPolicy{})
	21	+	aPC, _ := newPairWithCapacity(link, 64)
	22	+	closeErr := errors.New("packet connection close failed")
	23	+	tracePath := t.TempDir() + "/dgram-close.jsonl"
	24	+	observer, closer, err := observability.NewJSONL(tracePath, dgramObserverClock{now: time.Unix(0, 521)}, "dgram-process")
	25	+	if err != nil {
	26	+	t.Fatalf("NewJSONL() error = %v", err)
	27	+	}
	28	+	receiveStarted := make(chan struct{})
	29	+	signalingObserver := &dgramStartObserver{RuntimeObserver: observer, receiveStarted: receiveStarted}
	30	+	reporter := ports.NewSerializedRuntimeObserver(signalingObserver, 64)
	31	+	defer reporter.Close()
	32	+	transport, err := NewTransportWithOptions(&dgramCloseFailPC{fakePC: aPC, err: closeErr}, testAddr("b"), key(), 1, 2, Options{Clock: behaviorClock, ResendAfter: time.Hour, Heartbeat: time.Hour}, WithRuntimeObserver(reporter))
	33	+	if err != nil {
	34	+	t.Fatalf("NewTransportWithOptions() error = %v", err)
	35	+	}
	36	+	if err := transport.Send(ports.Frame{Type: ports.MsgOutput, Payload: []byte("write failure")}); err == nil {
	37	+	t.Fatal("Send() error = nil on packet write failure")
	38	+	}
	39	+	recvDone := make(chan error, 1)
	40	+	go func() { _, err := transport.Recv(); recvDone <- err }()
	41	+	<-receiveStarted
	42	+	if err := transport.Close(); !errors.Is(err, closeErr) {
	43	+	t.Fatalf("Close() error = %v, want %v", err, closeErr)
	44	+	}
	45	+	if err := <-recvDone; err == nil {
	46	+	t.Fatal("Recv() error = nil after failed Close()")
	47	+	}
	48	+	reporter.Close()
	49	+	if err := closer.Close(); err != nil {
	50	+	t.Fatalf("trace Close() error = %v", err)
	51	+	}
	52	+	assertDgramFailedSpanPairs(t, tracePath)
	53	+	}
	54	+
	55	+	type dgramStartObserver struct {
	56	+	ports.RuntimeObserver
	57	+	receiveStarted chan struct{}
	58	+	once sync.Once
	59	+	}
	60	+
	61	+	func (o *dgramStartObserver) ObserveRuntime(mark ports.RuntimeMark) {
	62	+	if mark.Kind == ports.RuntimeAdapterReceiveStart {
	63	+	o.once.Do(func() { close(o.receiveStarted) })
	64	+	}
	65	+	o.RuntimeObserver.ObserveRuntime(mark)
	66	+	}
	67	+
	68	+	type dgramCloseFailPC struct {
	69	+	*fakePC
	70	+	err error
	71	+	}
	72	+
	73	+	func (p *dgramCloseFailPC) WriteTo([]byte, net.Addr) (int, error) {
	74	+	return 0, errors.New("packet write failed")
	75	+	}
	76	+
	77	+	func (p *dgramCloseFailPC) Close() error {
	78	+	_ = p.fakePC.Close()
	79	+	return p.err
	80	+	}
	81	+
	82	+	func assertDgramFailedSpanPairs(t *testing.T, path string) {
	83	+	t.Helper()
	84	+	data, err := os.ReadFile(path)
	85	+	if err != nil {
	86	+	t.Fatalf("ReadFile(trace): %v", err)
	87	+	}
	88	+	type mark struct {
	89	+	ProcessID string `json:"process_id"`
	90	+	Scenario string `json:"scenario"`
	91	+	Run uint64 `json:"run"`
	92	+	Sequence uint64 `json:"sequence"`
	93	+	RequestID uint64 `json:"request_id"`
	94	+	Epoch uint64 `json:"epoch"`
	95	+	Kind string `json:"kind"`
	96	+	Valid bool `json:"valid"`
	97	+	}
	98	+	type spanKey struct {
	99	+	processID, scenario string
	100	+	run, sequence, requestID, epoch uint64
	101	+	kind string
	102	+	}
	103	+	starts := make(map[spanKey]mark)
	104	+	pairs := map[string]int{}
	105	+	for _, line := range strings.Split(strings.TrimSpace(string(data)), "\n") {
	106	+	var m mark
	107	+	if err := json.Unmarshal([]byte(line), &m); err != nil {
	108	+	t.Fatalf("Unmarshal(trace): %v", err)
	109	+	}
	110	+	key := spanKey{m.ProcessID, m.Scenario, m.Run, m.Sequence, m.RequestID, m.Epoch, m.Kind}
	111	+	switch m.Kind {
	112	+	case "adapter_receive_start", "adapter_send_start":
	113	+	if _, exists := starts[key]; exists {
	114	+	t.Fatalf("duplicate adapter start: %#v", m)
	115	+	}
	116	+	starts[key] = m
	117	+	case "adapter_receive_end", "adapter_send_end":
	118	+	key.kind = strings.TrimSuffix(m.Kind, "_end") + "_start"
	119	+	start, ok := starts[key]
	120	+	if !ok \|\| m.Valid {
	121	+	t.Fatalf("unmatched or valid failed end: start=%#v end=%#v", start, m)
	122	+	}
	123	+	pairs[strings.TrimSuffix(m.Kind, "_end")]++
	124	+	delete(starts, key)
	125	+	}
	126	+	}
	127	+	if len(starts) != 0 {
	128	+	t.Fatalf("unmatched adapter starts: %#v", starts)
	129	+	}
	130	+	if pairs["adapter_send"] != 1 \|\| pairs["adapter_receive"] != 1 \|\| len(pairs) != 2 {
	131	+	t.Fatalf("failed adapter span pairs=%v, want exactly one send and one receive", pairs)
	132	+	}
	133	+	}
	134	+
	135	+	func TestTransportObservabilityDgramKeepsBehaviorClockSeparate(t *testing.T) {
	136	+	behaviorClock := newManualClock(time.Unix(0, 401))
	137	+	link := newSimulatedLink(behaviorClock, packetPolicy{})
	138	+	aPC, bPC := newPairWithCapacity(link, 64)
	139	+	tracePath := t.TempDir() + "/dgram.jsonl"
	140	+	observer, closer, err := observability.NewJSONL(tracePath, dgramObserverClock{now: time.Unix(0, 509)}, "dgram-process")
	141	+	if err != nil {
	142	+	t.Fatalf("NewJSONL() error = %v", err)
	143	+	}
	144	+	reporter := ports.NewSerializedRuntimeObserver(observer, 64)
	145	+	defer reporter.Close()
	146	+
	147	+	a, err := NewTransportWithOptions(aPC, testAddr("b"), key(), 1, 2, Options{Clock: behaviorClock, ResendAfter: time.Hour, Heartbeat: time.Hour}, WithRuntimeObserver(reporter))
	148	+	if err != nil {
	149	+	t.Fatalf("NewTransportWithOptions(a) error = %v", err)
	150	+	}
	151	+	defer func() { _ = a.Close() }()
	152	+	b, err := NewTransportWithOptions(bPC, testAddr("a"), key(), 2, 1, Options{Clock: behaviorClock, ResendAfter: time.Hour, Heartbeat: time.Hour}, WithRuntimeObserver(reporter))
	153	+	if err != nil {
	154	+	t.Fatalf("NewTransportWithOptions(b) error = %v", err)
	155	+	}
	156	+	defer func() { _ = b.Close() }()
	157	+
	158	+	want := ports.Frame{Type: ports.MsgOutput, Payload: []byte("datagram carriage remains exact")}
	159	+	if err := a.Send(want); err != nil {
	160	+	t.Fatalf("Send() error = %v", err)
	161	+	}
	162	+	gotCh := make(chan ports.Frame, 1)
	163	+	errCh := make(chan error, 1)
	164	+	go func() {
	165	+	frame, recvErr := b.Recv()
	166	+	if recvErr != nil {
	167	+	errCh <- recvErr
	168	+	return
	169	+	}
	170	+	gotCh <- frame
	171	+	}()
	172	+	select {
	173	+	case err := <-errCh:
	174	+	t.Fatalf("Recv() error = %v", err)
	175	+	case got := <-gotCh:
	176	+	if !reflect.DeepEqual(got, want) {
	177	+	t.Fatalf("carried frame = %#v, want %#v", got, want)
	178	+	}
	179	+	case <-time.After(time.Second):
	180	+	t.Fatal("Recv() did not receive deterministic simulated datagram")
	181	+	}
	182	+	reporter.Close()
	183	+	if err := closer.Close(); err != nil {
	184	+	t.Fatalf("trace Close() error = %v", err)
	185	+	}
	186	+	trace, err := os.ReadFile(tracePath)
	187	+	if err != nil {
	188	+	t.Fatalf("ReadFile(trace): %v", err)
	189	+	}
	190	+	for _, kind := range []string{"adapter_send_start", "adapter_send_end", "adapter_receive_start", "adapter_receive_end"} {
	191	+	if !strings.Contains(string(trace), `"kind":"`+kind+`"`) {
	192	+	t.Errorf("trace missing %s", kind)
	193	+	}
	194	+	}
	195	+	}
	196	+
	197	+	type dgramObserverClock struct{ now time.Time }
	198	+
	199	+	func (c dgramObserverClock) Now() time.Time { return c.now }
	200	+	func (c dgramObserverClock) NewTimer(d time.Duration) ports.Timer {
	201	+	return dgramObserverTimer{timer: time.NewTimer(d)}
	202	+	}
	203	+
	204	+	type dgramObserverTimer struct{ timer *time.Timer }
	205	+
	206	+	func (t dgramObserverTimer) C() <-chan time.Time { return t.timer.C }
	207	+	func (t dgramObserverTimer) Reset(d time.Duration) bool { return t.timer.Reset(d) }
	208	+	func (t dgramObserverTimer) Stop() bool { return t.timer.Stop() }

A internal/adapters/dgram/transport_replay_test.go +31 -0

@@ -0,0 +1,31 @@
	1	+	package dgram
	2	+
	3	+	import (
	4	+	"testing"
	5	+
	6	+	"github.com/bnema/vev/internal/ports"
	7	+	"github.com/bnema/vev/internal/testutil/replaytest"
	8	+	"github.com/stretchr/testify/require"
	9	+	)
	10	+
	11	+	func TestTransportReplay(t *testing.T) {
	12	+	replaytest.Run(t, func(t *testing.T, frames []ports.Frame) []ports.Frame {
	13	+	link := newSimulatedLink(fixedClock{}, packetPolicy{})
	14	+	leftPC, rightPC := newPairWithCapacity(link, 32)
	15	+	left, err := NewTransport(leftPC, testAddr("b"), key(), 1, 2)
	16	+	require.NoError(t, err)
	17	+	right, err := NewTransport(rightPC, testAddr("a"), key(), 2, 1)
	18	+	require.NoError(t, err)
	19	+	t.Cleanup(func() { _ = left.Close(); _ = right.Close() })
	20	+
	21	+	got := make([]ports.Frame, 0, len(frames))
	22	+	for _, frame := range frames {
	23	+	require.NoError(t, left.Send(frame))
	24	+	link.flush(rightPC)
	25	+	received, recvErr := right.Recv()
	26	+	require.NoError(t, recvErr)
	27	+	got = append(got, received)
	28	+	}
	29	+	return got
	30	+	})
	31	+	}

M internal/adapters/ipc/dial.go +7 -5

 package ipc
 import (
+	"context"
 	"fmt"
 	"net"
 	"path/filepath"
 	"github.com/bnema/vev/internal/ports"
 )
-// Dial connects to the daemon's socket in dir (dir/daemon.sock) and returns
+// DialContext connects to the daemon's socket in dir (dir/daemon.sock) and
-// a Transport speaking vev's framed protocol. A dial failure (no daemon
+// returns a Transport speaking vev's framed protocol. A dial failure (no daemon
 // listening, or a stale/absent socket) is reported as a wrapped error so
 // callers can decide whether to spawn a daemon and retry.
-func Dial(dir string) (ports.Transport, error) {
+func DialContext(ctx context.Context, dir string, opts ...Option) (ports.Transport, error) {
 	sockPath := filepath.Join(dir, socketFileName)
-	conn, err := net.Dial("unix", sockPath)
+	var dialer net.Dialer
+	conn, err := dialer.DialContext(ctx, "unix", sockPath)
 	if err != nil {
 		return nil, fmt.Errorf("ipc: dial %s: %w", sockPath, err)
 	}
-	return NewTransport(conn), nil
+	return NewTransport(conn, opts...), nil
 }

A internal/adapters/ipc/dial_test.go +80 -0

@@ -0,0 +1,80 @@
	1	+	package ipc
	2	+
	3	+	import (
	4	+	"context"
	5	+	"errors"
	6	+	"path/filepath"
	7	+	"testing"
	8	+	"time"
	9	+
	10	+	"github.com/bnema/vev/internal/ports"
	11	+	)
	12	+
	13	+	type dialTestObserver struct{}
	14	+
	15	+	func (dialTestObserver) ObserveRuntime(ports.RuntimeMark) {}
	16	+
	17	+	func TestDialContextCanceledOrExpired(t *testing.T) {
	18	+	tests := []struct {
	19	+	name string
	20	+	ctx func() (context.Context, context.CancelFunc)
	21	+	want error
	22	+	}{
	23	+	{
	24	+	name: "canceled",
	25	+	ctx: func() (context.Context, context.CancelFunc) {
	26	+	ctx, cancel := context.WithCancel(context.Background())
	27	+	cancel()
	28	+	return ctx, func() {}
	29	+	},
	30	+	want: context.Canceled,
	31	+	},
	32	+	{
	33	+	name: "deadline exceeded",
	34	+	ctx: func() (context.Context, context.CancelFunc) {
	35	+	return context.WithDeadline(context.Background(), time.Now().Add(-time.Second))
	36	+	},
	37	+	want: context.DeadlineExceeded,
	38	+	},
	39	+	}
	40	+
	41	+	for _, tt := range tests {
	42	+	t.Run(tt.name, func(t *testing.T) {
	43	+	ctx, cancel := tt.ctx()
	44	+	defer cancel()
	45	+
	46	+	transport, err := DialContext(ctx, t.TempDir())
	47	+	if transport != nil {
	48	+	t.Fatal("DialContext returned a transport for a completed context")
	49	+	}
	50	+	if !errors.Is(err, tt.want) {
	51	+	t.Fatalf("DialContext error = %v, want wrapping %v", err, tt.want)
	52	+	}
	53	+	})
	54	+	}
	55	+	}
	56	+
	57	+	func TestDialContextPreservesOptions(t *testing.T) {
	58	+	dir := filepath.Join(t.TempDir(), "vev")
	59	+	listener, err := Listen(dir)
	60	+	if err != nil {
	61	+	t.Fatalf("Listen: %v", err)
	62	+	}
	63	+	defer func() { _ = listener.Close() }()
	64	+
	65	+	observer := ports.NewSerializedRuntimeObserver(dialTestObserver{}, 1)
	66	+	defer observer.Close()
	67	+	transport, err := DialContext(context.Background(), dir, WithRuntimeObserver(observer))
	68	+	if err != nil {
	69	+	t.Fatalf("DialContext: %v", err)
	70	+	}
	71	+	defer func() { _ = transport.Close() }()
	72	+
	73	+	got, ok := transport.(*unixTransport)
	74	+	if !ok {
	75	+	t.Fatalf("DialContext transport = %T, want *unixTransport", transport)
	76	+	}
	77	+	if got.observer != observer {
	78	+	t.Fatal("DialContext did not apply runtime observer option")
	79	+	}
	80	+	}

M internal/adapters/ipc/listener.go +4 -3

 type unixListener struct {
 	ln   *net.UnixListener
 	addr string
+	opts []Option
 }
 // Listen creates the socket directory (0700) if needed and starts
 // the previous owner died without cleaning up, so the stale socket file is
 // unlinked and bind is retried once. A successful dial means a live daemon
 // owns the socket, and Listen returns ErrDaemonRunning.
-func Listen(dir string) (ports.Listener, error) {
+func Listen(dir string, opts ...Option) (ports.Listener, error) {
 	if err := safedir.EnsurePrivate(dir); err != nil {
 		return nil, fmt.Errorf("ipc: securing socket directory: %w", err)
 	}
 		return nil, fmt.Errorf("ipc: chmod %s: %w", sockPath, err)
 	}
-	return &unixListener{ln: ln, addr: sockPath}, nil
+	return &unixListener{ln: ln, addr: sockPath, opts: opts}, nil
 }
 // bindUnix binds and listens on sockPath.
 	if err != nil {
 		return nil, err
 	}
-	return NewTransport(conn), nil
+	return NewTransport(conn, l.opts...), nil
 }
 // Close stops accepting connections. The underlying net.UnixListener

M internal/adapters/ipc/transport.go +276 -17


@@ -5,6 +5,7 @@ import (
5		"errors"	5		"errors"
6		"io"	6		"io"
7		"net"	7		"net"
			8	+	"sync"
8			9
9		"github.com/bnema/vev/internal/ports"	10		"github.com/bnema/vev/internal/ports"
10		)	11		)

@@ -26,17 +27,54 @@ var ErrZeroLengthFrame = errors.New("ipc: zero-length frame")
26		// maxFrameLen, and by Send when a payload is too large to encode.	27		// maxFrameLen, and by Send when a payload is too large to encode.
27		var ErrFrameTooLarge = errors.New("ipc: frame exceeds maximum length")	28		var ErrFrameTooLarge = errors.New("ipc: frame exceeds maximum length")
28			29
			30	+	// ErrBackpressure means the bounded IPC egress queue is full. Callers using
			31	+	// SendAsync must treat it as a failed render emission rather than spawning an
			32	+	// unbounded queue of their own.
			33	+	var ErrBackpressure = errors.New("ipc: egress queue full")
			34	+
			35	+	var errClosed = errors.New("ipc: closed")
			36	+
			37	+	// sendQueueCapacity bounds frames accepted while a peer is not reading. It is
			38	+	// deliberately small: render output can be large and must never turn a wedged
			39	+	// peer into an unbounded memory commitment.
			40	+	const sendQueueCapacity = 8
			41	+
29		// unixTransport implements ports.Transport over a net.Conn (in practice an	42		// unixTransport implements ports.Transport over a net.Conn (in practice an
30		// AF_UNIX SOCK_STREAM connection, but any net.Conn works — this also lets	43		// AF_UNIX SOCK_STREAM connection, but any net.Conn works — this also lets
31		// tests exercise it over net.Pipe).	44		// tests exercise it over net.Pipe).
32		//	45		//
33	-	// unixTransport is safe for one concurrent caller of Send and one	46	+	// unixTransport has one writer goroutine which owns every conn.Write. Send
34	-	// concurrent caller of Recv at the same time (i.e. a single reader	47	+	// waits for that owner's wire attempt; SendAsync only performs a bounded,
35	-	// goroutine and a single writer goroutine); it is not safe for concurrent	48	+	// ordered enqueue. Recv still has one caller at a time.
36	-	// calls to Send from multiple goroutines, nor Recv from multiple	49	+	type Option func(*unixTransport)
37	-	// goroutines.	50	+
			51	+	// WithRuntimeObserver enables process-local transport marks. It accepts no
			52	+	// clock: timestamp ownership belongs to the configured observer.
			53	+	func WithRuntimeObserver(observer ports.SerializedRuntimeObserver) Option {
			54	+	return func(t *unixTransport) { t.observer = observer }
			55	+	}
			56	+
38		type unixTransport struct {	57		type unixTransport struct {
39	-	conn net.Conn	58	+	conn net.Conn
			59	+	observer ports.SerializedRuntimeObserver
			60	+
			61	+	operationMu sync.Mutex
			62	+	operationCount int
			63	+	observedClosing bool
			64	+	operationsDone chan struct{}
			65	+
			66	+	// egressMu synchronizes Close with enqueue. Synchronous enqueuers are
			67	+	// tracked separately so the writer cannot drain the queue until each has
			68	+	// observed closure or finished enqueueing.
			69	+	egressMu sync.Mutex
			70	+	egressClosing bool
			71	+	egressSenders int
			72	+	egressSendersDone chan struct{}
			73	+	egress chan sendRequest
			74	+	done chan struct{}
			75	+	writerDone chan struct{}
			76	+	closeOnce sync.Once
			77	+	closeErr error
40			78
41		// readBuf is reused across Recv calls (grow-once strategy): it grows	79		// readBuf is reused across Recv calls (grow-once strategy): it grows
42		// to fit the largest frame seen so far and is never shrunk. The Frame	80		// to fit the largest frame seen so far and is never shrunk. The Frame

@@ -47,42 +85,197 @@ type unixTransport struct {
47			85
48		// NewTransport wraps conn as a ports.Transport speaking vev's framed	86		// NewTransport wraps conn as a ports.Transport speaking vev's framed
49		// binary protocol.	87		// binary protocol.
50	-	func NewTransport(conn net.Conn) ports.Transport {	88	+	func NewTransport(conn net.Conn, opts ...Option) ports.Transport {
51	-	return &unixTransport{conn: conn}	89	+	t := &unixTransport{
			90	+	conn: conn,
			91	+	egress: make(chan sendRequest, sendQueueCapacity),
			92	+	done: make(chan struct{}),
			93	+	writerDone: make(chan struct{}),
			94	+	}
			95	+	for _, opt := range opts {
			96	+	if opt != nil {
			97	+	opt(t)
			98	+	}
			99	+	}
			100	+	go t.writeLoop()
			101	+	return t
			102	+	}
			103	+
			104	+	type sendRequest struct {
			105	+	data []byte
			106	+	done chan error // nil for SendAsync
			107	+	end func(bool)
52		}	108		}
53			109
54	-	// Send writes f as a single frame: a 4-byte big-endian length (covering	110	+	// Send queues f for the sole writer and waits for its wire attempt.
55	-	// the type byte and payload), the type byte, then the payload — assembled
56	-	// into one buffer and written with a single Write call.
57		func (t *unixTransport) Send(f ports.Frame) error {	111		func (t *unixTransport) Send(f ports.Frame) error {
			112	+	data, err := marshalFrame(f)
			113	+	if err != nil {
			114	+	return err
			115	+	}
			116	+	end := t.beginOperation(ports.RuntimeAdapterSendStart, uint64(len(f.Payload)))
			117	+	result := make(chan error, 1)
			118	+	if err := t.enqueueWait(sendRequest{data: data, done: result, end: end}); err != nil {
			119	+	end(false)
			120	+	return err
			121	+	}
			122	+	select {
			123	+	case err := <-result:
			124	+	return err
			125	+	case <-t.done:
			126	+	return errClosed
			127	+	}
			128	+	}
			129	+
			130	+	// SendAsync accepts f for ordered background transmission without waiting for
			131	+	// a socket write. The queue is bounded and ownership includes a payload copy.
			132	+	func (t *unixTransport) SendAsync(f ports.Frame) error {
			133	+	data, err := marshalFrame(f)
			134	+	if err != nil {
			135	+	return err
			136	+	}
			137	+	end := t.beginOperation(ports.RuntimeAdapterSendStart, uint64(len(f.Payload)))
			138	+	if err := t.enqueueAsync(sendRequest{data: data, end: end}); err != nil {
			139	+	end(false)
			140	+	return err
			141	+	}
			142	+	return nil
			143	+	}
			144	+
			145	+	func marshalFrame(f ports.Frame) ([]byte, error) {
58		n := 1 + len(f.Payload) // type + payload	146		n := 1 + len(f.Payload) // type + payload
59		if n > maxFrameLen {	147		if n > maxFrameLen {
60	-	return ErrFrameTooLarge	148	+	return nil, ErrFrameTooLarge
61		}	149		}
62	-
63		buf := make([]byte, frameHeaderLen+n)	150		buf := make([]byte, frameHeaderLen+n)
64		binary.BigEndian.PutUint32(buf[:frameHeaderLen], uint32(n))	151		binary.BigEndian.PutUint32(buf[:frameHeaderLen], uint32(n))
65		buf[frameHeaderLen] = byte(f.Type)	152		buf[frameHeaderLen] = byte(f.Type)
66		copy(buf[frameHeaderLen+1:], f.Payload)	153		copy(buf[frameHeaderLen+1:], f.Payload)
			154	+	return buf, nil
			155	+	}
			156	+
			157	+	// enqueueWait waits for space in the egress queue. Close wakes waiters and
			158	+	// the writer waits for all in-flight enqueues before draining queued requests.
			159	+	func (t *unixTransport) enqueueWait(req sendRequest) error {
			160	+	t.egressMu.Lock()
			161	+	if t.egressClosing {
			162	+	t.egressMu.Unlock()
			163	+	return errClosed
			164	+	}
			165	+	if t.egressSenders == 0 {
			166	+	t.egressSendersDone = make(chan struct{})
			167	+	}
			168	+	t.egressSenders++
			169	+	t.egressMu.Unlock()
			170	+
			171	+	select {
			172	+	case <-t.done:
			173	+	t.finishEgressSender()
			174	+	return errClosed
			175	+	case t.egress <- req:
			176	+	t.finishEgressSender()
			177	+	return nil
			178	+	}
			179	+	}
67			180
68	-	_, err := t.conn.Write(buf)	181	+	// enqueueAsync performs a bounded, nonblocking enqueue for SendAsync.
69	-	return err	182	+	func (t *unixTransport) enqueueAsync(req sendRequest) error {
			183	+	t.egressMu.Lock()
			184	+	defer t.egressMu.Unlock()
			185	+	if t.egressClosing {
			186	+	return errClosed
			187	+	}
			188	+	select {
			189	+	case t.egress <- req:
			190	+	return nil
			191	+	default:
			192	+	return ErrBackpressure
			193	+	}
			194	+	}
			195	+
			196	+	func (t *unixTransport) finishEgressSender() {
			197	+	t.egressMu.Lock()
			198	+	defer t.egressMu.Unlock()
			199	+	t.egressSenders--
			200	+	if t.egressSenders == 0 {
			201	+	close(t.egressSendersDone)
			202	+	}
			203	+	}
			204	+
			205	+	func (t *unixTransport) waitForEgressSenders() {
			206	+	t.egressMu.Lock()
			207	+	done := t.egressSendersDone
			208	+	t.egressMu.Unlock()
			209	+	if done != nil {
			210	+	<-done
			211	+	}
			212	+	}
			213	+
			214	+	func (t *unixTransport) writeLoop() {
			215	+	defer close(t.writerDone)
			216	+	for {
			217	+	select {
			218	+	case req := <-t.egress:
			219	+	t.completeWrite(req, writeAll(t.conn, req.data))
			220	+	case <-t.done:
			221	+	t.waitForEgressSenders()
			222	+	t.egressMu.Lock()
			223	+	for {
			224	+	select {
			225	+	case req := <-t.egress:
			226	+	t.completeWrite(req, errClosed)
			227	+	default:
			228	+	t.egressMu.Unlock()
			229	+	return
			230	+	}
			231	+	}
			232	+	}
			233	+	}
			234	+	}
			235	+
			236	+	func (t *unixTransport) completeWrite(req sendRequest, err error) {
			237	+	if req.end != nil {
			238	+	req.end(err == nil)
			239	+	}
			240	+	if req.done != nil {
			241	+	req.done <- err
			242	+	}
			243	+	}
			244	+
			245	+	func writeAll(w io.Writer, data []byte) error {
			246	+	for len(data) > 0 {
			247	+	n, err := w.Write(data)
			248	+	if n > 0 {
			249	+	data = data[n:]
			250	+	}
			251	+	if err != nil {
			252	+	return err
			253	+	}
			254	+	if n == 0 {
			255	+	return io.ErrShortWrite
			256	+	}
			257	+	}
			258	+	return nil
70		}	259		}
71			260
72		// Recv reads one frame: a 4-byte length, then a body of that many bytes	261		// Recv reads one frame: a 4-byte length, then a body of that many bytes
73		// (type byte + payload). It blocks until a full frame arrives, the	262		// (type byte + payload). It blocks until a full frame arrives, the
74		// connection is closed (io.EOF), or an error occurs.	263		// connection is closed (io.EOF), or an error occurs.
75		func (t *unixTransport) Recv() (ports.Frame, error) {	264		func (t *unixTransport) Recv() (ports.Frame, error) {
			265	+	end := t.beginOperation(ports.RuntimeAdapterReceiveStart, 0)
76		var hdr [frameHeaderLen]byte	266		var hdr [frameHeaderLen]byte
77		if _, err := io.ReadFull(t.conn, hdr[:]); err != nil {	267		if _, err := io.ReadFull(t.conn, hdr[:]); err != nil {
			268	+	end(false)
78		return ports.Frame{}, err	269		return ports.Frame{}, err
79		}	270		}
80			271
81		n := binary.BigEndian.Uint32(hdr[:])	272		n := binary.BigEndian.Uint32(hdr[:])
82		if n == 0 {	273		if n == 0 {
			274	+	end(false)
83		return ports.Frame{}, ErrZeroLengthFrame	275		return ports.Frame{}, ErrZeroLengthFrame
84		}	276		}
85		if n > maxFrameLen {	277		if n > maxFrameLen {
			278	+	end(false)
86		return ports.Frame{}, ErrFrameTooLarge	279		return ports.Frame{}, ErrFrameTooLarge
87		}	280		}
88			281

@@ -92,6 +285,7 @@ func (t *unixTransport) Recv() (ports.Frame, error) {
92		t.readBuf = t.readBuf[:n]	285		t.readBuf = t.readBuf[:n]
93		}	286		}
94		if _, err := io.ReadFull(t.conn, t.readBuf); err != nil {	287		if _, err := io.ReadFull(t.conn, t.readBuf); err != nil {
			288	+	end(false)
95		return ports.Frame{}, err	289		return ports.Frame{}, err
96		}	290		}
97			291

@@ -100,10 +294,75 @@ func (t *unixTransport) Recv() (ports.Frame, error) {
100		payload = make([]byte, n-1)	294		payload = make([]byte, n-1)
101		copy(payload, t.readBuf[1:])	295		copy(payload, t.readBuf[1:])
102		}	296		}
			297	+	end(true)
103		return ports.Frame{Type: ports.MsgType(t.readBuf[0]), Payload: payload}, nil	298		return ports.Frame{Type: ports.MsgType(t.readBuf[0]), Payload: payload}, nil
104		}	299		}
105			300
106	-	// Close closes the underlying connection.	301	+	func (t *unixTransport) beginOperation(start ports.RuntimeMarkKind, bytes uint64) func(bool) {
			302	+	if t.observer == nil {
			303	+	return func(bool) {}
			304	+	}
			305	+	if !t.beginObservedOperation() {
			306	+	return func(bool) {}
			307	+	}
			308	+	correlation := ports.NewRuntimeCorrelation()
			309	+	t.observer.ObserveRuntime(ports.NewRuntimeMarkWithCorrelation("ipc", correlation, start, bytes, true))
			310	+	end := ports.RuntimeAdapterSendEnd
			311	+	if start == ports.RuntimeAdapterReceiveStart {
			312	+	end = ports.RuntimeAdapterReceiveEnd
			313	+	}
			314	+	return func(valid bool) {
			315	+	defer t.finishObservedOperation()
			316	+	t.observer.ObserveRuntime(ports.NewRuntimeMarkWithCorrelation("ipc", correlation, end, bytes, valid))
			317	+	}
			318	+	}
			319	+
			320	+	func (t *unixTransport) beginObservedOperation() bool {
			321	+	t.operationMu.Lock()
			322	+	defer t.operationMu.Unlock()
			323	+	if t.observedClosing {
			324	+	return false
			325	+	}
			326	+	if t.operationCount == 0 {
			327	+	t.operationsDone = make(chan struct{})
			328	+	}
			329	+	t.operationCount++
			330	+	return true
			331	+	}
			332	+
			333	+	func (t *unixTransport) finishObservedOperation() {
			334	+	t.operationMu.Lock()
			335	+	defer t.operationMu.Unlock()
			336	+	t.operationCount--
			337	+	if t.operationCount == 0 {
			338	+	close(t.operationsDone)
			339	+	}
			340	+	}
			341	+
			342	+	func (t *unixTransport) beginShutdown() <-chan struct{} {
			343	+	t.operationMu.Lock()
			344	+	defer t.operationMu.Unlock()
			345	+	t.observedClosing = true
			346	+	if t.operationCount == 0 {
			347	+	return nil
			348	+	}
			349	+	return t.operationsDone
			350	+	}
			351	+
			352	+	// Close interrupts Recv and the writer's current Write, fails queued frames,
			353	+	// and waits until the sole writer and all observed operations have stopped.
107		func (t *unixTransport) Close() error {	354		func (t *unixTransport) Close() error {
108	-	return t.conn.Close()	355	+	t.closeOnce.Do(func() {
			356	+	operationsDone := t.beginShutdown()
			357	+	t.egressMu.Lock()
			358	+	t.egressClosing = true
			359	+	close(t.done)
			360	+	t.egressMu.Unlock()
			361	+	t.closeErr = t.conn.Close()
			362	+	<-t.writerDone
			363	+	if operationsDone != nil {
			364	+	<-operationsDone
			365	+	}
			366	+	})
			367	+	return t.closeErr
109		}	368		}

A internal/adapters/ipc/transport_observability_test.go +197 -0

@@ -0,0 +1,197 @@
	1	+	package ipc
	2	+
	3	+	import (
	4	+	"encoding/json"
	5	+	"net"
	6	+	"os"
	7	+	"reflect"
	8	+	"strings"
	9	+	"sync"
	10	+	"testing"
	11	+	"time"
	12	+
	13	+	"github.com/bnema/vev/internal/adapters/observability"
	14	+	"github.com/bnema/vev/internal/ports"
	15	+	)
	16	+
	17	+	func TestTransportObservabilityIPCEOFAndCloseEndFailedSpans(t *testing.T) {
	18	+	trace := t.TempDir() + "/ipc-shutdown.jsonl"
	19	+	observer, closer, err := observability.NewJSONL(trace, ipcRuntimeClock{now: time.Unix(0, 223)}, "ipc-process")
	20	+	if err != nil {
	21	+	t.Fatalf("NewJSONL() error = %v", err)
	22	+	}
	23	+	reporter := ports.NewSerializedRuntimeObserver(observer, 64)
	24	+	defer reporter.Close()
	25	+
	26	+	// EOF from the peer and a locally closed blocked receive are distinct
	27	+	// shutdown paths; both must record a failed, correlated end.
	28	+	for _, closeLocal := range []bool{false, true} {
	29	+	left, right := net.Pipe()
	30	+	var started <-chan struct{}
	31	+	conn := left
	32	+	if closeLocal {
	33	+	observedConn := &ipcReadStartedConn{Conn: left, started: make(chan struct{})}
	34	+	conn, started = observedConn, observedConn.started
	35	+	}
	36	+	transport := NewTransport(conn, WithRuntimeObserver(reporter))
	37	+	recvDone := make(chan error, 1)
	38	+	go func() { _, err := transport.Recv(); recvDone <- err }()
	39	+	if closeLocal {
	40	+	<-started
	41	+	if err := transport.Close(); err != nil {
	42	+	t.Fatalf("Close() error = %v", err)
	43	+	}
	44	+	} else if err := right.Close(); err != nil {
	45	+	t.Fatalf("peer Close() error = %v", err)
	46	+	}
	47	+	if err := <-recvDone; err == nil {
	48	+	t.Fatal("Recv() error = nil after transport shutdown")
	49	+	}
	50	+	_ = left.Close()
	51	+	_ = right.Close()
	52	+	}
	53	+
	54	+	// A peer shutdown also makes Send fail after its start mark.
	55	+	left, right := net.Pipe()
	56	+	transport := NewTransport(left, WithRuntimeObserver(reporter))
	57	+	_ = right.Close()
	58	+	if err := transport.Send(ports.Frame{Type: ports.MsgOutput, Payload: []byte("failed")}); err == nil {
	59	+	t.Fatal("Send() error = nil after peer close")
	60	+	}
	61	+	_ = transport.Close()
	62	+	reporter.Close()
	63	+	if err := closer.Close(); err != nil {
	64	+	t.Fatalf("trace Close() error = %v", err)
	65	+	}
	66	+	assertIPCFailedSpanPairs(t, trace)
	67	+	}
	68	+
	69	+	type ipcReadStartedConn struct {
	70	+	net.Conn
	71	+	started chan struct{}
	72	+	once sync.Once
	73	+	}
	74	+
	75	+	func (c *ipcReadStartedConn) Read(p []byte) (int, error) {
	76	+	c.once.Do(func() { close(c.started) })
	77	+	return c.Conn.Read(p)
	78	+	}
	79	+
	80	+	func assertIPCFailedSpanPairs(t *testing.T, path string) {
	81	+	t.Helper()
	82	+	data, err := os.ReadFile(path)
	83	+	if err != nil {
	84	+	t.Fatalf("ReadFile(trace): %v", err)
	85	+	}
	86	+	type mark struct {
	87	+	ProcessID string `json:"process_id"`
	88	+	Component string `json:"component"`
	89	+	Scenario string `json:"scenario"`
	90	+	Run uint64 `json:"run"`
	91	+	Sequence uint64 `json:"sequence"`
	92	+	RequestID uint64 `json:"request_id"`
	93	+	Epoch uint64 `json:"epoch"`
	94	+	Kind string `json:"kind"`
	95	+	Valid bool `json:"valid"`
	96	+	}
	97	+	type spanKey struct {
	98	+	processID, scenario string
	99	+	run, sequence, requestID, epoch uint64
	100	+	}
	101	+	starts := make(map[spanKey]mark)
	102	+	startCounts := map[string]int{}
	103	+	for _, line := range strings.Split(strings.TrimSpace(string(data)), "\n") {
	104	+	var m mark
	105	+	if err := json.Unmarshal([]byte(line), &m); err != nil {
	106	+	t.Fatalf("Unmarshal(trace): %v", err)
	107	+	}
	108	+	key := spanKey{m.ProcessID, m.Scenario, m.Run, m.Sequence, m.RequestID, m.Epoch}
	109	+	switch m.Kind {
	110	+	case "adapter_receive_start", "adapter_send_start":
	111	+	if m.ProcessID != "ipc-process" \|\| m.Component != "ipc" \|\| m.Scenario != "runtime" \|\| m.Run != 1 \|\| m.Sequence == 0 \|\| m.RequestID != m.Sequence \|\| m.Epoch != m.Sequence \|\| !m.Valid {
	112	+	t.Fatalf("adapter start has invalid identity: %#v", m)
	113	+	}
	114	+	if _, exists := starts[key]; exists {
	115	+	t.Fatalf("duplicate adapter start for %#v", key)
	116	+	}
	117	+	starts[key] = m
	118	+	startCounts[m.Kind]++
	119	+	case "adapter_receive_end", "adapter_send_end":
	120	+	start, ok := starts[key]
	121	+	if !ok \|\| strings.TrimSuffix(start.Kind, "_start") != strings.TrimSuffix(m.Kind, "_end") {
	122	+	t.Fatalf("unmatched adapter end %#v", m)
	123	+	}
	124	+	if m.Valid {
	125	+	t.Fatalf("failed adapter end is valid: %#v", m)
	126	+	}
	127	+	delete(starts, key)
	128	+	}
	129	+	}
	130	+	if len(starts) != 0 {
	131	+	t.Fatalf("unmatched adapter starts: %#v", starts)
	132	+	}
	133	+	if got := startCounts["adapter_receive_start"]; got != 2 {
	134	+	t.Errorf("adapter receive starts = %d, want 2", got)
	135	+	}
	136	+	if got := startCounts["adapter_send_start"]; got != 1 {
	137	+	t.Errorf("adapter send starts = %d, want 1", got)
	138	+	}
	139	+	}
	140	+
	141	+	func TestTransportObservabilityIPCMarksCarriageWithoutChangingBytes(t *testing.T) {
	142	+	trace := t.TempDir() + "/ipc.jsonl"
	143	+	observer, closer, err := observability.NewJSONL(trace, ipcRuntimeClock{now: time.Unix(0, 211)}, "ipc-process")
	144	+	if err != nil {
	145	+	t.Fatalf("NewJSONL() error = %v", err)
	146	+	}
	147	+	reporter := ports.NewSerializedRuntimeObserver(observer, 64)
	148	+	defer reporter.Close()
	149	+
	150	+	left, right := net.Pipe()
	151	+	defer func() { _ = left.Close() }()
	152	+	defer func() { _ = right.Close() }()
	153	+	client := NewTransport(left, WithRuntimeObserver(reporter))
	154	+	server := NewTransport(right, WithRuntimeObserver(reporter))
	155	+	want := ports.Frame{Type: ports.MsgOutput, Payload: []byte("wire bytes stay exact")}
	156	+
	157	+	var wg sync.WaitGroup
	158	+	wg.Go(func() {
	159	+	if err := client.Send(want); err != nil {
	160	+	t.Errorf("Send() error = %v", err)
	161	+	}
	162	+	})
	163	+	got, err := server.Recv()
	164	+	wg.Wait()
	165	+	if err != nil {
	166	+	t.Fatalf("Recv() error = %v", err)
	167	+	}
	168	+	if !reflect.DeepEqual(got, want) {
	169	+	t.Fatalf("carried frame = %#v, want %#v", got, want)
	170	+	}
	171	+	reporter.Close()
	172	+	if err := closer.Close(); err != nil {
	173	+	t.Fatalf("trace Close() error = %v", err)
	174	+	}
	175	+	traceBytes, err := os.ReadFile(trace)
	176	+	if err != nil {
	177	+	t.Fatalf("ReadFile(trace): %v", err)
	178	+	}
	179	+	for _, kind := range []string{"adapter_send_start", "adapter_send_end", "adapter_receive_start", "adapter_receive_end"} {
	180	+	if !strings.Contains(string(traceBytes), `"kind":"`+kind+`"`) {
	181	+	t.Errorf("trace missing %s", kind)
	182	+	}
	183	+	}
	184	+	}
	185	+
	186	+	type ipcRuntimeClock struct{ now time.Time }
	187	+
	188	+	func (c ipcRuntimeClock) Now() time.Time { return c.now }
	189	+	func (c ipcRuntimeClock) NewTimer(d time.Duration) ports.Timer {
	190	+	return ipcRuntimeTimer{timer: time.NewTimer(d)}
	191	+	}
	192	+
	193	+	type ipcRuntimeTimer struct{ timer *time.Timer }
	194	+
	195	+	func (t ipcRuntimeTimer) C() <-chan time.Time { return t.timer.C }
	196	+	func (t ipcRuntimeTimer) Reset(d time.Duration) bool { return t.timer.Reset(d) }
	197	+	func (t ipcRuntimeTimer) Stop() bool { return t.timer.Stop() }

A internal/adapters/ipc/transport_replay_test.go +40 -0

@@ -0,0 +1,40 @@
	1	+	package ipc
	2	+
	3	+	import (
	4	+	"net"
	5	+	"sync"
	6	+	"testing"
	7	+
	8	+	"github.com/bnema/vev/internal/ports"
	9	+	"github.com/bnema/vev/internal/testutil/replaytest"
	10	+	)
	11	+
	12	+	func TestTransportReplay(t *testing.T) {
	13	+	replaytest.Run(t, func(t *testing.T, frames []ports.Frame) []ports.Frame {
	14	+	left, right := net.Pipe()
	15	+	t.Cleanup(func() { _ = left.Close(); _ = right.Close() })
	16	+	sender, receiver := NewTransport(left), NewTransport(right)
	17	+
	18	+	var wg sync.WaitGroup
	19	+	wg.Go(func() {
	20	+	for _, frame := range frames {
	21	+	if err := sender.Send(frame); err != nil {
	22	+	t.Errorf("Send: %v", err)
	23	+	return
	24	+	}
	25	+	}
	26	+	})
	27	+	got := make([]ports.Frame, 0, len(frames))
	28	+	for range frames {
	29	+	frame, err := receiver.Recv()
	30	+	if err != nil {
	31	+	t.Errorf("Recv: %v", err)
	32	+	_ = receiver.Close()
	33	+	break
	34	+	}
	35	+	got = append(got, frame)
	36	+	}
	37	+	wg.Wait()
	38	+	return got
	39	+	})
	40	+	}

M internal/adapters/ipc/transport_test.go +285 -0


@@ -6,8 +6,10 @@ import (
6	"io"	6		"io"
7	"net"	7		"net"
8	"reflect"	8		"reflect"
		9	+	"runtime"
9	"sync"	10		"sync"
10	"testing"	11		"testing"
		12	+	"time"
11		13
12	"github.com/bnema/vev/internal/domain"	14		"github.com/bnema/vev/internal/domain"
13	"github.com/bnema/vev/internal/ports"	15		"github.com/bnema/vev/internal/ports"

@@ -231,3 +233,286 @@ func TestTransportSendOversizePayloadRejected(t *testing.T) {
231	t.Fatalf("client.Send() error = %v, want ErrFrameTooLarge", err)	233		t.Fatalf("client.Send() error = %v, want ErrFrameTooLarge", err)
232	}	234		}
233	}	235		}
		236	+
		237	+	func TestTransportAsyncEgressPreservesWelcomeBeforeOutput(t *testing.T) {
		238	+	const timeout = time.Second
		239	+
		240	+	serverConn, clientConn := net.Pipe()
		241	+	defer func() { _ = serverConn.Close() }()
		242	+	defer func() { _ = clientConn.Close() }()
		243	+	if err := serverConn.SetDeadline(time.Now().Add(timeout)); err != nil {
		244	+	t.Fatalf("SetDeadline server connection: %v", err)
		245	+	}
		246	+	if err := clientConn.SetDeadline(time.Now().Add(timeout)); err != nil {
		247	+	t.Fatalf("SetDeadline client connection: %v", err)
		248	+	}
		249	+
		250	+	server := NewTransport(serverConn)
		251	+	defer closeTransport(t, server)
		252	+	client := NewTransport(clientConn)
		253	+	defer closeTransport(t, client)
		254	+	async, ok := server.(ports.AsyncTransport)
		255	+	if !ok {
		256	+	t.Fatal("IPC transport does not implement AsyncTransport")
		257	+	}
		258	+
		259	+	welcome := ports.Frame{Type: ports.MsgWelcome, Payload: []byte("welcome")}
		260	+	sent := make(chan error, 1)
		261	+	go func() { sent <- server.Send(welcome) }()
		262	+	got, err := client.Recv()
		263	+	if err != nil {
		264	+	t.Fatalf("Recv welcome: %v", err)
		265	+	}
		266	+	if !reflect.DeepEqual(got, welcome) {
		267	+	t.Fatalf("welcome = %#v, want %#v", got, welcome)
		268	+	}
		269	+	select {
		270	+	case err := <-sent:
		271	+	if err != nil {
		272	+	t.Fatalf("Send welcome: %v", err)
		273	+	}
		274	+	case <-time.After(timeout):
		275	+	t.Fatal("Send welcome did not return")
		276	+	}
		277	+
		278	+	payload := []byte("first")
		279	+	if err := async.SendAsync(ports.Frame{Type: ports.MsgOutput, Payload: payload}); err != nil {
		280	+	t.Fatalf("SendAsync first output: %v", err)
		281	+	}
		282	+	payload[0] = 'X' // async ownership must not retain caller memory.
		283	+	if err := async.SendAsync(ports.Frame{Type: ports.MsgOutput, Payload: []byte("second")}); err != nil {
		284	+	t.Fatalf("SendAsync second output: %v", err)
		285	+	}
		286	+	for _, want := range []ports.Frame{
		287	+	{Type: ports.MsgOutput, Payload: []byte("first")},
		288	+	{Type: ports.MsgOutput, Payload: []byte("second")},
		289	+	} {
		290	+	got, err := client.Recv()
		291	+	if err != nil {
		292	+	t.Fatalf("Recv output: %v", err)
		293	+	}
		294	+	if !reflect.DeepEqual(got, want) {
		295	+	t.Fatalf("output = %#v, want %#v", got, want)
		296	+	}
		297	+	}
		298	+	}
		299	+
		300	+	func TestTransportSendWaitsForEgressCapacity(t *testing.T) {
		301	+	const timeout = time.Second
		302	+
		303	+	serverConn, clientConn := net.Pipe()
		304	+	defer func() { _ = serverConn.Close() }()
		305	+	defer func() { _ = clientConn.Close() }()
		306	+	if err := clientConn.SetDeadline(time.Now().Add(timeout)); err != nil {
		307	+	t.Fatalf("SetDeadline client connection: %v", err)
		308	+	}
		309	+
		310	+	server := NewTransport(serverConn)
		311	+	defer closeTransport(t, server)
		312	+	transport, ok := server.(*unixTransport)
		313	+	if !ok {
		314	+	t.Fatal("IPC transport is not a unixTransport")
		315	+	}
		316	+	client := NewTransport(clientConn)
		317	+	defer closeTransport(t, client)
		318	+	async, ok := server.(ports.AsyncTransport)
		319	+	if !ok {
		320	+	t.Fatal("IPC transport does not implement AsyncTransport")
		321	+	}
		322	+
		323	+	queued := fillAsyncEgress(t, transport, async, timeout)
		324	+	want := ports.Frame{Type: ports.MsgOutput, Payload: []byte("synchronous")}
		325	+	sent := make(chan error, 1)
		326	+	go func() { sent <- server.Send(want) }()
		327	+	waitForEgressSender(t, transport, timeout)
		328	+
		329	+	select {
		330	+	case err := <-sent:
		331	+	t.Fatalf("Send returned before capacity was available: %v", err)
		332	+	default:
		333	+	}
		334	+
		335	+	for _, want := range append(queued, want) {
		336	+	got, err := client.Recv()
		337	+	if err != nil {
		338	+	t.Fatalf("Recv: %v", err)
		339	+	}
		340	+	if !reflect.DeepEqual(got, want) {
		341	+	t.Fatalf("Recv = %#v, want %#v", got, want)
		342	+	}
		343	+	}
		344	+	select {
		345	+	case err := <-sent:
		346	+	if err != nil {
		347	+	t.Fatalf("Send after draining: %v", err)
		348	+	}
		349	+	case <-time.After(timeout):
		350	+	t.Fatal("Send did not return after draining")
		351	+	}
		352	+	}
		353	+
		354	+	func TestTransportCloseInterruptsSendWaitingForEgressCapacity(t *testing.T) {
		355	+	const timeout = time.Second
		356	+
		357	+	serverConn, clientConn := net.Pipe()
		358	+	defer func() { _ = clientConn.Close() }()
		359	+
		360	+	server := NewTransport(serverConn)
		361	+	defer closeTransport(t, server)
		362	+	transport, ok := server.(*unixTransport)
		363	+	if !ok {
		364	+	t.Fatal("IPC transport is not a unixTransport")
		365	+	}
		366	+	async, ok := server.(ports.AsyncTransport)
		367	+	if !ok {
		368	+	t.Fatal("IPC transport does not implement AsyncTransport")
		369	+	}
		370	+	fillAsyncEgress(t, transport, async, timeout)
		371	+
		372	+	sent := make(chan error, 1)
		373	+	go func() { sent <- server.Send(ports.Frame{Type: ports.MsgOutput, Payload: []byte("synchronous")}) }()
		374	+	waitForEgressSender(t, transport, timeout)
		375	+
		376	+	closeTransport(t, server)
		377	+	select {
		378	+	case err := <-sent:
		379	+	if !errors.Is(err, errClosed) {
		380	+	t.Fatalf("Send after Close = %v, want errClosed", err)
		381	+	}
		382	+	case <-time.After(timeout):
		383	+	t.Fatal("Close did not interrupt Send")
		384	+	}
		385	+	}
		386	+
		387	+	func waitForEgressSender(t testing.T, transport unixTransport, timeout time.Duration) {
		388	+	t.Helper()
		389	+
		390	+	timer := time.NewTimer(timeout)
		391	+	defer timer.Stop()
		392	+	for {
		393	+	transport.egressMu.Lock()
		394	+	senders := transport.egressSenders
		395	+	transport.egressMu.Unlock()
		396	+	if senders > 0 {
		397	+	return
		398	+	}
		399	+	select {
		400	+	case <-timer.C:
		401	+	t.Fatal("Send did not register as an egress capacity waiter")
		402	+	default:
		403	+	runtime.Gosched()
		404	+	}
		405	+	}
		406	+	}
		407	+
		408	+	func closeTransport(t *testing.T, transport ports.Transport) {
		409	+	t.Helper()
		410	+
		411	+	closed := make(chan error, 1)
		412	+	go func() { closed <- transport.Close() }()
		413	+	select {
		414	+	case err := <-closed:
		415	+	if err != nil {
		416	+	t.Fatalf("Close: %v", err)
		417	+	}
		418	+	case <-time.After(time.Second):
		419	+	t.Fatal("Close did not return")
		420	+	}
		421	+	}
		422	+
		423	+	func fillAsyncEgress(t testing.T, transport unixTransport, async ports.AsyncTransport, timeout time.Duration) []ports.Frame {
		424	+	t.Helper()
		425	+
		426	+	first := ports.Frame{Type: ports.MsgOutput, Payload: []byte{0}}
		427	+	if err := async.SendAsync(first); err != nil {
		428	+	t.Fatalf("SendAsync blocked frame: %v", err)
		429	+	}
		430	+	waitForEgressDequeue(t, transport, timeout)
		431	+
		432	+	queued := []ports.Frame{first}
		433	+	for i := 1; ; i++ {
		434	+	frame := ports.Frame{Type: ports.MsgOutput, Payload: []byte{byte(i)}}
		435	+	err := async.SendAsync(frame)
		436	+	switch {
		437	+	case err == nil:
		438	+	queued = append(queued, frame)
		439	+	case errors.Is(err, ErrBackpressure):
		440	+	return queued
		441	+	default:
		442	+	t.Fatalf("SendAsync queued frame: %v", err)
		443	+	}
		444	+	}
		445	+	}
		446	+
		447	+	func waitForEgressDequeue(t testing.T, transport unixTransport, timeout time.Duration) {
		448	+	t.Helper()
		449	+
		450	+	timer := time.NewTimer(timeout)
		451	+	defer timer.Stop()
		452	+	for {
		453	+	if len(transport.egress) == 0 {
		454	+	return
		455	+	}
		456	+	select {
		457	+	case <-timer.C:
		458	+	t.Fatal("IPC writer did not dequeue the blocked frame")
		459	+	default:
		460	+	runtime.Gosched()
		461	+	}
		462	+	}
		463	+	}
		464	+
		465	+	func TestTransportAsyncEgressIsBoundedAndCloseInterruptsWorkers(t *testing.T) {
		466	+	serverConn, clientConn := net.Pipe()
		467	+	defer func() { _ = clientConn.Close() }()
		468	+
		469	+	tr := NewTransport(serverConn)
		470	+	defer closeTransport(t, tr)
		471	+	transport, ok := tr.(*unixTransport)
		472	+	if !ok {
		473	+	t.Fatal("IPC transport is not a unixTransport")
		474	+	}
		475	+	async, ok := tr.(ports.AsyncTransport)
		476	+	if !ok {
		477	+	t.Fatal("IPC transport does not implement AsyncTransport")
		478	+	}
		479	+
		480	+	// The first write blocks in net.Pipe because the peer never drains it.
		481	+	if err := async.SendAsync(ports.Frame{Type: ports.MsgOutput, Payload: []byte("blocked")}); err != nil {
		482	+	t.Fatalf("SendAsync blocked frame: %v", err)
		483	+	}
		484	+	var backpressure error
		485	+	for range sendQueueCapacity + 2 {
		486	+	err := async.SendAsync(ports.Frame{Type: ports.MsgOutput, Payload: []byte("queued")})
		487	+	if errors.Is(err, ErrBackpressure) {
		488	+	backpressure = err
		489	+	break
		490	+	}
		491	+	if err != nil {
		492	+	t.Fatalf("SendAsync unexpected error: %v", err)
		493	+	}
		494	+	}
		495	+	if !errors.Is(backpressure, ErrBackpressure) {
		496	+	t.Fatalf("SendAsync never reported bounded backpressure, got %v", backpressure)
		497	+	}
		498	+
		499	+	recvDone := make(chan error, 1)
		500	+	go func() {
		501	+	_, err := tr.Recv()
		502	+	recvDone <- err
		503	+	}()
		504	+	closeTransport(t, tr)
		505	+	select {
		506	+	case err := <-recvDone:
		507	+	if err == nil {
		508	+	t.Fatal("Recv succeeded after Close")
		509	+	}
		510	+	case <-time.After(time.Second):
		511	+	t.Fatal("Close did not interrupt Recv")
		512	+	}
		513	+	select {
		514	+	case <-transport.writerDone:
		515	+	default:
		516	+	t.Fatal("Close returned before the IPC writer terminated")
		517	+	}
		518	+	}

A internal/adapters/observability/jsonl.go +100 -0

@@ -0,0 +1,100 @@
	1	+	// Package observability contains concrete process-local trace sinks.
	2	+	package observability
	3	+
	4	+	import (
	5	+	"encoding/json"
	6	+	"io"
	7	+	"os"
	8	+	"sync"
	9	+
	10	+	"github.com/bnema/vev/internal/ports"
	11	+	)
	12	+
	13	+	// NewJSONL opens path as the sole timestamp owner for one OS process. Marks
	14	+	// carrying a caller timestamp are rejected rather than rewritten, preventing
	15	+	// accidental arithmetic across clock domains.
	16	+	func NewJSONL(path string, clock ports.Clock, processID string) (ports.RuntimeObserver, io.Closer, error) {
	17	+	if path == "" \|\| clock == nil \|\| processID == "" {
	18	+	return nil, nil, os.ErrInvalid
	19	+	}
	20	+	f, err := os.OpenFile(path, os.O_WRONLY\|os.O_CREATE\|os.O_APPEND, 0o600)
	21	+	if err != nil {
	22	+	return nil, nil, err
	23	+	}
	24	+	j := &jsonl{file: f, clock: clock, processID: processID}
	25	+	return j, j, nil
	26	+	}
	27	+
	28	+	type jsonl struct {
	29	+	mu sync.Mutex
	30	+	file *os.File
	31	+	clock ports.Clock
	32	+	processID string
	33	+	closed bool
	34	+	err error
	35	+	}
	36	+
	37	+	type record struct {
	38	+	Schema uint16 `json:"schema"`
	39	+	ProcessID string `json:"process_id"`
	40	+	Component string `json:"component"`
	41	+	Scenario string `json:"scenario"`
	42	+	Run uint64 `json:"run"`
	43	+	Sequence uint64 `json:"sequence"`
	44	+	RequestID uint64 `json:"request_id"`
	45	+	Epoch uint64 `json:"epoch"`
	46	+	Kind ports.RuntimeMarkKind `json:"kind"`
	47	+	Tick int64 `json:"tick"`
	48	+	Bytes uint64 `json:"bytes"`
	49	+	Fragments uint64 `json:"fragments"`
	50	+	Retransmits uint64 `json:"retransmits"`
	51	+	Pending uint64 `json:"pending"`
	52	+	AckRTTNanos int64 `json:"ack_rtt_nanos"`
	53	+	Valid bool `json:"valid"`
	54	+	}
	55	+
	56	+	func (j *jsonl) ObserveRuntime(mark ports.RuntimeMark) {
	57	+	// Tick is a write-once value owned by this adapter. Reject it before any
	58	+	// mutation so callers cannot smuggle a timestamp from their own clock.
	59	+	if j == nil \|\| mark.Tick != 0 {
	60	+	return
	61	+	}
	62	+	mark.ProcessID = j.processID
	63	+	if !ports.RuntimeMarkValid(mark) {
	64	+	return
	65	+	}
	66	+	j.mu.Lock()
	67	+	defer j.mu.Unlock()
	68	+	if j.closed {
	69	+	return
	70	+	}
	71	+	r := record{Schema: mark.Schema, ProcessID: mark.ProcessID, Component: mark.Component, Scenario: mark.Scenario, Run: mark.Run, Sequence: mark.Sequence, RequestID: mark.RequestID, Epoch: mark.Epoch, Kind: mark.Kind, Tick: j.clock.Now().UnixNano(), Bytes: mark.Bytes, Fragments: mark.Fragments, Retransmits: mark.Retransmits, Pending: mark.Pending, AckRTTNanos: mark.AckRTTNanos, Valid: mark.Valid}
	72	+	encoded, err := json.Marshal(r)
	73	+	if err != nil {
	74	+	j.recordError(err)
	75	+	return
	76	+	}
	77	+	if _, err := j.file.Write(append(encoded, '\n')); err != nil {
	78	+	j.recordError(err)
	79	+	}
	80	+	}
	81	+
	82	+	func (j *jsonl) recordError(err error) {
	83	+	if err != nil && j.err == nil {
	84	+	j.err = err
	85	+	}
	86	+	}
	87	+
	88	+	func (j *jsonl) Close() error {
	89	+	j.mu.Lock()
	90	+	defer j.mu.Unlock()
	91	+	if j.closed {
	92	+	return nil
	93	+	}
	94	+	j.closed = true
	95	+	closeErr := j.file.Close()
	96	+	if j.err != nil {
	97	+	return j.err
	98	+	}
	99	+	return closeErr
	100	+	}

A internal/adapters/observability/jsonl_test.go +193 -0

@@ -0,0 +1,193 @@
	1	+	package observability
	2	+
	3	+	import (
	4	+	"bufio"
	5	+	"encoding/json"
	6	+	"os"
	7	+	"sync"
	8	+	"testing"
	9	+	"time"
	10	+
	11	+	"github.com/bnema/vev/internal/ports"
	12	+	)
	13	+
	14	+	func TestTransportObservabilityJSONLOwnsTicksAndSchema(t *testing.T) {
	15	+	path := t.TempDir() + "/daemon.jsonl"
	16	+	observer, closer, err := NewJSONL(path, fixedRuntimeClock{now: time.Unix(0, 41)}, "daemon-process")
	17	+	if err != nil {
	18	+	t.Fatalf("NewJSONL() error = %v", err)
	19	+	}
	20	+
	21	+	observer.ObserveRuntime(validRuntimeMark(ports.RuntimeComposeStart))
	22	+	// A producer's clock domain must never enter a JSONL record. Rejection is
	23	+	// observable as no second record, rather than a silently rewritten mark.
	24	+	producerTicked := validRuntimeMark(ports.RuntimeComposeEnd)
	25	+	producerTicked.Tick = 99
	26	+	observer.ObserveRuntime(producerTicked)
	27	+	if err := closer.Close(); err != nil {
	28	+	t.Fatalf("Close() error = %v", err)
	29	+	}
	30	+
	31	+	records := readJSONLRecords(t, path)
	32	+	if len(records) != 1 {
	33	+	t.Fatalf("JSONL records = %d, want 1 (caller Tick must be rejected)", len(records))
	34	+	}
	35	+	wantKeys := map[string]bool{
	36	+	"schema": true, "process_id": true, "component": true, "scenario": true,
	37	+	"run": true, "sequence": true, "request_id": true, "epoch": true,
	38	+	"kind": true, "tick": true, "bytes": true, "fragments": true,
	39	+	"retransmits": true, "pending": true, "ack_rtt_nanos": true, "valid": true,
	40	+	}
	41	+	if len(records[0]) != len(wantKeys) {
	42	+	t.Fatalf("JSONL fields = %#v, want exactly canonical schema", records[0])
	43	+	}
	44	+	for key := range wantKeys {
	45	+	if _, ok := records[0][key]; !ok {
	46	+	t.Errorf("JSONL missing %q", key)
	47	+	}
	48	+	}
	49	+	if got := records[0]["tick"]; got != float64(41) {
	50	+	t.Errorf("JSONL tick = %v, want observer clock tick 41", got)
	51	+	}
	52	+	if got := records[0]["process_id"]; got != "daemon-process" {
	53	+	t.Errorf("JSONL process_id = %v, want observer process", got)
	54	+	}
	55	+	}
	56	+
	57	+	func TestTransportObservabilityJSONLRejectsInvalidRequiredFields(t *testing.T) {
	58	+	for _, mark := range []ports.RuntimeMark{
	59	+	{},
	60	+	{Schema: 1, ProcessID: "p", Component: "daemon", Scenario: "s", Run: 1, Sequence: 1, Kind: ports.RuntimeEmitStart},
	61	+	{Schema: 2, ProcessID: "p", Component: "daemon", Scenario: "s", Run: 1, Sequence: 1, Kind: ports.RuntimeEmitStart, Valid: true},
	62	+	} {
	63	+	path := t.TempDir() + "/invalid.jsonl"
	64	+	observer, closer, err := NewJSONL(path, fixedRuntimeClock{now: time.Unix(0, 5)}, "p")
	65	+	if err != nil {
	66	+	t.Fatalf("NewJSONL() error = %v", err)
	67	+	}
	68	+	observer.ObserveRuntime(mark)
	69	+	if err := closer.Close(); err != nil {
	70	+	t.Fatalf("Close() error = %v", err)
	71	+	}
	72	+	if records := readJSONLRecords(t, path); len(records) != 0 {
	73	+	t.Errorf("invalid mark %#v wrote %#v", mark, records)
	74	+	}
	75	+	}
	76	+	}
	77	+
	78	+	func TestTransportObservabilityJSONLWriteErrorSurfacesOnClose(t *testing.T) {
	79	+	// /dev/full accepts Open but rejects every write, exercising the concrete
	80	+	// observer lifecycle without changing the one-way RuntimeObserver port.
	81	+	observer, closer, err := NewJSONL("/dev/full", fixedRuntimeClock{now: time.Unix(0, 1)}, "p")
	82	+	if err != nil {
	83	+	t.Skipf("/dev/full unavailable: %v", err)
	84	+	}
	85	+	observer.ObserveRuntime(validRuntimeMark(ports.RuntimeEmitStart))
	86	+	if err := closer.Close(); err == nil {
	87	+	t.Fatal("Close() error = nil, want surfaced JSONL write error")
	88	+	}
	89	+	}
	90	+
	91	+	func TestTransportObservabilitySameProcessSpanClockDomains(t *testing.T) {
	92	+	for _, tc := range []struct {
	93	+	name string
	94	+	start ports.RuntimeMarkKind
	95	+	end ports.RuntimeMarkKind
	96	+	field string
	97	+	}{
	98	+	{"diff", ports.RuntimeDiffStart, ports.RuntimeDiffEnd, "diff_duration"},
	99	+	{"queue", ports.RuntimeQueueEnqueued, ports.RuntimeQueueDequeued, "queue_wait"},
	100	+	{"ack", ports.RuntimeACKBlockedStart, ports.RuntimeACKBlockedEnd, "ack_blocked_interval"},
	101	+	{"send", ports.RuntimeAdapterSendStart, ports.RuntimeAdapterSendEnd, "adapter_send_duration"},
	102	+	{"receive", ports.RuntimeAdapterReceiveStart, ports.RuntimeAdapterReceiveEnd, "adapter_receive_duration"},
	103	+	} {
	104	+	t.Run(tc.name, func(t *testing.T) {
	105	+	start := validRuntimeMark(tc.start)
	106	+	start.Tick = 11
	107	+	end := start
	108	+	end.Kind, end.Tick = tc.end, 19
	109	+	if got, err := sameProcessDuration(start, end); err != nil \|\| got != 8 {
	110	+	t.Fatalf("%s = %d, %v; want 8, nil", tc.field, got, err)
	111	+	}
	112	+
	113	+	end.ProcessID = "other-process"
	114	+	if _, err := sameProcessDuration(start, end); err == nil {
	115	+	t.Fatalf("%s accepted cross-process ticks", tc.field)
	116	+	}
	117	+	})
	118	+	}
	119	+	}
	120	+
	121	+	func TestTransportObservabilityJSONLSerializesConcurrentMarks(t *testing.T) {
	122	+	path := t.TempDir() + "/concurrent.jsonl"
	123	+	observer, closer, err := NewJSONL(path, fixedRuntimeClock{now: time.Unix(0, 73)}, "ipc-process")
	124	+	if err != nil {
	125	+	t.Fatalf("NewJSONL() error = %v", err)
	126	+	}
	127	+	var wg sync.WaitGroup
	128	+	for i := uint64(0); i < 32; i++ {
	129	+	wg.Add(1)
	130	+	go func(sequence uint64) {
	131	+	defer wg.Done()
	132	+	mark := validRuntimeMark(ports.RuntimeAdapterSendEnd)
	133	+	mark.Sequence = sequence + 1
	134	+	observer.ObserveRuntime(mark)
	135	+	}(i)
	136	+	}
	137	+	wg.Wait()
	138	+	if err := closer.Close(); err != nil {
	139	+	t.Fatalf("Close() error = %v", err)
	140	+	}
	141	+	if records := readJSONLRecords(t, path); len(records) != 32 {
	142	+	t.Fatalf("concurrently serialized records = %d, want 32", len(records))
	143	+	}
	144	+	}
	145	+
	146	+	func sameProcessDuration(start, end ports.RuntimeMark) (int64, error) {
	147	+	if start.ProcessID == "" \|\| start.ProcessID != end.ProcessID {
	148	+	return 0, os.ErrInvalid
	149	+	}
	150	+	if start.Scenario != end.Scenario \|\| start.Run != end.Run \|\| start.Sequence != end.Sequence \|\| start.RequestID != end.RequestID \|\| start.Epoch != end.Epoch \|\| end.Tick < start.Tick {
	151	+	return 0, os.ErrInvalid
	152	+	}
	153	+	return end.Tick - start.Tick, nil
	154	+	}
	155	+
	156	+	func validRuntimeMark(kind ports.RuntimeMarkKind) ports.RuntimeMark {
	157	+	return ports.RuntimeMark{Schema: 1, ProcessID: "daemon-process", Component: "daemon", Scenario: "resize", Run: 7, Sequence: 9, RequestID: 11, Epoch: 13, Kind: kind, Valid: true}
	158	+	}
	159	+
	160	+	func readJSONLRecords(t *testing.T, path string) []map[string]any {
	161	+	t.Helper()
	162	+	f, err := os.Open(path)
	163	+	if err != nil {
	164	+	t.Fatalf("Open(%q): %v", path, err)
	165	+	}
	166	+	defer func() { _ = f.Close() }()
	167	+	var records []map[string]any
	168	+	scanner := bufio.NewScanner(f)
	169	+	for scanner.Scan() {
	170	+	var record map[string]any
	171	+	if err := json.Unmarshal(scanner.Bytes(), &record); err != nil {
	172	+	t.Fatalf("invalid JSONL record %q: %v", scanner.Text(), err)
	173	+	}
	174	+	records = append(records, record)
	175	+	}
	176	+	if err := scanner.Err(); err != nil {
	177	+	t.Fatalf("reading JSONL: %v", err)
	178	+	}
	179	+	return records
	180	+	}
	181	+
	182	+	type fixedRuntimeClock struct{ now time.Time }
	183	+
	184	+	func (c fixedRuntimeClock) Now() time.Time { return c.now }
	185	+	func (c fixedRuntimeClock) NewTimer(d time.Duration) ports.Timer {
	186	+	return fixedRuntimeTimer{timer: time.NewTimer(d)}
	187	+	}
	188	+
	189	+	type fixedRuntimeTimer struct{ timer *time.Timer }
	190	+
	191	+	func (t fixedRuntimeTimer) C() <-chan time.Time { return t.timer.C }
	192	+	func (t fixedRuntimeTimer) Reset(d time.Duration) bool { return t.timer.Reset(d) }
	193	+	func (t fixedRuntimeTimer) Stop() bool { return t.timer.Stop() }

M internal/adapters/pty/pty_linux.go +30 -6

 // x/sys/unix plus the standard library.
 import (
+	"context"
 	"errors"
 	"fmt"
 	"io"
 // verbatim (nil means inherit the current process environment); the caller
 // decides TERM and friends. sz sets the terminal window size before the child
 // starts, so the child observes the correct dimensions on its first query.
-func (Factory) Open(command string, args []string, env []string, dir string, sz domain.Size) (ports.PTY, error) {
+func (Factory) Open(ctx context.Context, command string, args []string, env []string, dir string, sz domain.Size) (ports.PTY, error) {
+	if err := ctx.Err(); err != nil {
+		return nil, err
+	}
 	masterFd, err := unix.Open("/dev/ptmx", unix.O_RDWR|unix.O_NOCTTY|unix.O_CLOEXEC, 0)
 	if err != nil {
 		return nil, fmt.Errorf("pty: open /dev/ptmx: %w", err)
 		}
 	}
-	cmd := exec.Command(command, args...)
+	cmd := exec.CommandContext(ctx, command, args...)
 	cmd.Env = env
 	cmd.Dir = platform.DirOrHome(dir)
 	cmd.Stdin = slave
 		Setctty: true,
 		Ctty:    0,
 	}
+	// CommandContext's default cancellation kills only the direct child. The
+	// child leads a process group, so cancellation must target that group too.
+	cmd.Cancel = func() error {
+		if cmd.Process == nil {
+			return os.ErrProcessDone
+		}
+		if err := signalProcessGroup(cmd.Process.Pid, syscall.SIGKILL); err != nil {
+			if errors.Is(err, syscall.ESRCH) {
+				return os.ErrProcessDone
+			}
+			return err
+		}
+		return nil
+	}
 	if err := cmd.Start(); err != nil {
 		return nil, fmt.Errorf("pty: start %q: %w", command, err)
 	p.closeOnce.Do(func() {
 		pid := p.cmd.Process.Pid
-		// Negative pid targets the whole process group (the child is the group
+		// Ignore errors: the child may already be gone.
-		// leader thanks to Setsid). Ignore errors: the child may already be gone.
+		_ = signalProcessGroup(pid, syscall.SIGHUP)
-		_ = unix.Kill(-pid, unix.SIGHUP)
 		// Reap the child (exit status is intentionally discarded; the port has no
 		// Wait — "child gone" surfaces to callers as io.EOF on Read).
 		select {
 		case <-reaped:
 		case <-time.After(killGracePeriod):
-			_ = unix.Kill(-pid, unix.SIGKILL)
+			_ = signalProcessGroup(pid, syscall.SIGKILL)
 			<-reaped
 		}
 	return p.closeErr
 }
+// signalProcessGroup signals the process group led by pid. Open creates each
+// child as its own session and process-group leader, so a negative pid reaches
+// both the command and every descendant that remains in its process group.
+func signalProcessGroup(pid int, signal syscall.Signal) error {
+	return unix.Kill(-pid, signal)
+}
 // setWinsize applies sz to the terminal referenced by fd via TIOCSWINSZ.
 func setWinsize(fd int, sz domain.Size) error {
 	return unix.IoctlSetWinsize(fd, unix.TIOCSWINSZ, &unix.Winsize{

M internal/adapters/pty/pty_linux_test.go +44 -7

 import (
 	"bytes"
+	"context"
 	"errors"
 	"io"
 	"os"
 		t.Skip("skipping pty integration test in -short mode")
 	}
 	f := newFactory()
-	p, err := f.Open("sh", []string{"-c", "printf hello"}, os.Environ(), "", domain.Size{Cols: 80, Rows: 24})
+	p, err := f.Open(context.Background(), "sh", []string{"-c", "printf hello"}, os.Environ(), "", domain.Size{Cols: 80, Rows: 24})
 	require.NoError(t, err)
 	t.Cleanup(func() { _ = p.Close() })
 		t.Skip("skipping pty integration test in -short mode")
 	}
 	f := newFactory()
-	p, err := f.Open("cat", nil, os.Environ(), "", domain.Size{Cols: 80, Rows: 24})
+	p, err := f.Open(context.Background(), "cat", nil, os.Environ(), "", domain.Size{Cols: 80, Rows: 24})
 	require.NoError(t, err)
 	t.Cleanup(func() { _ = p.Close() })
 		t.Skip("skipping pty integration test in -short mode")
 	}
 	f := newFactory()
-	p, err := f.Open("sh", []string{"-c", "sleep 2"}, os.Environ(), "", domain.Size{Cols: 80, Rows: 24})
+	p, err := f.Open(context.Background(), "sh", []string{"-c", "sleep 2"}, os.Environ(), "", domain.Size{Cols: 80, Rows: 24})
 	require.NoError(t, err)
 	t.Cleanup(func() { _ = p.Close() })
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
 			f := newFactory()
-			p, err := f.Open("sh", []string{"-c", tt.script}, os.Environ(), "", tt.initial)
+			p, err := f.Open(context.Background(), "sh", []string{"-c", tt.script}, os.Environ(), "", tt.initial)
 			require.NoError(t, err)
 			t.Cleanup(func() { _ = p.Close() })
 		t.Skip("skipping pty integration test in -short mode")
 	}
 	f := newFactory()
-	p, err := f.Open("sh", []string{"-c", "sleep 30"}, os.Environ(), "", domain.Size{Cols: 80, Rows: 24})
+	p, err := f.Open(context.Background(), "sh", []string{"-c", "sleep 30"}, os.Environ(), "", domain.Size{Cols: 80, Rows: 24})
 	require.NoError(t, err)
 	pid := p.Pid()
 		t.Skip("skipping pty integration test in -short mode")
 	}
 	f := newFactory()
-	p, err := f.Open("sh", []string{"-c", "sleep 60"}, os.Environ(), "", domain.Size{Cols: 80, Rows: 24})
+	p, err := f.Open(context.Background(), "sh", []string{"-c", "sleep 60"}, os.Environ(), "", domain.Size{Cols: 80, Rows: 24})
 	require.NoError(t, err)
 	pid := p.Pid()
 	}, 3*time.Second, 20*time.Millisecond)
 }
+func TestOpen_ContextCancellationTerminatesProcessGroup(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping pty integration test in -short mode")
+	}
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	// The background sleep ignores SIGHUP so killing only the session leader
+	// cannot accidentally pass via the terminal's hangup behavior.
+	p, err := newFactory().Open(ctx, "sh", []string{"-c", `(trap '' HUP; exec sleep 60) & child=$!; printf '%s\n' "$child"; wait "$child"`}, os.Environ(), "", domain.Size{Cols: 80, Rows: 24})
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = p.Close() })
+	buf := make([]byte, 64)
+	n, err := p.Read(buf)
+	require.NoError(t, err)
+	descendantPID, err := strconv.Atoi(strings.TrimSpace(string(buf[:n])))
+	require.NoError(t, err)
+	cancel()
+	// The direct shell may remain as a zombie until Close reaps it; a zombie has
+	// exited, so it and the background sleep must not remain running.
+	require.Eventually(t, func() bool {
+		return processTerminated(p.Pid()) && processTerminated(descendantPID)
+	}, 3*time.Second, 20*time.Millisecond)
+}
+func processTerminated(pid int) bool {
+	if errors.Is(syscall.Kill(pid, 0), syscall.ESRCH) {
+		return true
+	}
+	status, err := os.ReadFile("/proc/" + strconv.Itoa(pid) + "/status")
+	return err != nil || strings.Contains(string(status), "State:\tZ")
+}
 func TestResize_DeliversSIGWINCH(t *testing.T) {
 	if testing.Short() {
 		t.Skip("skipping pty integration test in -short mode")
 	// deliver SIGWINCH to the foreground process group.
 	script := `trap 'echo GOTWINCH; exit 0' WINCH; echo READY; i=0; while [ $i -lt 200 ]; do sleep 0.05; i=$((i+1)); done`
 	f := newFactory()
-	p, err := f.Open("sh", []string{"-c", script}, os.Environ(), "", domain.Size{Cols: 80, Rows: 24})
+	p, err := f.Open(context.Background(), "sh", []string{"-c", script}, os.Environ(), "", domain.Size{Cols: 80, Rows: 24})
 	require.NoError(t, err)
 	t.Cleanup(func() { _ = p.Close() })

M internal/adapters/remote/remote_factory.go +19 -8

 )
 // DialerFactory selects the requested remote transport adapter.
-type DialerFactory struct{}
+type DialerFactory struct {
+	observer ports.SerializedRuntimeObserver
+}
 var _ ports.RemoteDialerFactory = DialerFactory{}
 func NewDialerFactory() DialerFactory { return DialerFactory{} }
-func (DialerFactory) DialerForRemote(target string, session string, mode ports.RemoteTransportMode, log *slog.Logger) (ports.Dialer, error) {
+// NewDialerFactoryWithRuntimeObserver wires the one process-local sink into
+// whichever concrete remote carriage adapter is selected.
+func NewDialerFactoryWithRuntimeObserver(observer ports.SerializedRuntimeObserver) DialerFactory {
+	return DialerFactory{observer: observer}
+}
+func (f DialerFactory) DialerForRemote(target string, session string, mode ports.RemoteTransportMode, log *slog.Logger) (ports.Dialer, error) {
 	switch mode {
 	case ports.RemoteTransportUDP:
 		if log != nil {
 			log.Info("remote transport selected", "mode", mode, "target", target, "session", session)
 		}
-		return dgram.NewRemoteDialerWithLogger(target, session, log), nil
+		dialer := dgram.NewRemoteDialerWithLogger(target, session, log)
+		dialer.RuntimeObserver = f.observer
+		return dialer, nil
 	case ports.RemoteTransportStdio:
 		if log != nil {
 			log.Info("remote transport selected", "mode", mode, "target", target, "session", session)
 		}
-		return stdioDialer{target: target, session: session, log: log}, nil
+		return stdioDialer{target: target, session: session, log: log, observer: f.observer}, nil
 	default:
 		return nil, fmt.Errorf("vev: unsupported remote transport %q", mode)
 	}
 }
 type stdioDialer struct {
-	target  string
+	target   string
-	session string
+	session  string
-	log     *slog.Logger
+	log      *slog.Logger
+	observer ports.SerializedRuntimeObserver
 }
 func (d stdioDialer) Dial(ctx context.Context) (ports.Transport, error) {
-	return sshstdio.DialContext(ctx, d.target, d.session, d.log)
+	return sshstdio.DialContextWithRuntimeObserver(ctx, d.target, d.session, d.log, sshstdio.WithRuntimeObserver(d.observer))
 }

M internal/adapters/sshstdio/transport.go +128 -15


@@ -34,12 +34,26 @@ var (
34		type closeFunc func() error	34		type closeFunc func() error
35		type eofErrFunc func() error	35		type eofErrFunc func() error
36			36
			37	+	type Option func(*transport)
			38	+
			39	+	// WithRuntimeObserver enables process-local transport marks; it deliberately
			40	+	// takes only an observer so a carriage adapter never owns trace time.
			41	+	func WithRuntimeObserver(observer ports.SerializedRuntimeObserver) Option {
			42	+	return func(t *transport) { t.observer = observer }
			43	+	}
			44	+
37		// NewTransport wraps separate reader/writer streams as a framed Transport.	45		// NewTransport wraps separate reader/writer streams as a framed Transport.
38	-	func NewTransport(r io.Reader, w io.Writer, closeFn closeFunc) ports.Transport {	46	+	func NewTransport(r io.Reader, w io.Writer, closeFn closeFunc, opts ...Option) ports.Transport {
39	-	return newTransport(r, w, closeFn, nil)	47	+	t := newTransport(r, w, closeFn, nil)
			48	+	for _, opt := range opts {
			49	+	if opt != nil {
			50	+	opt(t)
			51	+	}
			52	+	}
			53	+	return t
40		}	54		}
41			55
42	-	func newTransport(r io.Reader, w io.Writer, closeFn closeFunc, eofErr eofErrFunc) ports.Transport {	56	+	func newTransport(r io.Reader, w io.Writer, closeFn closeFunc, eofErr eofErrFunc) *transport {
43		if closeFn == nil {	57		if closeFn == nil {
44		closeFn = func() error { return nil }	58		closeFn = func() error { return nil }
45		}	59		}

@@ -47,18 +61,25 @@ func newTransport(r io.Reader, w io.Writer, closeFn closeFunc, eofErr eofErrFunc
47		}	61		}
48			62
49		type transport struct {	63		type transport struct {
50	-	r io.Reader	64	+	r io.Reader
51	-	w io.Writer	65	+	w io.Writer
52	-	close closeFunc	66	+	close closeFunc
53	-	eofErr eofErrFunc	67	+	eofErr eofErrFunc
			68	+	observer ports.SerializedRuntimeObserver
			69	+	operationMu sync.Mutex
			70	+	operationCount int
			71	+	closing bool
			72	+	operationsDone chan struct{}
54			73
55		mu sync.Mutex	74		mu sync.Mutex
56		readBuf []byte	75		readBuf []byte
57		}	76		}
58			77
59		func (t *transport) Send(f ports.Frame) error {	78		func (t *transport) Send(f ports.Frame) error {
			79	+	end := t.beginOperation(ports.RuntimeAdapterSendStart, uint64(len(f.Payload)))
60		n := 1 + len(f.Payload)	80		n := 1 + len(f.Payload)
61		if n > maxFrameLen {	81		if n > maxFrameLen {
			82	+	end(false)
62		return ErrFrameTooLarge	83		return ErrFrameTooLarge
63		}	84		}
64			85

@@ -68,22 +89,28 @@ func (t *transport) Send(f ports.Frame) error {
68		copy(buf[frameHeaderLen+1:], f.Payload)	89		copy(buf[frameHeaderLen+1:], f.Payload)
69			90
70		t.mu.Lock()	91		t.mu.Lock()
71	-	defer t.mu.Unlock()
72		_, err := t.w.Write(buf)	92		_, err := t.w.Write(buf)
			93	+	t.mu.Unlock()
			94	+	end(err == nil)
73		return err	95		return err
74		}	96		}
75			97
76		func (t *transport) Recv() (ports.Frame, error) {	98		func (t *transport) Recv() (ports.Frame, error) {
			99	+	end := t.beginOperation(ports.RuntimeAdapterReceiveStart, 0)
77		var hdr [frameHeaderLen]byte	100		var hdr [frameHeaderLen]byte
78		if _, err := io.ReadFull(t.r, hdr[:]); err != nil {	101		if _, err := io.ReadFull(t.r, hdr[:]); err != nil {
79	-	return ports.Frame{}, t.mapEOFError(err)	102	+	err = t.mapEOFError(err)
			103	+	end(false)
			104	+	return ports.Frame{}, err
80		}	105		}
81			106
82		n := binary.BigEndian.Uint32(hdr[:])	107		n := binary.BigEndian.Uint32(hdr[:])
83		if n == 0 {	108		if n == 0 {
			109	+	end(false)
84		return ports.Frame{}, ErrZeroLengthFrame	110		return ports.Frame{}, ErrZeroLengthFrame
85		}	111		}
86		if n > maxFrameLen {	112		if n > maxFrameLen {
			113	+	end(false)
87		return ports.Frame{}, ErrFrameTooLarge	114		return ports.Frame{}, ErrFrameTooLarge
88		}	115		}
89			116

@@ -93,13 +120,35 @@ func (t *transport) Recv() (ports.Frame, error) {
93		t.readBuf = t.readBuf[:n]	120		t.readBuf = t.readBuf[:n]
94		}	121		}
95		if _, err := io.ReadFull(t.r, t.readBuf); err != nil {	122		if _, err := io.ReadFull(t.r, t.readBuf); err != nil {
96	-	return ports.Frame{}, t.mapEOFError(err)	123	+	err = t.mapEOFError(err)
			124	+	end(false)
			125	+	return ports.Frame{}, err
97		}	126		}
98			127
99		payload := append([]byte(nil), t.readBuf[1:]...)	128		payload := append([]byte(nil), t.readBuf[1:]...)
			129	+	end(true)
100		return ports.Frame{Type: ports.MsgType(t.readBuf[0]), Payload: payload}, nil	130		return ports.Frame{Type: ports.MsgType(t.readBuf[0]), Payload: payload}, nil
101		}	131		}
102			132
			133	+	func (t *transport) beginOperation(start ports.RuntimeMarkKind, bytes uint64) func(bool) {
			134	+	if t.observer == nil {
			135	+	return func(bool) {}
			136	+	}
			137	+	if !t.beginObservedOperation() {
			138	+	return func(bool) {}
			139	+	}
			140	+	correlation := ports.NewRuntimeCorrelation()
			141	+	t.observer.ObserveRuntime(ports.NewRuntimeMarkWithCorrelation("sshstdio", correlation, start, bytes, true))
			142	+	end := ports.RuntimeAdapterSendEnd
			143	+	if start == ports.RuntimeAdapterReceiveStart {
			144	+	end = ports.RuntimeAdapterReceiveEnd
			145	+	}
			146	+	return func(valid bool) {
			147	+	defer t.finishObservedOperation()
			148	+	t.observer.ObserveRuntime(ports.NewRuntimeMarkWithCorrelation("sshstdio", correlation, end, bytes, valid))
			149	+	}
			150	+	}
			151	+
103		func (t *transport) mapEOFError(err error) error {	152		func (t *transport) mapEOFError(err error) error {
104		if t.eofErr == nil \|\| (!errors.Is(err, io.EOF) && !errors.Is(err, io.ErrUnexpectedEOF)) {	153		if t.eofErr == nil \|\| (!errors.Is(err, io.EOF) && !errors.Is(err, io.ErrUnexpectedEOF)) {
105		return err	154		return err

@@ -110,7 +159,56 @@ func (t *transport) mapEOFError(err error) error {
110		return err	159		return err
111		}	160		}
112			161
113	-	func (t *transport) Close() error { return t.close() }	162	+	func (t *transport) Close() error {
			163	+	done := t.beginShutdown()
			164	+	// Closing the reader first releases a Recv blocked in io.ReadFull even if
			165	+	// the subprocess shutdown callback reports an error. This lets Recv emit
			166	+	// its failed end mark before the process observer is closed.
			167	+	var readErr error
			168	+	if r, ok := t.r.(io.Closer); ok {
			169	+	readErr = r.Close()
			170	+	}
			171	+	closeErr := t.close()
			172	+	if done != nil {
			173	+	<-done
			174	+	}
			175	+	if closeErr != nil {
			176	+	return closeErr
			177	+	}
			178	+	return readErr
			179	+	}
			180	+
			181	+	func (t *transport) beginObservedOperation() bool {
			182	+	t.operationMu.Lock()
			183	+	defer t.operationMu.Unlock()
			184	+	if t.closing {
			185	+	return false
			186	+	}
			187	+	if t.operationCount == 0 {
			188	+	t.operationsDone = make(chan struct{})
			189	+	}
			190	+	t.operationCount++
			191	+	return true
			192	+	}
			193	+
			194	+	func (t *transport) finishObservedOperation() {
			195	+	t.operationMu.Lock()
			196	+	defer t.operationMu.Unlock()
			197	+	t.operationCount--
			198	+	if t.operationCount == 0 {
			199	+	close(t.operationsDone)
			200	+	}
			201	+	}
			202	+
			203	+	func (t *transport) beginShutdown() <-chan struct{} {
			204	+	t.operationMu.Lock()
			205	+	defer t.operationMu.Unlock()
			206	+	t.closing = true
			207	+	if t.operationCount == 0 {
			208	+	return nil
			209	+	}
			210	+	return t.operationsDone
			211	+	}
114			212
115		type processWaiter struct {	213		type processWaiter struct {
116		cmd *exec.Cmd	214		cmd *exec.Cmd

@@ -222,18 +320,27 @@ func Dial(target, session string) (ports.Transport, error) {
222		return DialContext(context.Background(), target, session)	320		return DialContext(context.Background(), target, session)
223		}	321		}
224			322
			323	+	// DialContextWithRuntimeObserver is the option-bearing production constructor.
			324	+	func DialContextWithRuntimeObserver(ctx context.Context, target, session string, logger *slog.Logger, opts ...Option) (ports.Transport, error) {
			325	+	return dialContext(ctx, target, session, logger, opts...)
			326	+	}
			327	+
225		// DialContext is like Dial, but the context is propagated to ssh startup so a	328		// DialContext is like Dial, but the context is propagated to ssh startup so a
226		// canceled attach attempt interrupts the local ssh process. Callers may pass a	329		// canceled attach attempt interrupts the local ssh process. Callers may pass a
227		// logger to record ssh start failures and non-clean exits without logging the	330		// logger to record ssh start failures and non-clean exits without logging the
228		// generated command line.	331		// generated command line.
229		func DialContext(ctx context.Context, target, session string, logger ...*slog.Logger) (ports.Transport, error) {	332		func DialContext(ctx context.Context, target, session string, logger ...*slog.Logger) (ports.Transport, error) {
230	-	if err := ctx.Err(); err != nil {
231	-	return nil, err
232	-	}
233		var log *slog.Logger	333		var log *slog.Logger
234		if len(logger) > 0 {	334		if len(logger) > 0 {
235		log = logger[0]	335		log = logger[0]
236		}	336		}
			337	+	return dialContext(ctx, target, session, log)
			338	+	}
			339	+
			340	+	func dialContext(ctx context.Context, target, session string, log *slog.Logger, opts ...Option) (ports.Transport, error) {
			341	+	if err := ctx.Err(); err != nil {
			342	+	return nil, err
			343	+	}
237		spec := BuildCommand(target, session)	344		spec := BuildCommand(target, session)
238		cmd := exec.CommandContext(ctx, spec.Path, spec.Args...)	345		cmd := exec.CommandContext(ctx, spec.Path, spec.Args...)
239		stdin, err := cmd.StdinPipe()	346		stdin, err := cmd.StdinPipe()

@@ -254,5 +361,11 @@ func DialContext(ctx context.Context, target, session string, logger ...*slog.Lo
254		}	361		}
255			362
256		waiter := newProcessWaiter(cmd, stdin, &stderr, sshCloseTimeout, log, target, session)	363		waiter := newProcessWaiter(cmd, stdin, &stderr, sshCloseTimeout, log, target, session)
257	-	return newTransport(stdout, stdin, waiter.close, waiter.eofErr), nil	364	+	transport := newTransport(stdout, stdin, waiter.close, waiter.eofErr)
			365	+	for _, opt := range opts {
			366	+	if opt != nil {
			367	+	opt(transport)
			368	+	}
			369	+	}
			370	+	return transport, nil
258		}	371		}

A internal/adapters/sshstdio/transport_observability_test.go +182 -0

@@ -0,0 +1,182 @@
	1	+	package sshstdio
	2	+
	3	+	import (
	4	+	"bytes"
	5	+	"encoding/json"
	6	+	"errors"
	7	+	"os"
	8	+	"strings"
	9	+	"sync"
	10	+	"testing"
	11	+	"time"
	12	+
	13	+	"github.com/bnema/vev/internal/adapters/observability"
	14	+	"github.com/bnema/vev/internal/ports"
	15	+	)
	16	+
	17	+	func TestTransportObservabilitySSHStdioEOFEndsReceive(t *testing.T) {
	18	+	tracePath := t.TempDir() + "/ssh-eof.jsonl"
	19	+	observer, closer, err := observability.NewJSONL(tracePath, sshRuntimeClock{now: time.Unix(0, 309)}, "ssh-process")
	20	+	if err != nil {
	21	+	t.Fatalf("NewJSONL() error = %v", err)
	22	+	}
	23	+	reporter := ports.NewSerializedRuntimeObserver(observer, 64)
	24	+	defer reporter.Close()
	25	+	if _, err := NewTransport(bytes.NewReader(nil), &bytes.Buffer{}, nil, WithRuntimeObserver(reporter)).Recv(); err == nil {
	26	+	t.Fatal("Recv() error = nil at EOF")
	27	+	}
	28	+	reporter.Close()
	29	+	if err := closer.Close(); err != nil {
	30	+	t.Fatalf("trace Close() error = %v", err)
	31	+	}
	32	+	assertSSHReceiveFailurePair(t, tracePath)
	33	+	}
	34	+
	35	+	func TestTransportObservabilitySSHStdioCloseEndsBlockedReceive(t *testing.T) {
	36	+	tracePath := t.TempDir() + "/ssh-close.jsonl"
	37	+	observer, closer, err := observability.NewJSONL(tracePath, sshRuntimeClock{now: time.Unix(0, 311)}, "ssh-process")
	38	+	if err != nil {
	39	+	t.Fatalf("NewJSONL() error = %v", err)
	40	+	}
	41	+	reporter := ports.NewSerializedRuntimeObserver(observer, 64)
	42	+	defer reporter.Close()
	43	+	reader := &sshShutdownReader{entered: make(chan struct{}), done: make(chan struct{})}
	44	+	shutdownErr := errors.New("ssh shutdown failed")
	45	+	transport := NewTransport(reader, &bytes.Buffer{}, func() error { return shutdownErr }, WithRuntimeObserver(reporter))
	46	+	recvDone := make(chan error, 1)
	47	+	go func() { _, err := transport.Recv(); recvDone <- err }()
	48	+	<-reader.entered
	49	+	if err := transport.Close(); !errors.Is(err, shutdownErr) {
	50	+	t.Fatalf("Close() error = %v, want %v", err, shutdownErr)
	51	+	}
	52	+	if err := <-recvDone; err == nil {
	53	+	t.Fatal("Recv() error = nil after Close()")
	54	+	}
	55	+	reporter.Close()
	56	+	if err := closer.Close(); err != nil {
	57	+	t.Fatalf("trace Close() error = %v", err)
	58	+	}
	59	+	assertSSHReceiveFailurePair(t, tracePath)
	60	+	}
	61	+
	62	+	type sshShutdownReader struct {
	63	+	entered chan struct{}
	64	+	done chan struct{}
	65	+	once sync.Once
	66	+	}
	67	+
	68	+	func (r *sshShutdownReader) Read([]byte) (int, error) {
	69	+	r.once.Do(func() { close(r.entered) })
	70	+	<-r.done
	71	+	return 0, errors.New("reader closed")
	72	+	}
	73	+
	74	+	func (r *sshShutdownReader) Close() error {
	75	+	r.once.Do(func() { close(r.entered) })
	76	+	close(r.done)
	77	+	return nil
	78	+	}
	79	+
	80	+	func assertSSHReceiveFailurePair(t *testing.T, path string) {
	81	+	t.Helper()
	82	+	data, err := os.ReadFile(path)
	83	+	if err != nil {
	84	+	t.Fatalf("ReadFile(trace): %v", err)
	85	+	}
	86	+	var starts, ends []struct {
	87	+	ProcessID string `json:"process_id"`
	88	+	Scenario string `json:"scenario"`
	89	+	Run uint64 `json:"run"`
	90	+	Sequence uint64 `json:"sequence"`
	91	+	RequestID uint64 `json:"request_id"`
	92	+	Epoch uint64 `json:"epoch"`
	93	+	Kind string `json:"kind"`
	94	+	Valid bool `json:"valid"`
	95	+	}
	96	+	for _, line := range strings.Split(strings.TrimSpace(string(data)), "\n") {
	97	+	var mark struct {
	98	+	ProcessID string `json:"process_id"`
	99	+	Scenario string `json:"scenario"`
	100	+	Run uint64 `json:"run"`
	101	+	Sequence uint64 `json:"sequence"`
	102	+	RequestID uint64 `json:"request_id"`
	103	+	Epoch uint64 `json:"epoch"`
	104	+	Kind string `json:"kind"`
	105	+	Valid bool `json:"valid"`
	106	+	}
	107	+	if err := json.Unmarshal([]byte(line), &mark); err != nil {
	108	+	t.Fatalf("Unmarshal(trace): %v", err)
	109	+	}
	110	+	switch mark.Kind {
	111	+	case "adapter_receive_start":
	112	+	starts = append(starts, mark)
	113	+	case "adapter_receive_end":
	114	+	ends = append(ends, mark)
	115	+	}
	116	+	}
	117	+	if len(starts) != 1 \|\| len(ends) != 1 {
	118	+	t.Fatalf("receive spans start=%d end=%d, want exactly one pair", len(starts), len(ends))
	119	+	}
	120	+	start, end := starts[0], ends[0]
	121	+	if start.ProcessID != end.ProcessID \|\| start.Scenario != end.Scenario \|\| start.Run != end.Run \|\| start.Sequence != end.Sequence \|\| start.RequestID != end.RequestID \|\| start.Epoch != end.Epoch {
	122	+	t.Fatalf("mismatched receive correlation: start=%+v end=%+v", start, end)
	123	+	}
	124	+	if !start.Valid \|\| end.Valid {
	125	+	t.Fatalf("receive validity start=%t end=%t, want true/false", start.Valid, end.Valid)
	126	+	}
	127	+	}
	128	+
	129	+	func TestTransportObservabilitySSHStdioPreservesCarriage(t *testing.T) {
	130	+	tracePath := t.TempDir() + "/ssh.jsonl"
	131	+	observer, closer, err := observability.NewJSONL(tracePath, sshRuntimeClock{now: time.Unix(0, 307)}, "ssh-process")
	132	+	if err != nil {
	133	+	t.Fatalf("NewJSONL() error = %v", err)
	134	+	}
	135	+	reporter := ports.NewSerializedRuntimeObserver(observer, 64)
	136	+	defer reporter.Close()
	137	+
	138	+	frame := ports.Frame{Type: ports.MsgOutput, Payload: []byte("stdio bytes stay exact")}
	139	+	var observed, baseline bytes.Buffer
	140	+	if err := NewTransport(bytes.NewReader(nil), &observed, nil, WithRuntimeObserver(reporter)).Send(frame); err != nil {
	141	+	t.Fatalf("observed Send() error = %v", err)
	142	+	}
	143	+	if err := NewTransport(bytes.NewReader(nil), &baseline, nil).Send(frame); err != nil {
	144	+	t.Fatalf("baseline Send() error = %v", err)
	145	+	}
	146	+	if !bytes.Equal(observed.Bytes(), baseline.Bytes()) {
	147	+	t.Fatalf("observer changed SSH stdio wire bytes: got %x, want %x", observed.Bytes(), baseline.Bytes())
	148	+	}
	149	+	received, err := NewTransport(bytes.NewReader(baseline.Bytes()), &bytes.Buffer{}, nil, WithRuntimeObserver(reporter)).Recv()
	150	+	if err != nil {
	151	+	t.Fatalf("observed Recv() error = %v", err)
	152	+	}
	153	+	if received.Type != frame.Type \|\| !bytes.Equal(received.Payload, frame.Payload) {
	154	+	t.Fatalf("observed Recv() = %#v, want %#v", received, frame)
	155	+	}
	156	+	reporter.Close()
	157	+	if err := closer.Close(); err != nil {
	158	+	t.Fatalf("trace Close() error = %v", err)
	159	+	}
	160	+	trace, err := os.ReadFile(tracePath)
	161	+	if err != nil {
	162	+	t.Fatalf("ReadFile(trace): %v", err)
	163	+	}
	164	+	for _, kind := range []string{"adapter_send_start", "adapter_send_end", "adapter_receive_start", "adapter_receive_end"} {
	165	+	if !strings.Contains(string(trace), `"kind":"`+kind+`"`) {
	166	+	t.Errorf("trace missing %s", kind)
	167	+	}
	168	+	}
	169	+	}
	170	+
	171	+	type sshRuntimeClock struct{ now time.Time }
	172	+
	173	+	func (c sshRuntimeClock) Now() time.Time { return c.now }
	174	+	func (c sshRuntimeClock) NewTimer(d time.Duration) ports.Timer {
	175	+	return sshRuntimeTimer{timer: time.NewTimer(d)}
	176	+	}
	177	+
	178	+	type sshRuntimeTimer struct{ timer *time.Timer }
	179	+
	180	+	func (t sshRuntimeTimer) C() <-chan time.Time { return t.timer.C }
	181	+	func (t sshRuntimeTimer) Reset(d time.Duration) bool { return t.timer.Reset(d) }
	182	+	func (t sshRuntimeTimer) Stop() bool { return t.timer.Stop() }

A internal/adapters/sshstdio/transport_replay_test.go +42 -0

@@ -0,0 +1,42 @@
	1	+	package sshstdio
	2	+
	3	+	import (
	4	+	"io"
	5	+	"sync"
	6	+	"testing"
	7	+
	8	+	"github.com/bnema/vev/internal/ports"
	9	+	"github.com/bnema/vev/internal/testutil/replaytest"
	10	+	)
	11	+
	12	+	func TestTransportReplay(t *testing.T) {
	13	+	replaytest.Run(t, func(t *testing.T, frames []ports.Frame) []ports.Frame {
	14	+	clientRead, serverWrite := io.Pipe()
	15	+	serverRead, clientWrite := io.Pipe()
	16	+	sender := NewTransport(clientRead, clientWrite, func() error { return clientWrite.Close() })
	17	+	receiver := NewTransport(serverRead, serverWrite, func() error { return serverRead.Close() })
	18	+	t.Cleanup(func() { _ = sender.Close(); _ = receiver.Close() })
	19	+
	20	+	var wg sync.WaitGroup
	21	+	wg.Go(func() {
	22	+	for _, frame := range frames {
	23	+	if err := sender.Send(frame); err != nil {
	24	+	t.Errorf("Send: %v", err)
	25	+	return
	26	+	}
	27	+	}
	28	+	})
	29	+	got := make([]ports.Frame, 0, len(frames))
	30	+	for range frames {
	31	+	frame, err := receiver.Recv()
	32	+	if err != nil {
	33	+	t.Errorf("Recv: %v", err)
	34	+	_ = receiver.Close()
	35	+	break
	36	+	}
	37	+	got = append(got, frame)
	38	+	}
	39	+	wg.Wait()
	40	+	return got
	41	+	})
	42	+	}

M internal/app/daemon_integration_test.go +5 -5


@@ -74,7 +74,7 @@ func recvPump(tr ports.Transport) *pump {
74		// attach dials, handshakes, and returns the transport plus its frame pump.	74		// attach dials, handshakes, and returns the transport plus its frame pump.
75		func attach(t testing.T, dir string, intent uint8, name string, sz domain.Size) (ports.Transport, pump) {	75		func attach(t testing.T, dir string, intent uint8, name string, sz domain.Size) (ports.Transport, pump) {
76		t.Helper()	76		t.Helper()
77	-	tr, err := ipc.Dial(dir)	77	+	tr, err := ipc.DialContext(context.Background(), dir)
78		require.NoError(t, err)	78		require.NoError(t, err)
79		hello := ports.Hello{Version: ports.ProtocolVersion, Intent: intent, Name: name, Size: sz, TermEnv: "xterm-256color"}	79		hello := ports.Hello{Version: ports.ProtocolVersion, Intent: intent, Name: name, Size: sz, TermEnv: "xterm-256color"}
80		require.NoError(t, tr.Send(ports.Frame{Type: ports.MsgHello, Payload: ports.MarshalHello(hello)}))	80		require.NoError(t, tr.Send(ports.Frame{Type: ports.MsgHello, Payload: ports.MarshalHello(hello)}))

@@ -91,7 +91,7 @@ func attach(t *testing.T, dir string, intent uint8, name string, sz domain.Size)
91			91
92		func listRemoteSessions(t *testing.T, dir string) ports.Sessions {	92		func listRemoteSessions(t *testing.T, dir string) ports.Sessions {
93		t.Helper()	93		t.Helper()
94	-	tr, err := ipc.Dial(dir)	94	+	tr, err := ipc.DialContext(context.Background(), dir)
95		require.NoError(t, err)	95		require.NoError(t, err)
96		defer func() { _ = tr.Close() }()	96		defer func() { _ = tr.Close() }()
97		require.NoError(t, tr.Send(ports.Frame{Type: ports.MsgList, Payload: ports.MarshalList(ports.List{})}))	97		require.NoError(t, tr.Send(ports.Frame{Type: ports.MsgList, Payload: ports.MarshalList(ports.List{})}))

@@ -104,7 +104,7 @@ func listRemoteSessions(t *testing.T, dir string) ports.Sessions {
104		}	104		}
105			105
106		func killAll(dir string) error {	106		func killAll(dir string) error {
107	-	tr, err := ipc.Dial(dir)	107	+	tr, err := ipc.DialContext(context.Background(), dir)
108		if err != nil {	108		if err != nil {
109		return err	109		return err
110		}	110		}

@@ -248,7 +248,7 @@ func TestIntegration_CommandPaletteRenamesEphemeralSession(t *testing.T) {
248		text = awaitScreenText(t, p, sz, " work ")	248		text = awaitScreenText(t, p, sz, " work ")
249		require.NotContains(t, text, "work*")	249		require.NotContains(t, text, "work*")
250			250
251	-	listTr, err := ipc.Dial(dir)	251	+	listTr, err := ipc.DialContext(context.Background(), dir)
252		require.NoError(t, err)	252		require.NoError(t, err)
253		defer func() { _ = listTr.Close() }()	253		defer func() { _ = listTr.Close() }()
254		require.NoError(t, listTr.Send(ports.Frame{Type: ports.MsgList, Payload: ports.MarshalList(ports.List{})}))	254		require.NoError(t, listTr.Send(ports.Frame{Type: ports.MsgList, Payload: ports.MarshalList(ports.List{})}))

@@ -367,7 +367,7 @@ func TestIntegration_KillAllShutsDownDaemon(t *testing.T) {
367		tr2, _ := attach(t, dir, ports.IntentNew, "two", sz)	367		tr2, _ := attach(t, dir, ports.IntentNew, "two", sz)
368		require.NoError(t, tr2.Close())	368		require.NoError(t, tr2.Close())
369			369
370	-	killTr, err := ipc.Dial(dir)	370	+	killTr, err := ipc.DialContext(context.Background(), dir)
371		require.NoError(t, err)	371		require.NoError(t, err)
372		defer func() { _ = killTr.Close() }()	372		defer func() { _ = killTr.Close() }()
373		require.NoError(t, killTr.Send(ports.Frame{Type: ports.MsgKill, Payload: ports.MarshalKill(ports.Kill{All: true})}))	373		require.NoError(t, killTr.Send(ports.Frame{Type: ports.MsgKill, Payload: ports.MarshalKill(ports.Kill{All: true})}))

M internal/app/run.go +175 -23


@@ -22,6 +22,7 @@ import (
22		"path/filepath"	22		"path/filepath"
23		"strconv"	23		"strconv"
24		"strings"	24		"strings"
			25	+	"sync"
25		"syscall"	26		"syscall"
26		"text/tabwriter"	27		"text/tabwriter"
27		"time"	28		"time"

@@ -31,6 +32,7 @@ import (
31		"github.com/bnema/vev/internal/adapters/config"	32		"github.com/bnema/vev/internal/adapters/config"
32		"github.com/bnema/vev/internal/adapters/dgram"	33		"github.com/bnema/vev/internal/adapters/dgram"
33		"github.com/bnema/vev/internal/adapters/ipc"	34		"github.com/bnema/vev/internal/adapters/ipc"
			35	+	"github.com/bnema/vev/internal/adapters/observability"
34		"github.com/bnema/vev/internal/adapters/pty"	36		"github.com/bnema/vev/internal/adapters/pty"
35		remoteadapter "github.com/bnema/vev/internal/adapters/remote"	37		remoteadapter "github.com/bnema/vev/internal/adapters/remote"
36		"github.com/bnema/vev/internal/adapters/shellcmd"	38		"github.com/bnema/vev/internal/adapters/shellcmd"

@@ -238,6 +240,106 @@ func dispatch(ctx context.Context, cmd command) error {
238		}	240		}
239		}	241		}
240			242
			243	+	// performanceTrace creates one serialized timestamp owner for this process.
			244	+	// An empty trace environment leaves all production behavior and wire bytes
			245	+	// unchanged.
			246	+	// Composition-root factory seams keep observer propagation testable without
			247	+	// opening real transports.
			248	+	var (
			249	+	newPerformanceTrace = performanceTrace
			250	+	newRemoteDialerFactoryWithRuntimeObserver = func(observer ports.SerializedRuntimeObserver) ports.RemoteDialerFactory {
			251	+	return remoteadapter.NewDialerFactoryWithRuntimeObserver(observer)
			252	+	}
			253	+	runClientWithRuntimeObserver = client.RunWithRuntimeObserver
			254	+	)
			255	+
			256	+	func performanceTrace(clk ports.Clock) (ports.SerializedRuntimeObserver, io.Closer, error) {
			257	+	return performanceTraceWithFactories(clk, observability.NewJSONL, ports.NewRuntimeCorrelationObserver)
			258	+	}
			259	+
			260	+	// performanceTraceWithFactories keeps setup rollback behavior directly
			261	+	// testable without requiring an operating-system file close failure.
			262	+	func performanceTraceWithFactories(
			263	+	clk ports.Clock,
			264	+	newSink func(string, ports.Clock, string) (ports.RuntimeObserver, io.Closer, error),
			265	+	newCorrelation func(ports.RuntimeObserver, ports.RuntimeCorrelationInputs) (ports.RuntimeObserver, error),
			266	+	) (ports.SerializedRuntimeObserver, io.Closer, error) {
			267	+	path, processID := os.Getenv("VEV_PERF_TRACE"), os.Getenv("VEV_PERF_PROCESS_ID")
			268	+	if path == "" {
			269	+	return nil, nil, nil
			270	+	}
			271	+	observer, closer, err := newSink(path, clk, processID)
			272	+	if err != nil {
			273	+	return nil, nil, err
			274	+	}
			275	+	// The harness supplies these manifest fields to every launched role. Keep a
			276	+	// valid standalone trace for operators that set only the original trace
			277	+	// variables, while ensuring harness traces match their process mapping.
			278	+	inputs := ports.RuntimeCorrelationInputs{Scenario: os.Getenv("VEV_PERF_SCENARIO"), Run: 1}
			279	+	if inputs.Scenario == "" {
			280	+	inputs.Scenario = "runtime"
			281	+	}
			282	+	if rawRun := os.Getenv("VEV_PERF_RUN"); rawRun != "" {
			283	+	inputs.Run, err = strconv.ParseUint(rawRun, 10, 64)
			284	+	if err != nil {
			285	+	setupErr := fmt.Errorf("invalid VEV_PERF_RUN %q: %w", rawRun, err)
			286	+	return nil, nil, closeTraceAfterSetupFailure(closer, setupErr)
			287	+	}
			288	+	if inputs.Run == 0 {
			289	+	setupErr := fmt.Errorf("invalid VEV_PERF_RUN %q", rawRun)
			290	+	return nil, nil, closeTraceAfterSetupFailure(closer, setupErr)
			291	+	}
			292	+	}
			293	+	observer, err = newCorrelation(observer, inputs)
			294	+	if err != nil {
			295	+	setupErr := fmt.Errorf("configure runtime trace correlation: %w", err)
			296	+	return nil, nil, closeTraceAfterSetupFailure(closer, setupErr)
			297	+	}
			298	+	reporter := ports.NewSerializedRuntimeObserver(observer, runtimeTraceQueueDepth)
			299	+	return reporter, &runtimeTraceCloser{reporter: reporter, closer: closer}, nil
			300	+	}
			301	+
			302	+	func closeTraceAfterSetupFailure(closer io.Closer, setupErr error) error {
			303	+	if closer == nil {
			304	+	return setupErr
			305	+	}
			306	+	if closeErr := closer.Close(); closeErr != nil {
			307	+	return errors.Join(setupErr, fmt.Errorf("close performance trace after setup failure: %w", closeErr))
			308	+	}
			309	+	return setupErr
			310	+	}
			311	+
			312	+	// runtimeTraceQueueDepth bounds all process-local trace producer handoffs.
			313	+	// A full queue emits the serialized diagnostic gap rather than delaying a
			314	+	// terminal, transport, or ACK progress path.
			315	+	const runtimeTraceQueueDepth = 256
			316	+
			317	+	// runtimeTraceCloser owns the two-stage trace shutdown: drain the reporter
			318	+	// before closing the concrete timestamp/file owner. sync.Once makes every
			319	+	// deferred and explicit close path safe without closing shared workers twice.
			320	+	type runtimeTraceCloser struct {
			321	+	reporter ports.SerializedRuntimeObserver
			322	+	closer io.Closer
			323	+	once sync.Once
			324	+	err error
			325	+	}
			326	+
			327	+	func (c *runtimeTraceCloser) Close() error {
			328	+	if c == nil {
			329	+	return nil
			330	+	}
			331	+	c.once.Do(func() {
			332	+	if c.reporter != nil {
			333	+	c.reporter.Flush()
			334	+	c.reporter.Close()
			335	+	}
			336	+	if c.closer != nil {
			337	+	c.err = c.closer.Close()
			338	+	}
			339	+	})
			340	+	return c.err
			341	+	}
			342	+
241		func configureLogging(component logging.Component, rotateAtRuntime bool) (*slog.Logger, io.Closer, error) {	343		func configureLogging(component logging.Component, rotateAtRuntime bool) (*slog.Logger, io.Closer, error) {
242		return logging.Setup(logging.Config{	344		return logging.Setup(logging.Config{
243		Dir: platform.StateDir(),	345		Dir: platform.StateDir(),

@@ -278,14 +380,23 @@ func pprofAddrIsLoopback(addr string) bool {
278		// entered by an auto-spawned child): it sets up logging, binds the socket,	380		// entered by an auto-spawned child): it sets up logging, binds the socket,
279		// constructs the daemon, and serves until the last session exits or a	381		// constructs the daemon, and serves until the last session exits or a
280		// termination signal arrives (graceful shutdown notifies attached clients).	382		// termination signal arrives (graceful shutdown notifies attached clients).
281	-	func runDaemon() error {	383	+	func runDaemon() (retErr error) {
282		log, logCloser, err := configureLogging(logging.Daemon, true)	384		log, logCloser, err := configureLogging(logging.Daemon, true)
283		if err != nil {	385		if err != nil {
284		return err	386		return err
285		}	387		}
286		defer func() { _ = logCloser.Close() }()	388		defer func() { _ = logCloser.Close() }()
287			389
288	-	ln, err := ipc.Listen(ipc.SocketDir())	390	+	clk := clock.New()
			391	+	observer, observerCloser, err := newPerformanceTrace(clk)
			392	+	if err != nil {
			393	+	return fmt.Errorf("vev: performance trace: %w", err)
			394	+	}
			395	+	if observerCloser != nil {
			396	+	defer func() { retErr = errors.Join(retErr, observerCloser.Close()) }()
			397	+	}
			398	+	// Every accepted local connection shares this process's serialized sink.
			399	+	ln, err := ipc.Listen(ipc.SocketDir(), ipc.WithRuntimeObserver(observer))
289		if err != nil {	400		if err != nil {
290		log.Error("daemon listen failed", "socket_dir", ipc.SocketDir(), "err", err)	401		log.Error("daemon listen failed", "socket_dir", ipc.SocketDir(), "err", err)
291		return fmt.Errorf("vev: daemon listen: %w", err)	402		return fmt.Errorf("vev: daemon listen: %w", err)

@@ -309,6 +420,9 @@ func runDaemon() error {
309			420
310		log.Info("daemon starting", "socket", ln.Addr())	421		log.Info("daemon starting", "socket", ln.Addr())
311		daemonOpts := []daemon.Option(nil)	422		daemonOpts := []daemon.Option(nil)
			423	+	if observer != nil {
			424	+	daemonOpts = append(daemonOpts, daemon.WithRuntimeObserver(observer))
			425	+	}
312		configPath := platform.ConfigPath()	426		configPath := platform.ConfigPath()
313		cfg, warnings, err := config.Load(configPath)	427		cfg, warnings, err := config.Load(configPath)
314		if err != nil {	428		if err != nil {

@@ -327,7 +441,6 @@ func runDaemon() error {
327		log.Info("session persistence enabled", "path", storePath)	441		log.Info("session persistence enabled", "path", storePath)
328		daemonOpts = append(daemonOpts, daemon.WithStore(store))	442		daemonOpts = append(daemonOpts, daemon.WithStore(store))
329		}	443		}
330	-	clk := clock.New()
331		d := daemon.New(pty.NewFactory(), clk, log, daemonOpts...)	444		d := daemon.New(pty.NewFactory(), clk, log, daemonOpts...)
332		watchCtx, stopWatch := context.WithCancel(ctx)	445		watchCtx, stopWatch := context.WithCancel(ctx)
333		defer stopWatch()	446		defer stopWatch()

@@ -350,20 +463,32 @@ func runDaemon() error {
350		// runAttach dials (auto-spawning the daemon if needed) and runs the client	463		// runAttach dials (auto-spawning the daemon if needed) and runs the client
351		// attach loop. Logging goes to the shared file: the client must never write	464		// attach loop. Logging goes to the shared file: the client must never write
352		// to the console while the terminal is raw.	465		// to the console while the terminal is raw.
353	-	func runAttach(ctx context.Context, intent uint8, name, remoteTarget string) error {	466	+	func runAttach(ctx context.Context, intent uint8, name, remoteTarget string) (retErr error) {
354		log, logCloser, err := configureLogging(logging.Client, false)	467		log, logCloser, err := configureLogging(logging.Client, false)
355		if err != nil {	468		if err != nil {
356		return err	469		return err
357		}	470		}
358		defer func() { _ = logCloser.Close() }()	471		defer func() { _ = logCloser.Close() }()
359			472
			473	+	clk := clock.New()
			474	+	observer, observerCloser, err := newPerformanceTrace(clk)
			475	+	if err != nil {
			476	+	return fmt.Errorf("vev: performance trace: %w", err)
			477	+	}
			478	+	if observerCloser != nil {
			479	+	defer func() { retErr = errors.Join(retErr, observerCloser.Close()) }()
			480	+	}
360		return runAttachWithDeps(ctx, intent, name, remoteTarget, os.Getenv("VEV"), log, runAttachDeps{	481		return runAttachWithDeps(ctx, intent, name, remoteTarget, os.Getenv("VEV"), log, runAttachDeps{
361	-	localDialer: defaultLocalDialer,	482	+	localDialer: func() ports.Dialer {
362	-	remoteDialerFactory: defaultRemoteDialerFactory(),	483	+	return localDaemonDialer{dir: ipc.SocketDir(), observer: observer}
			484	+	},
			485	+	remoteDialerFactory: newRemoteDialerFactoryWithRuntimeObserver(observer),
363		selectedRemoteTransport: os.Getenv(envRemoteTransport),	486		selectedRemoteTransport: os.Getenv(envRemoteTransport),
364	-	runClient: client.Run,	487	+	runClient: func(ctx context.Context, dialer ports.Dialer, terminal ports.Terminal, behaviorClock ports.Clock, intent uint8, name string, remote bool, clipboard ports.ClipboardReader, log *slog.Logger) error {
365	-	createDetached: createDetachedLocalSession,	488	+	return runClientWithRuntimeObserver(ctx, dialer, terminal, behaviorClock, intent, name, remote, clipboard, log, observer)
366	-	clipboard: clipboard.New(),	489	+	},
			490	+	createDetached: createDetachedLocalSession,
			491	+	clipboard: clipboard.New(),
367		})	492		})
368		}	493		}
369			494

@@ -499,10 +624,19 @@ func isDaemonVersionDrift(err error) bool {
499		return protocolErr.Code == ports.ErrVersionMismatch \|\| protocolErr.Text == legacyMalformedHelloSignal	624		return protocolErr.Code == ports.ErrVersionMismatch \|\| protocolErr.Text == legacyMalformedHelloSignal
500		}	625		}
501			626
502	-	type localDaemonDialer struct{ dir string }	627	+	type localDaemonDialer struct {
			628	+	dir string
			629	+	observer ports.SerializedRuntimeObserver
			630	+	}
503			631
504		func (d localDaemonDialer) Dial(ctx context.Context) (ports.Transport, error) {	632		func (d localDaemonDialer) Dial(ctx context.Context) (ports.Transport, error) {
505	-	return ensureDaemon(ctx, d.dir, realDial, realSpawn, defaultBackoff)	633	+	dial := realDial
			634	+	if d.observer != nil {
			635	+	dial = func(ctx context.Context, dir string) (ports.Transport, error) {
			636	+	return ipc.DialContext(ctx, dir, ipc.WithRuntimeObserver(d.observer))
			637	+	}
			638	+	}
			639	+	return ensureDaemon(ctx, d.dir, dial, realSpawn, defaultBackoff)
506		}	640		}
507			641
508		func detachedLocalHello(name, cwd string) ports.Hello {	642		func detachedLocalHello(name, cwd string) ports.Hello {

@@ -584,7 +718,7 @@ func requestDaemonStop(ctx context.Context) error {
584			718
585		done := make(chan error, 1)	719		done := make(chan error, 1)
586		go func() {	720		go func() {
587	-	transport, err := realDial(ipc.SocketDir())	721	+	transport, err := realDial(ctx, ipc.SocketDir())
588		if err != nil {	722		if err != nil {
589		done <- fmt.Errorf("vev: no daemon running")	723		done <- fmt.Errorf("vev: no daemon running")
590		return	724		return

@@ -621,7 +755,7 @@ func requestDaemonStop(ctx context.Context) error {
621		// runStdio is the hidden remote-side mode used by `ssh host vev _stdio`: it	755		// runStdio is the hidden remote-side mode used by `ssh host vev _stdio`: it
622		// connects to the per-user daemon (auto-spawning it if needed) and proxies the	756		// connects to the per-user daemon (auto-spawning it if needed) and proxies the
623		// framed protocol between process stdio and the daemon socket.	757		// framed protocol between process stdio and the daemon socket.
624	-	func runStdio(ctx context.Context) error {	758	+	func runStdio(ctx context.Context) (retErr error) {
625		log, logCloser, err := configureLogging(logging.Stdio, false)	759		log, logCloser, err := configureLogging(logging.Stdio, false)
626		if err != nil {	760		if err != nil {
627		return err	761		return err

@@ -629,13 +763,22 @@ func runStdio(ctx context.Context) error {
629		defer func() { _ = logCloser.Close() }()	763		defer func() { _ = logCloser.Close() }()
630		log.Debug("stdio proxy starting")	764		log.Debug("stdio proxy starting")
631			765
632	-	transport, err := ensureDaemon(ctx, ipc.SocketDir(), realDial, realSpawn, defaultBackoff)	766	+	clk := clock.New()
			767	+	observer, observerCloser, err := newPerformanceTrace(clk)
			768	+	if err != nil {
			769	+	return fmt.Errorf("vev: performance trace: %w", err)
			770	+	}
			771	+	if observerCloser != nil {
			772	+	defer func() { retErr = errors.Join(retErr, observerCloser.Close()) }()
			773	+	}
			774	+	transport, err := ensureDaemon(ctx, ipc.SocketDir(), func(ctx context.Context, dir string) (ports.Transport, error) {
			775	+	return ipc.DialContext(ctx, dir, ipc.WithRuntimeObserver(observer))
			776	+	}, realSpawn, defaultBackoff)
633		if err != nil {	777		if err != nil {
634		return err	778		return err
635		}	779		}
636		defer func() { _ = transport.Close() }()	780		defer func() { _ = transport.Close() }()
637	-		781	+	stdio := sshstdio.NewTransport(os.Stdin, os.Stdout, nil, sshstdio.WithRuntimeObserver(observer))
638	-	stdio := sshstdio.NewTransport(os.Stdin, os.Stdout, nil)
639		return proxyTransports(ctx, stdio, transport, log)	782		return proxyTransports(ctx, stdio, transport, log)
640		}	783		}
641			784

@@ -719,7 +862,7 @@ func runUDPBootstrap(ctx context.Context, session string) error {
719		}	862		}
720			863
721		// runUDPProxy is the detached long-lived remote-side datagram proxy.	864		// runUDPProxy is the detached long-lived remote-side datagram proxy.
722	-	func runUDPProxy(ctx context.Context, session string, ready io.Writer) error {	865	+	func runUDPProxy(ctx context.Context, session string, ready io.Writer) (retErr error) {
723		log, logCloser, err := configureLogging(logging.Stdio, false)	866		log, logCloser, err := configureLogging(logging.Stdio, false)
724		if err != nil {	867		if err != nil {
725		return err	868		return err

@@ -727,6 +870,13 @@ func runUDPProxy(ctx context.Context, session string, ready io.Writer) error {
727		defer func() { _ = logCloser.Close() }()	870		defer func() { _ = logCloser.Close() }()
728		log.Debug("udp proxy starting", "session", session)	871		log.Debug("udp proxy starting", "session", session)
729			872
			873	+	observer, observerCloser, err := newPerformanceTrace(clock.New())
			874	+	if err != nil {
			875	+	return fmt.Errorf("vev: performance trace: %w", err)
			876	+	}
			877	+	if observerCloser != nil {
			878	+	defer func() { retErr = errors.Join(retErr, observerCloser.Close()) }()
			879	+	}
730		portRange, err := parseUDPPortRange(os.Getenv(envUDPPortRange))	880		portRange, err := parseUDPPortRange(os.Getenv(envUDPPortRange))
731		if err != nil {	881		if err != nil {
732		return err	882		return err

@@ -740,14 +890,16 @@ func runUDPProxy(ctx context.Context, session string, ready io.Writer) error {
740		if _, err := rand.Read(key); err != nil {	890		if _, err := rand.Read(key); err != nil {
741		return err	891		return err
742		}	892		}
743	-	daemonTr, err := ensureDaemon(ctx, ipc.SocketDir(), realDial, realSpawn, defaultBackoff)	893	+	daemonTr, err := ensureDaemon(ctx, ipc.SocketDir(), func(ctx context.Context, dir string) (ports.Transport, error) {
			894	+	return ipc.DialContext(ctx, dir, ipc.WithRuntimeObserver(observer))
			895	+	}, realSpawn, defaultBackoff)
744		if err != nil {	896		if err != nil {
745		return err	897		return err
746		}	898		}
747		defer func() { _ = daemonTr.Close() }()	899		defer func() { _ = daemonTr.Close() }()
748		udpOptions := udpProxyClientTransportOptions	900		udpOptions := udpProxyClientTransportOptions
749		udpOptions.Observe = dgram.DiagnosticLogObserver(log)	901		udpOptions.Observe = dgram.DiagnosticLogObserver(log)
750	-	dg, err := dgram.NewTransportWithOptions(conn, nil, key, 2, 1, udpOptions)	902	+	dg, err := dgram.NewTransportWithOptions(conn, nil, key, 2, 1, udpOptions, dgram.WithRuntimeObserver(observer))
751		if err != nil {	903		if err != nil {
752		return err	904		return err
753		}	905		}

@@ -890,8 +1042,8 @@ func proxyTransports(ctx context.Context, a, b ports.Transport, log *slog.Logger
890			1042
891		// runList prints the daemon's session listing. With no daemon running, it	1043		// runList prints the daemon's session listing. With no daemon running, it
892		// falls back to the persisted stopped-session records.	1044		// falls back to the persisted stopped-session records.
893	-	func runList(_ context.Context) error {	1045	+	func runList(ctx context.Context) error {
894	-	transport, err := realDial(ipc.SocketDir())	1046	+	transport, err := realDial(ctx, ipc.SocketDir())
895		if err != nil {	1047		if err != nil {
896		records, loadErr := persist.LoadReadOnly(platform.StateDir())	1048		records, loadErr := persist.LoadReadOnly(platform.StateDir())
897		if loadErr != nil {	1049		if loadErr != nil {

@@ -955,8 +1107,8 @@ func printSessions(w io.Writer, sessions []ports.SessionInfo) {
955		}	1107		}
956			1108
957		// runKill asks the daemon to terminate a named session, every session, or the daemon.	1109		// runKill asks the daemon to terminate a named session, every session, or the daemon.
958	-	func runKill(_ context.Context, name string, all, daemon bool) error {	1110	+	func runKill(ctx context.Context, name string, all, daemon bool) error {
959	-	transport, err := realDial(ipc.SocketDir())	1111	+	transport, err := realDial(ctx, ipc.SocketDir())
960		if err != nil {	1112		if err != nil {
961		if name != "" && !all && !daemon {	1113		if name != "" && !all && !daemon {
962		storePath := persist.StorePath(platform.StateDir())	1114		storePath := persist.StorePath(platform.StateDir())

M internal/app/spawn.go +14 -5

 // dialFunc dials the daemon socket in dir. Injected so tests can drive the
 // spawn/backoff logic without a real socket.
-type dialFunc func(dir string) (ports.Transport, error)
+type dialFunc func(ctx context.Context, dir string) (ports.Transport, error)
 // spawnFunc launches a detached daemon process. Injected so tests never
 // re-exec a real binary.
 // spawner via an mkdir lock (losers simply wait), then retry-dials with
 // backoff until the socket is live or the budget expires.
 func ensureDaemon(ctx context.Context, dir string, dial dialFunc, spawn spawnFunc, cfg backoffConfig) (ports.Transport, error) {
-	if t, err := dial(dir); err == nil {
+	if err := ctx.Err(); err != nil {
+		return nil, err
+	}
+	if t, err := dial(ctx, dir); err == nil {
 		slog.Debug("daemon already reachable", "socket_dir", dir)
 		return t, nil
 	}
 		// rather than spawning a second daemon.
 		defer release()
 		slog.Info("spawning daemon", "socket_dir", dir)
+		if err := ctx.Err(); err != nil {
+			return nil, err
+		}
 		if err := spawn(); err != nil {
 			slog.Error("daemon spawn failed", "err", err)
 			return nil, fmt.Errorf("vev: spawning daemon: %w", err)
 	deadline := time.Now().Add(cfg.total)
 	backoff := cfg.initial
 	for {
-		if t, err := dial(dir); err == nil {
+		if err := ctx.Err(); err != nil {
+			return nil, err
+		}
+		if t, err := dial(ctx, dir); err == nil {
 			return t, nil
 		}
 		if !time.Now().Before(deadline) {
 }
 // realDial is the production dialer.
-func realDial(dir string) (ports.Transport, error) {
+func realDial(ctx context.Context, dir string) (ports.Transport, error) {
-	return ipc.Dial(dir)
+	return ipc.DialContext(ctx, dir)
 }
 // realSpawn re-execs this binary as a detached daemon: a new session

M internal/app/spawn_test.go +53 -6


@@ -91,7 +91,7 @@ func TestAcquireSpawnLockTakesOverStaleLock(t *testing.T) {
91		func TestRetryDialSucceedsAfterTransientFailures(t *testing.T) {	91		func TestRetryDialSucceedsAfterTransientFailures(t *testing.T) {
92		want := portsmocks.NewMockTransport(t)	92		want := portsmocks.NewMockTransport(t)
93		var calls atomic.Int32	93		var calls atomic.Int32
94	-	dial := func(string) (ports.Transport, error) {	94	+	dial := func(context.Context, string) (ports.Transport, error) {
95		if calls.Add(1) <= 3 {	95		if calls.Add(1) <= 3 {
96		return nil, errDialFailed	96		return nil, errDialFailed
97		}	97		}

@@ -111,7 +111,7 @@ func TestRetryDialSucceedsAfterTransientFailures(t *testing.T) {
111		}	111		}
112			112
113		func TestRetryDialPermanentFailure(t *testing.T) {	113		func TestRetryDialPermanentFailure(t *testing.T) {
114	-	dial := func(string) (ports.Transport, error) { return nil, errDialFailed }	114	+	dial := func(context.Context, string) (ports.Transport, error) { return nil, errDialFailed }
115		_, err := retryDial(context.Background(), t.TempDir(), dial, fastBackoff)	115		_, err := retryDial(context.Background(), t.TempDir(), dial, fastBackoff)
116		if !errors.Is(err, ErrDaemonUnreachable) {	116		if !errors.Is(err, ErrDaemonUnreachable) {
117		t.Fatalf("retryDial error = %v, want ErrDaemonUnreachable", err)	117		t.Fatalf("retryDial error = %v, want ErrDaemonUnreachable", err)

@@ -121,18 +121,65 @@ func TestRetryDialPermanentFailure(t *testing.T) {
121		func TestRetryDialContextCancelled(t *testing.T) {	121		func TestRetryDialContextCancelled(t *testing.T) {
122		ctx, cancel := context.WithCancel(context.Background())	122		ctx, cancel := context.WithCancel(context.Background())
123		cancel()	123		cancel()
124	-	dial := func(string) (ports.Transport, error) { return nil, errDialFailed }	124	+	var calls atomic.Int32
			125	+	dial := func(context.Context, string) (ports.Transport, error) {
			126	+	calls.Add(1)
			127	+	return nil, errDialFailed
			128	+	}
125		_, err := retryDial(ctx, t.TempDir(), dial, fastBackoff)	129		_, err := retryDial(ctx, t.TempDir(), dial, fastBackoff)
126		if !errors.Is(err, context.Canceled) {	130		if !errors.Is(err, context.Canceled) {
127		t.Fatalf("retryDial error = %v, want context.Canceled", err)	131		t.Fatalf("retryDial error = %v, want context.Canceled", err)
128		}	132		}
			133	+	if calls.Load() != 0 {
			134	+	t.Fatalf("dial called %d times for canceled context, want 0", calls.Load())
			135	+	}
			136	+	}
			137	+
			138	+	func TestEnsureDaemonContextCancelledDoesNotDialOrSpawn(t *testing.T) {
			139	+	ctx, cancel := context.WithCancel(context.Background())
			140	+	cancel()
			141	+	dial := func(context.Context, string) (ports.Transport, error) {
			142	+	t.Fatal("ensureDaemon dialed with a canceled context")
			143	+	return nil, nil
			144	+	}
			145	+	spawn := func() error {
			146	+	t.Fatal("ensureDaemon spawned with a canceled context")
			147	+	return nil
			148	+	}
			149	+
			150	+	_, err := ensureDaemon(ctx, t.TempDir(), dial, spawn, fastBackoff)
			151	+	if !errors.Is(err, context.Canceled) {
			152	+	t.Fatalf("ensureDaemon error = %v, want context.Canceled", err)
			153	+	}
			154	+	}
			155	+
			156	+	func TestEnsureDaemonDoesNotSpawnWhenContextCancelsAfterFailedDial(t *testing.T) {
			157	+	ctx, cancel := context.WithCancel(context.Background())
			158	+	defer cancel()
			159	+	var spawnCalls atomic.Int32
			160	+	dial := func(context.Context, string) (ports.Transport, error) {
			161	+	cancel()
			162	+	return nil, errDialFailed
			163	+	}
			164	+	spawn := func() error {
			165	+	spawnCalls.Add(1)
			166	+	return nil
			167	+	}
			168	+
			169	+	_, err := ensureDaemon(ctx, filepath.Join(t.TempDir(), "vev"), dial, spawn, fastBackoff)
			170	+	if !errors.Is(err, context.Canceled) {
			171	+	t.Fatalf("ensureDaemon error = %v, want context.Canceled", err)
			172	+	}
			173	+	if got := spawnCalls.Load(); got != 0 {
			174	+	t.Fatalf("spawn called %d times after canceled failed dial, want 0", got)
			175	+	}
129		}	176		}
130			177
131		func TestEnsureDaemonSpawnsThenDials(t *testing.T) {	178		func TestEnsureDaemonSpawnsThenDials(t *testing.T) {
132		dir := filepath.Join(t.TempDir(), "vev")	179		dir := filepath.Join(t.TempDir(), "vev")
133		want := portsmocks.NewMockTransport(t)	180		want := portsmocks.NewMockTransport(t)
134		var dialCalls, spawnCalls atomic.Int32	181		var dialCalls, spawnCalls atomic.Int32
135	-	dial := func(string) (ports.Transport, error) {	182	+	dial := func(context.Context, string) (ports.Transport, error) {
136		// First (pre-spawn) dial fails; the post-spawn retry succeeds.	183		// First (pre-spawn) dial fails; the post-spawn retry succeeds.
137		if dialCalls.Add(1) == 1 {	184		if dialCalls.Add(1) == 1 {
138		return nil, errDialFailed	185		return nil, errDialFailed

@@ -163,7 +210,7 @@ func TestEnsureDaemonSpawnsThenDials(t *testing.T) {
163		func TestEnsureDaemonReturnsExistingDaemon(t *testing.T) {	210		func TestEnsureDaemonReturnsExistingDaemon(t *testing.T) {
164		want := portsmocks.NewMockTransport(t)	211		want := portsmocks.NewMockTransport(t)
165		var spawnCalls atomic.Int32	212		var spawnCalls atomic.Int32
166	-	dial := func(string) (ports.Transport, error) { return want, nil }	213	+	dial := func(context.Context, string) (ports.Transport, error) { return want, nil }
167		spawn := func() error { spawnCalls.Add(1); return nil }	214		spawn := func() error { spawnCalls.Add(1); return nil }
168			215
169		got, err := ensureDaemon(context.Background(), t.TempDir(), dial, spawn, fastBackoff)	216		got, err := ensureDaemon(context.Background(), t.TempDir(), dial, spawn, fastBackoff)

@@ -180,7 +227,7 @@ func TestEnsureDaemonReturnsExistingDaemon(t *testing.T) {
180			227
181		func TestEnsureDaemonSpawnFailure(t *testing.T) {	228		func TestEnsureDaemonSpawnFailure(t *testing.T) {
182		spawnErr := errors.New("exec boom")	229		spawnErr := errors.New("exec boom")
183	-	dial := func(string) (ports.Transport, error) { return nil, errDialFailed }	230	+	dial := func(context.Context, string) (ports.Transport, error) { return nil, errDialFailed }
184		spawn := func() error { return spawnErr }	231		spawn := func() error { return spawnErr }
185			232
186		_, err := ensureDaemon(context.Background(), filepath.Join(t.TempDir(), "vev"), dial, spawn, fastBackoff)	233		_, err := ensureDaemon(context.Background(), filepath.Join(t.TempDir(), "vev"), dial, spawn, fastBackoff)

A internal/app/transport_observability_test.go +250 -0

@@ -0,0 +1,250 @@
	1	+	package app
	2	+
	3	+	import (
	4	+	"bufio"
	5	+	"context"
	6	+	"encoding/json"
	7	+	"errors"
	8	+	"io"
	9	+	"log/slog"
	10	+	"os"
	11	+	"path/filepath"
	12	+	"strconv"
	13	+	"strings"
	14	+	"testing"
	15	+	"time"
	16	+
	17	+	"github.com/bnema/vev/internal/ports"
	18	+	portsmocks "github.com/bnema/vev/internal/ports/mocks"
	19	+	"github.com/stretchr/testify/mock"
	20	+	)
	21	+
	22	+	func TestPerformanceTraceUsesHarnessProcessMapping(t *testing.T) {
	23	+	path := filepath.Join(t.TempDir(), "process.jsonl")
	24	+	t.Setenv("VEV_PERF_TRACE", path)
	25	+	t.Setenv("VEV_PERF_PROCESS_ID", "idle-local-r001-daemon-01")
	26	+	t.Setenv("VEV_PERF_SCENARIO", "1x4-idle-local")
	27	+	t.Setenv("VEV_PERF_RUN", "1")
	28	+
	29	+	clock := portsmocks.NewMockClock(t)
	30	+	clock.EXPECT().Now().Return(time.Unix(0, 1)).Once()
	31	+	observer, closer, err := performanceTrace(clock)
	32	+	if err != nil {
	33	+	t.Fatal(err)
	34	+	}
	35	+	if observer == nil \|\| closer == nil {
	36	+	t.Fatal("performance trace was not configured")
	37	+	}
	38	+	// performanceTrace's serialized return type makes a raw JSONL sink
	39	+	// unrepresentable to application transport wiring.
	40	+	observer.ObserveRuntime(ports.RuntimeMark{Schema: ports.RuntimeMarkSchema, Component: "daemon", Scenario: "runtime", Run: 1, Kind: ports.RuntimeEmitEnd, Valid: true})
	41	+	if err := closer.Close(); err != nil {
	42	+	t.Fatal(err)
	43	+	}
	44	+	if err := closer.Close(); err != nil {
	45	+	t.Fatal(err)
	46	+	}
	47	+	f, err := os.Open(path)
	48	+	if err != nil {
	49	+	t.Fatal(err)
	50	+	}
	51	+	defer func() {
	52	+	if err := f.Close(); err != nil {
	53	+	t.Errorf("close trace file: %v", err)
	54	+	}
	55	+	}()
	56	+	var mark struct {
	57	+	ProcessID string `json:"process_id"`
	58	+	Scenario string `json:"scenario"`
	59	+	Run uint64 `json:"run"`
	60	+	Sequence uint64 `json:"sequence"`
	61	+	RequestID uint64 `json:"request_id"`
	62	+	Epoch uint64 `json:"epoch"`
	63	+	}
	64	+	if err := json.NewDecoder(bufio.NewReader(f)).Decode(&mark); err != nil {
	65	+	t.Fatal(err)
	66	+	}
	67	+	if mark.ProcessID != "idle-local-r001-daemon-01" \|\| mark.Scenario != "1x4-idle-local" \|\| mark.Run != 1 \|\| mark.Sequence == 0 \|\| mark.RequestID == 0 \|\| mark.Epoch == 0 {
	68	+	t.Fatalf("mark does not match harness mapping: %+v", mark)
	69	+	}
	70	+	}
	71	+
	72	+	func TestPerformanceTraceWithFactoriesCreatesOneConfiguredObserver(t *testing.T) {
	73	+	t.Setenv("VEV_PERF_TRACE", "trace.jsonl")
	74	+	t.Setenv("VEV_PERF_PROCESS_ID", "client-01")
	75	+	t.Setenv("VEV_PERF_SCENARIO", "lossy-remote")
	76	+	t.Setenv("VEV_PERF_RUN", "7")
	77	+
	78	+	clock := portsmocks.NewMockClock(t)
	79	+	sink := portsmocks.NewMockRuntimeObserver(t)
	80	+	sink.EXPECT().ObserveRuntime(mock.MatchedBy(func(mark ports.RuntimeMark) bool {
	81	+	return mark.Component == "client" && mark.Scenario == "lossy-remote" && mark.Run == 7 &&
	82	+	mark.Sequence != 0 && mark.RequestID != 0 && mark.Epoch != 0 && mark.Kind == ports.RuntimeEmitEnd
	83	+	})).Once()
	84	+
	85	+	var sinkCalls, correlationCalls int
	86	+	observer, closer, err := performanceTraceWithFactories(clock,
	87	+	func(path string, gotClock ports.Clock, processID string) (ports.RuntimeObserver, io.Closer, error) {
	88	+	sinkCalls++
	89	+	if path != "trace.jsonl" \|\| gotClock != clock \|\| processID != "client-01" {
	90	+	t.Fatalf("trace sink config = (%q, %v, %q)", path, gotClock, processID)
	91	+	}
	92	+	return sink, errorCloser{}, nil
	93	+	},
	94	+	func(got ports.RuntimeObserver, inputs ports.RuntimeCorrelationInputs) (ports.RuntimeObserver, error) {
	95	+	correlationCalls++
	96	+	if got != sink \|\| inputs != (ports.RuntimeCorrelationInputs{Scenario: "lossy-remote", Run: 7}) {
	97	+	t.Fatalf("trace correlation config = (%v, %+v)", got, inputs)
	98	+	}
	99	+	return ports.NewRuntimeCorrelationObserver(got, inputs)
	100	+	},
	101	+	)
	102	+	if err != nil {
	103	+	t.Fatal(err)
	104	+	}
	105	+	if sinkCalls != 1 \|\| correlationCalls != 1 \|\| observer == nil \|\| closer == nil {
	106	+	t.Fatalf("trace setup = sink calls %d, correlation calls %d, observer %v, closer %v; want one configured observer", sinkCalls, correlationCalls, observer, closer)
	107	+	}
	108	+	observer.ObserveRuntime(ports.NewRuntimeMark("client", ports.RuntimeEmitEnd, 0, true))
	109	+	if err := closer.Close(); err != nil {
	110	+	t.Fatal(err)
	111	+	}
	112	+	}
	113	+
	114	+	func TestPerformanceTraceSetupRollbackJoinsCloseError(t *testing.T) {
	115	+	setupErr := errors.New("correlation setup failed")
	116	+	closeErr := errors.New("rollback close failed")
	117	+
	118	+	tests := []struct {
	119	+	name string
	120	+	rawRun string
	121	+	correlationErr error
	122	+	wantPrimaryText string
	123	+	wantPrimaryErr error
	124	+	wantNumError bool
	125	+	}{
	126	+	{
	127	+	name: "validation",
	128	+	rawRun: "not-a-run",
	129	+	wantPrimaryText: `invalid VEV_PERF_RUN "not-a-run"`,
	130	+	wantNumError: true,
	131	+	},
	132	+	{
	133	+	name: "correlation",
	134	+	rawRun: "1",
	135	+	correlationErr: setupErr,
	136	+	wantPrimaryText: "runtime trace correlation",
	137	+	wantPrimaryErr: setupErr,
	138	+	},
	139	+	}
	140	+
	141	+	for _, tc := range tests {
	142	+	t.Run(tc.name, func(t *testing.T) {
	143	+	t.Setenv("VEV_PERF_TRACE", "unused-by-test-sink")
	144	+	t.Setenv("VEV_PERF_PROCESS_ID", "test-process")
	145	+	t.Setenv("VEV_PERF_RUN", tc.rawRun)
	146	+
	147	+	newSink := func(string, ports.Clock, string) (ports.RuntimeObserver, io.Closer, error) {
	148	+	return runtimeObserverFunc(func(ports.RuntimeMark) {}), errorCloser{err: closeErr}, nil
	149	+	}
	150	+	newCorrelation := func(observer ports.RuntimeObserver, inputs ports.RuntimeCorrelationInputs) (ports.RuntimeObserver, error) {
	151	+	if tc.correlationErr != nil {
	152	+	return nil, tc.correlationErr
	153	+	}
	154	+	return ports.NewRuntimeCorrelationObserver(observer, inputs)
	155	+	}
	156	+
	157	+	observer, closer, err := performanceTraceWithFactories(nil, newSink, newCorrelation)
	158	+	if observer != nil \|\| closer != nil {
	159	+	t.Fatalf("performanceTrace() returned observer=%v closer=%v after setup failure", observer, closer)
	160	+	}
	161	+	if !errors.Is(err, closeErr) {
	162	+	t.Errorf("performanceTrace() error = %v, want joined rollback error %v", err, closeErr)
	163	+	}
	164	+	if tc.wantPrimaryErr != nil && !errors.Is(err, tc.wantPrimaryErr) {
	165	+	t.Errorf("performanceTrace() error = %v, want primary setup error %v", err, tc.wantPrimaryErr)
	166	+	}
	167	+	if tc.wantNumError {
	168	+	var numErr *strconv.NumError
	169	+	if !errors.As(err, &numErr) {
	170	+	t.Errorf("performanceTrace() error = %v, want discoverable parse error", err)
	171	+	}
	172	+	}
	173	+	if !strings.Contains(err.Error(), tc.wantPrimaryText) {
	174	+	t.Errorf("performanceTrace() error = %q, want context %q", err, tc.wantPrimaryText)
	175	+	}
	176	+	})
	177	+	}
	178	+	}
	179	+
	180	+	func TestRunAttachJoinsTraceCloseError(t *testing.T) {
	181	+	traceCloseErr := errors.New("trace close failed")
	182	+	original := newPerformanceTrace
	183	+	newPerformanceTrace = func(ports.Clock) (ports.SerializedRuntimeObserver, io.Closer, error) {
	184	+	return ports.NewSerializedRuntimeObserver(runtimeObserverFunc(func(ports.RuntimeMark) {}), 1), errorCloser{err: traceCloseErr}, nil
	185	+	}
	186	+	t.Cleanup(func() { newPerformanceTrace = original })
	187	+	t.Setenv("XDG_STATE_HOME", t.TempDir())
	188	+	t.Setenv("XDG_RUNTIME_DIR", t.TempDir())
	189	+	t.Setenv(envRemoteTransport, "invalid")
	190	+
	191	+	err := runAttach(context.Background(), ports.IntentEphemeral, "", "host")
	192	+	if !errors.Is(err, traceCloseErr) {
	193	+	t.Fatalf("runAttach() error = %v, want joined trace close error %v", err, traceCloseErr)
	194	+	}
	195	+	}
	196	+
	197	+	type errorCloser struct{ err error }
	198	+
	199	+	func (c errorCloser) Close() error { return c.err }
	200	+
	201	+	type runtimeObserverFunc func(ports.RuntimeMark)
	202	+
	203	+	func (f runtimeObserverFunc) ObserveRuntime(mark ports.RuntimeMark) { f(mark) }
	204	+
	205	+	func TestRunAttachPropagatesOneObserverToRemoteTransportFactory(t *testing.T) {
	206	+	t.Setenv("XDG_STATE_HOME", t.TempDir())
	207	+	t.Setenv("XDG_RUNTIME_DIR", t.TempDir())
	208	+	t.Setenv(envRemoteTransport, string(ports.RemoteTransportStdio))
	209	+
	210	+	observer := ports.NewSerializedRuntimeObserver(runtimeObserverFunc(func(ports.RuntimeMark) {}), 1)
	211	+	originalTrace := newPerformanceTrace
	212	+	traceCalls := 0
	213	+	newPerformanceTrace = func(ports.Clock) (ports.SerializedRuntimeObserver, io.Closer, error) {
	214	+	traceCalls++
	215	+	return observer, &runtimeTraceCloser{reporter: observer}, nil
	216	+	}
	217	+	t.Cleanup(func() { newPerformanceTrace = originalTrace })
	218	+
	219	+	factory := portsmocks.NewMockRemoteDialerFactory(t)
	220	+	factory.EXPECT().DialerForRemote("remote.example", "work", ports.RemoteTransportStdio, mock.Anything).Return(namedDialer{name: "remote"}, nil)
	221	+	originalFactory := newRemoteDialerFactoryWithRuntimeObserver
	222	+	factoryCalls := 0
	223	+	newRemoteDialerFactoryWithRuntimeObserver = func(got ports.SerializedRuntimeObserver) ports.RemoteDialerFactory {
	224	+	factoryCalls++
	225	+	if got != observer {
	226	+	t.Fatalf("remote transport observer = %v, want process observer %v", got, observer)
	227	+	}
	228	+	return factory
	229	+	}
	230	+	t.Cleanup(func() { newRemoteDialerFactoryWithRuntimeObserver = originalFactory })
	231	+
	232	+	originalRunClient := runClientWithRuntimeObserver
	233	+	runClientWithRuntimeObserver = func(_ context.Context, gotDialer ports.Dialer, _ ports.Terminal, _ ports.Clock, _ uint8, _ string, _ bool, _ ports.ClipboardReader, _ *slog.Logger, got ports.SerializedRuntimeObserver) error {
	234	+	if got != observer {
	235	+	t.Fatalf("client transport observer = %v, want process observer %v", got, observer)
	236	+	}
	237	+	if gotDialer != (namedDialer{name: "remote"}) {
	238	+	t.Fatalf("remote transport dialer = %v, want configured factory dialer", gotDialer)
	239	+	}
	240	+	return nil
	241	+	}
	242	+	t.Cleanup(func() { runClientWithRuntimeObserver = originalRunClient })
	243	+
	244	+	if err := runAttach(context.Background(), ports.IntentAttach, "work", "remote.example"); err != nil {
	245	+	t.Fatal(err)
	246	+	}
	247	+	if traceCalls != 1 \|\| factoryCalls != 1 {
	248	+	t.Fatalf("trace/factory calls = %d/%d, want exactly one observer and one transport factory", traceCalls, factoryCalls)
	249	+	}
	250	+	}

M internal/ports/mocks/mocks.go +93 -20

 }
 // Open provides a mock function for the type MockPTYFactory
-func (_mock *MockPTYFactory) Open(cmd string, args []string, env []string, dir string, sz domain.Size) (ports.PTY, error) {
+func (_mock *MockPTYFactory) Open(ctx context.Context, cmd string, args []string, env []string, dir string, sz domain.Size) (ports.PTY, error) {
-	ret := _mock.Called(cmd, args, env, dir, sz)
+	ret := _mock.Called(ctx, cmd, args, env, dir, sz)
 	if len(ret) == 0 {
 		panic("no return value specified for Open")
 	var r0 ports.PTY
 	var r1 error
-	if returnFunc, ok := ret.Get(0).(func(string, []string, []string, string, domain.Size) (ports.PTY, error)); ok {
+	if returnFunc, ok := ret.Get(0).(func(context.Context, string, []string, []string, string, domain.Size) (ports.PTY, error)); ok {
-		return returnFunc(cmd, args, env, dir, sz)
+		return returnFunc(ctx, cmd, args, env, dir, sz)
 	}
-	if returnFunc, ok := ret.Get(0).(func(string, []string, []string, string, domain.Size) ports.PTY); ok {
+	if returnFunc, ok := ret.Get(0).(func(context.Context, string, []string, []string, string, domain.Size) ports.PTY); ok {
-		r0 = returnFunc(cmd, args, env, dir, sz)
+		r0 = returnFunc(ctx, cmd, args, env, dir, sz)
 	} else {
 		if ret.Get(0) != nil {
 			r0 = ret.Get(0).(ports.PTY)
 		}
 	}
-	if returnFunc, ok := ret.Get(1).(func(string, []string, []string, string, domain.Size) error); ok {
+	if returnFunc, ok := ret.Get(1).(func(context.Context, string, []string, []string, string, domain.Size) error); ok {
-		r1 = returnFunc(cmd, args, env, dir, sz)
+		r1 = returnFunc(ctx, cmd, args, env, dir, sz)
 	} else {
 		r1 = ret.Error(1)
 	}
 }
 // Open is a helper method to define mock.On call
+//   - ctx context.Context
 //   - cmd string
 //   - args []string
 //   - env []string
 //   - dir string
 //   - sz domain.Size
-func (_e *MockPTYFactory_Expecter) Open(cmd any, args any, env any, dir any, sz any) *MockPTYFactory_Open_Call {
+func (_e *MockPTYFactory_Expecter) Open(ctx any, cmd any, args any, env any, dir any, sz any) *MockPTYFactory_Open_Call {
-	return &MockPTYFactory_Open_Call{Call: _e.mock.On("Open", cmd, args, env, dir, sz)}
+	return &MockPTYFactory_Open_Call{Call: _e.mock.On("Open", ctx, cmd, args, env, dir, sz)}
 }
-func (_c *MockPTYFactory_Open_Call) Run(run func(cmd string, args []string, env []string, dir string, sz domain.Size)) *MockPTYFactory_Open_Call {
+func (_c *MockPTYFactory_Open_Call) Run(run func(ctx context.Context, cmd string, args []string, env []string, dir string, sz domain.Size)) *MockPTYFactory_Open_Call {
 	_c.Call.Run(func(args mock.Arguments) {
-		var arg0 string
+		var arg0 context.Context
 		if args[0] != nil {
-			arg0 = args[0].(string)
+			arg0 = args[0].(context.Context)
 		}
-		var arg1 []string
+		var arg1 string
 		if args[1] != nil {
-			arg1 = args[1].([]string)
+			arg1 = args[1].(string)
 		}
 		var arg2 []string
 		if args[2] != nil {
 			arg2 = args[2].([]string)
 		}
-		var arg3 string
+		var arg3 []string
 		if args[3] != nil {
-			arg3 = args[3].(string)
+			arg3 = args[3].([]string)
 		}
-		var arg4 domain.Size
+		var arg4 string
 		if args[4] != nil {
-			arg4 = args[4].(domain.Size)
+			arg4 = args[4].(string)
+		}
+		var arg5 domain.Size
+		if args[5] != nil {
+			arg5 = args[5].(domain.Size)
 		}
 		run(
 			arg0,
 			arg2,
 			arg3,
 			arg4,
+			arg5,
 		)
 	})
 	return _c
 	return _c
 }
-func (_c *MockPTYFactory_Open_Call) RunAndReturn(run func(cmd string, args []string, env []string, dir string, sz domain.Size) (ports.PTY, error)) *MockPTYFactory_Open_Call {
+func (_c *MockPTYFactory_Open_Call) RunAndReturn(run func(ctx context.Context, cmd string, args []string, env []string, dir string, sz domain.Size) (ports.PTY, error)) *MockPTYFactory_Open_Call {
 	_c.Call.Return(run)
 	return _c
 }
 	_c.Call.Return(run)
 	return _c
 }
+// NewMockRuntimeObserver creates a new instance of MockRuntimeObserver. It also registers a testing interface on the mock and a cleanup function to assert the mocks expectations.
+// The first argument is typically a *testing.T value.
+func NewMockRuntimeObserver(t interface {
+	mock.TestingT
+	Cleanup(func())
+}) *MockRuntimeObserver {
+	mock := &MockRuntimeObserver{}
+	mock.Mock.Test(t)
+	t.Cleanup(func() { mock.AssertExpectations(t) })
+	return mock
+}
+// MockRuntimeObserver is an autogenerated mock type for the RuntimeObserver type
+type MockRuntimeObserver struct {
+	mock.Mock
+}
+type MockRuntimeObserver_Expecter struct {
+	mock *mock.Mock
+}
+func (_m *MockRuntimeObserver) EXPECT() *MockRuntimeObserver_Expecter {
+	return &MockRuntimeObserver_Expecter{mock: &_m.Mock}
+}
+// ObserveRuntime provides a mock function for the type MockRuntimeObserver
+func (_mock *MockRuntimeObserver) ObserveRuntime(runtimeMark ports.RuntimeMark) {
+	_mock.Called(runtimeMark)
+	return
+}
+// MockRuntimeObserver_ObserveRuntime_Call is a *mock.Call that shadows Run/Return methods with type explicit version for method 'ObserveRuntime'
+type MockRuntimeObserver_ObserveRuntime_Call struct {
+	*mock.Call
+}
+// ObserveRuntime is a helper method to define mock.On call
+//   - runtimeMark ports.RuntimeMark
+func (_e *MockRuntimeObserver_Expecter) ObserveRuntime(runtimeMark any) *MockRuntimeObserver_ObserveRuntime_Call {
+	return &MockRuntimeObserver_ObserveRuntime_Call{Call: _e.mock.On("ObserveRuntime", runtimeMark)}
+}
+func (_c *MockRuntimeObserver_ObserveRuntime_Call) Run(run func(runtimeMark ports.RuntimeMark)) *MockRuntimeObserver_ObserveRuntime_Call {
+	_c.Call.Run(func(args mock.Arguments) {
+		var arg0 ports.RuntimeMark
+		if args[0] != nil {
+			arg0 = args[0].(ports.RuntimeMark)
+		}
+		run(
+			arg0,
+		)
+	})
+	return _c
+}
+func (_c *MockRuntimeObserver_ObserveRuntime_Call) Return() *MockRuntimeObserver_ObserveRuntime_Call {
+	_c.Call.Return()
+	return _c
+}
+func (_c *MockRuntimeObserver_ObserveRuntime_Call) RunAndReturn(run func(runtimeMark ports.RuntimeMark)) *MockRuntimeObserver_ObserveRuntime_Call {
+	_c.Run(run)
+	return _c
+}

M internal/ports/ports.go +6 -2

 // PTYFactory creates PTYs by spawning a command attached to a new pseudo-terminal.
 type PTYFactory interface {
-	Open(cmd string, args []string, env []string, dir string, sz domain.Size) (PTY, error)
+	// Open creates a PTY while honoring ctx cancellation. Implementations must
+	// return promptly once ctx is cancelled and must not leave a child running.
+	Open(ctx context.Context, cmd string, args []string, env []string, dir string, sz domain.Size) (PTY, error)
 }
 // Terminal is the CLIENT-side controlling terminal.
 	Flush() error
 }
-// Transport is a framed message channel over a single connection.
+// Transport is a framed message channel over a single connection. Close must
+// be safe to call concurrently with Send and Recv, and must unblock active
+// Send and Recv calls.
 type Transport interface {
 	Send(Frame) error
 	Recv() (Frame, error) // blocking; io.EOF on close

A internal/ports/runtime_observer.go +298 -0

@@ -0,0 +1,298 @@
	1	+	package ports
	2	+
	3	+	import (
	4	+	"errors"
	5	+	"sync"
	6	+	"sync/atomic"
	7	+	)
	8	+
	9	+	// RuntimeMarkSchema is the version of the process-local performance trace.
	10	+	const RuntimeMarkSchema uint16 = 1
	11	+
	12	+	// RuntimeMarkKind identifies a process-local runtime boundary. It deliberately
	13	+	// contains no protocol payload, cells, or rendering policy.
	14	+	type RuntimeMarkKind string
	15	+
	16	+	const (
	17	+	RuntimeInputInjected RuntimeMarkKind = "input_injected"
	18	+	RuntimeResizeRequested RuntimeMarkKind = "resize_requested"
	19	+	RuntimeResizeCommitted RuntimeMarkKind = "resize_committed"
	20	+	RuntimeCaptureStart RuntimeMarkKind = "capture_start"
	21	+	RuntimeCaptureEnd RuntimeMarkKind = "capture_end"
	22	+	RuntimeComposeStart RuntimeMarkKind = "compose_start"
	23	+	RuntimeComposeEnd RuntimeMarkKind = "compose_end"
	24	+	RuntimeDiffStart RuntimeMarkKind = "diff_start"
	25	+	RuntimeDiffEnd RuntimeMarkKind = "diff_end"
	26	+	RuntimeQueueEnqueued RuntimeMarkKind = "queue_enqueued"
	27	+	RuntimeQueueDequeued RuntimeMarkKind = "queue_dequeued"
	28	+	RuntimeACKBlockedStart RuntimeMarkKind = "ack_blocked_start"
	29	+	RuntimeACKBlockedEnd RuntimeMarkKind = "ack_blocked_end"
	30	+	RuntimeEmitStart RuntimeMarkKind = "emit_start"
	31	+	RuntimeEmitEnd RuntimeMarkKind = "emit_end"
	32	+	RuntimeAdapterSendStart RuntimeMarkKind = "adapter_send_start"
	33	+	RuntimeAdapterSendEnd RuntimeMarkKind = "adapter_send_end"
	34	+	RuntimeAdapterReceiveStart RuntimeMarkKind = "adapter_receive_start"
	35	+	RuntimeAdapterReceiveEnd RuntimeMarkKind = "adapter_receive_end"
	36	+	RuntimeTerminalFlushed RuntimeMarkKind = "terminal_flushed"
	37	+	RuntimeTransportDiagnostic RuntimeMarkKind = "transport_diagnostic"
	38	+	)
	39	+
	40	+	// RuntimeMark is a process-local observation record. Tick is intentionally
	41	+	// left zero by producers: the concrete process observer owns timestamping.
	42	+	type RuntimeMark struct {
	43	+	Schema uint16
	44	+	ProcessID string
	45	+	Component string
	46	+	Scenario string
	47	+	Run uint64
	48	+	Sequence uint64
	49	+	RequestID uint64
	50	+	Epoch uint64
	51	+	Kind RuntimeMarkKind
	52	+	Tick int64
	53	+	Bytes uint64
	54	+	Fragments uint64
	55	+	Retransmits uint64
	56	+	Pending uint64
	57	+	AckRTTNanos int64
	58	+	Valid bool
	59	+	}
	60	+
	61	+	// RuntimeObserver receives process-local marks. Implementations must not let
	62	+	// a producer provide a timestamp in another clock domain.
	63	+	type RuntimeObserver interface{ ObserveRuntime(RuntimeMark) }
	64	+
	65	+	// SerializedRuntimeObserver is the only observer contract accepted by runtime
	66	+	// transport and use-case hot paths. ObserveRuntime must return without waiting
	67	+	// for sink I/O. Flush and Close wait until every accepted mark has reached the
	68	+	// concrete observer; Close is idempotent and rejects later marks.
	69	+	type SerializedRuntimeObserver interface {
	70	+	RuntimeObserver
	71	+	Flush()
	72	+	Close()
	73	+	}
	74	+
	75	+	// NewSerializedRuntimeObserver isolates producers from a potentially slow
	76	+	// observer with one ordered worker. capacity bounds queued marks. If the queue
	77	+	// fills, dropped marks are represented by one ordered diagnostic mark before
	78	+	// reporting resumes, so a trace can never silently claim completeness.
	79	+	func NewSerializedRuntimeObserver(observer RuntimeObserver, capacity int) SerializedRuntimeObserver {
	80	+	if observer == nil {
	81	+	return nil
	82	+	}
	83	+	if capacity < 1 {
	84	+	capacity = 1
	85	+	}
	86	+	o := &serializedRuntimeObserver{observer: observer, capacity: capacity, done: make(chan struct{})}
	87	+	o.cond = sync.NewCond(&o.mu)
	88	+	go o.run()
	89	+	return o
	90	+	}
	91	+
	92	+	// EnsureSerializedRuntimeObserver returns observer's existing serialized
	93	+	// owner when it has one. Otherwise it creates an owned worker. Callers close
	94	+	// the result only when owned is true, preventing nested workers and duplicate
	95	+	// lifecycle ownership when application wiring already serialized the process
	96	+	// observer.
	97	+	func EnsureSerializedRuntimeObserver(observer RuntimeObserver, capacity int) (serialized SerializedRuntimeObserver, owned bool) {
	98	+	if observer == nil {
	99	+	return nil, false
	100	+	}
	101	+	if serialized, ok := observer.(SerializedRuntimeObserver); ok {
	102	+	return serialized, false
	103	+	}
	104	+	return NewSerializedRuntimeObserver(observer, capacity), true
	105	+	}
	106	+
	107	+	type serializedRuntimeObserver struct {
	108	+	observer RuntimeObserver
	109	+	capacity int
	110	+
	111	+	mu sync.Mutex
	112	+	cond *sync.Cond
	113	+	marks []RuntimeMark
	114	+	dropped uint64
	115	+	closed bool
	116	+	active bool
	117	+	done chan struct{}
	118	+	}
	119	+
	120	+	func (o *serializedRuntimeObserver) ObserveRuntime(mark RuntimeMark) {
	121	+	if o == nil {
	122	+	return
	123	+	}
	124	+	o.mu.Lock()
	125	+	defer o.mu.Unlock()
	126	+	if o.closed {
	127	+	return
	128	+	}
	129	+	if o.dropped != 0 {
	130	+	// Reserve space for both the gap marker and the first post-gap mark.
	131	+	// This preserves the producer order around a bounded-queue loss.
	132	+	if len(o.marks)+2 > o.capacity {
	133	+	o.dropped++
	134	+	return
	135	+	}
	136	+	o.marks = append(o.marks, runtimeObserverGapMark(o.dropped), mark)
	137	+	o.dropped = 0
	138	+	o.cond.Signal()
	139	+	return
	140	+	}
	141	+	if len(o.marks) == o.capacity {
	142	+	o.dropped = 1
	143	+	return
	144	+	}
	145	+	o.marks = append(o.marks, mark)
	146	+	o.cond.Signal()
	147	+	}
	148	+
	149	+	func runtimeObserverGapMark(dropped uint64) RuntimeMark {
	150	+	return NewRuntimeMark("observer", RuntimeTransportDiagnostic, dropped, false)
	151	+	}
	152	+
	153	+	func (o *serializedRuntimeObserver) Flush() {
	154	+	if o == nil {
	155	+	return
	156	+	}
	157	+	o.mu.Lock()
	158	+	for len(o.marks) != 0 \|\| o.active \|\| o.dropped != 0 {
	159	+	o.cond.Signal()
	160	+	o.cond.Wait()
	161	+	}
	162	+	closed := o.closed
	163	+	o.mu.Unlock()
	164	+	if closed {
	165	+	<-o.done
	166	+	}
	167	+	}
	168	+
	169	+	func (o *serializedRuntimeObserver) Close() {
	170	+	if o == nil {
	171	+	return
	172	+	}
	173	+	o.mu.Lock()
	174	+	if !o.closed {
	175	+	o.closed = true
	176	+	o.cond.Broadcast()
	177	+	}
	178	+	o.mu.Unlock()
	179	+	<-o.done
	180	+	}
	181	+
	182	+	func (o *serializedRuntimeObserver) run() {
	183	+	defer close(o.done)
	184	+	for {
	185	+	o.mu.Lock()
	186	+	for len(o.marks) == 0 && o.dropped == 0 && !o.closed {
	187	+	o.cond.Wait()
	188	+	}
	189	+	if len(o.marks) == 0 && o.dropped != 0 {
	190	+	o.marks = append(o.marks, runtimeObserverGapMark(o.dropped))
	191	+	o.dropped = 0
	192	+	}
	193	+	if len(o.marks) == 0 && o.closed {
	194	+	o.cond.Broadcast()
	195	+	o.mu.Unlock()
	196	+	return
	197	+	}
	198	+	mark := o.marks[0]
	199	+	o.marks = o.marks[1:]
	200	+	o.active = true
	201	+	o.mu.Unlock()
	202	+
	203	+	o.observer.ObserveRuntime(mark)
	204	+
	205	+	o.mu.Lock()
	206	+	o.active = false
	207	+	o.cond.Broadcast()
	208	+	o.mu.Unlock()
	209	+	}
	210	+	}
	211	+
	212	+	// RuntimeMarkValid reports whether a mark has all correlation fields required
	213	+	// by the JSONL schema. Tick is checked separately by the timestamp owner.
	214	+	func RuntimeMarkValid(mark RuntimeMark) bool {
	215	+	if mark.Schema != RuntimeMarkSchema \|\| mark.Component == "" \|\| mark.Scenario == "" \|\| mark.Run == 0 \|\| mark.Sequence == 0 \|\| mark.RequestID == 0 \|\| mark.Epoch == 0 {
	216	+	return false
	217	+	}
	218	+	switch mark.Kind {
	219	+	case RuntimeInputInjected, RuntimeResizeRequested, RuntimeResizeCommitted,
	220	+	RuntimeCaptureStart, RuntimeCaptureEnd, RuntimeComposeStart, RuntimeComposeEnd,
	221	+	RuntimeDiffStart, RuntimeDiffEnd, RuntimeQueueEnqueued, RuntimeQueueDequeued,
	222	+	RuntimeACKBlockedStart, RuntimeACKBlockedEnd, RuntimeEmitStart, RuntimeEmitEnd,
	223	+	RuntimeAdapterSendStart, RuntimeAdapterSendEnd, RuntimeAdapterReceiveStart,
	224	+	RuntimeAdapterReceiveEnd, RuntimeTerminalFlushed, RuntimeTransportDiagnostic:
	225	+	return true
	226	+	default:
	227	+	return false
	228	+	}
	229	+	}
	230	+
	231	+	// RuntimeCorrelation identifies one attempted operation. It is deliberately
	232	+	// copied to both endpoints of a span; process ID remains owned by the sink.
	233	+	type RuntimeCorrelation struct {
	234	+	Scenario string
	235	+	Run, Sequence, RequestID, Epoch uint64
	236	+	}
	237	+
	238	+	// RuntimeCorrelationInputs are supplied once by process orchestration. They
	239	+	// are deliberately data-only: consumer APIs continue to receive only their
	240	+	// one RuntimeObserver argument and never a clock or tick.
	241	+	type RuntimeCorrelationInputs struct {
	242	+	Scenario string
	243	+	Run uint64
	244	+	}
	245	+
	246	+	// NewRuntimeCorrelationObserver binds every mark from one OS process to its
	247	+	// harness manifest scenario/run. Operation IDs remain producer-owned when
	248	+	// supplied; missing IDs are assigned from the process-wide allocator.
	249	+	func NewRuntimeCorrelationObserver(observer RuntimeObserver, inputs RuntimeCorrelationInputs) (RuntimeObserver, error) {
	250	+	if observer == nil \|\| inputs.Scenario == "" \|\| inputs.Run == 0 {
	251	+	return nil, errors.New("invalid runtime correlation inputs")
	252	+	}
	253	+	return &runtimeCorrelationObserver{observer: observer, inputs: inputs}, nil
	254	+	}
	255	+
	256	+	type runtimeCorrelationObserver struct {
	257	+	observer RuntimeObserver
	258	+	inputs RuntimeCorrelationInputs
	259	+	}
	260	+
	261	+	func (o *runtimeCorrelationObserver) ObserveRuntime(mark RuntimeMark) {
	262	+	if o == nil \|\| o.observer == nil {
	263	+	return
	264	+	}
	265	+	mark.Scenario, mark.Run = o.inputs.Scenario, o.inputs.Run
	266	+	if mark.Sequence == 0 {
	267	+	mark.Sequence = runtimeMarkSequence.Add(1)
	268	+	}
	269	+	if mark.RequestID == 0 {
	270	+	mark.RequestID = mark.Sequence
	271	+	}
	272	+	if mark.Epoch == 0 {
	273	+	mark.Epoch = mark.RequestID
	274	+	}
	275	+	o.observer.ObserveRuntime(mark)
	276	+	}
	277	+
	278	+	var runtimeMarkSequence atomic.Uint64
	279	+
	280	+	// NewRuntimeCorrelation makes a unique, process-local operation identity.
	281	+	// Components with a domain request/epoch should use those values directly via
	282	+	// NewRuntimeMarkWithCorrelation instead of inventing a second identity.
	283	+	func NewRuntimeCorrelation() RuntimeCorrelation {
	284	+	sequence := runtimeMarkSequence.Add(1)
	285	+	return RuntimeCorrelation{Scenario: "runtime", Run: 1, Sequence: sequence, RequestID: sequence, Epoch: sequence}
	286	+	}
	287	+
	288	+	// NewRuntimeMarkWithCorrelation returns a timestamp-free mark carrying the
	289	+	// supplied operation identity. Start/end callers must reuse correlation.
	290	+	func NewRuntimeMarkWithCorrelation(component string, correlation RuntimeCorrelation, kind RuntimeMarkKind, bytes uint64, valid bool) RuntimeMark {
	291	+	return RuntimeMark{Schema: RuntimeMarkSchema, Component: component, Scenario: correlation.Scenario, Run: correlation.Run, Sequence: correlation.Sequence, RequestID: correlation.RequestID, Epoch: correlation.Epoch, Kind: kind, Bytes: bytes, Valid: valid}
	292	+	}
	293	+
	294	+	// NewRuntimeMark is retained for one-shot marks. Unlike the old helper it
	295	+	// never returns constant correlation IDs.
	296	+	func NewRuntimeMark(component string, kind RuntimeMarkKind, bytes uint64, valid bool) RuntimeMark {
	297	+	return NewRuntimeMarkWithCorrelation(component, NewRuntimeCorrelation(), kind, bytes, valid)
	298	+	}

A internal/ports/runtime_observer_test.go +204 -0

@@ -0,0 +1,204 @@
	1	+	package ports
	2	+
	3	+	import (
	4	+	"reflect"
	5	+	"sync"
	6	+	"testing"
	7	+	"time"
	8	+	)
	9	+
	10	+	func TestRuntimeObserverContract(t *testing.T) {
	11	+	wantFields := []struct {
	12	+	name string
	13	+	typ reflect.Type
	14	+	}{
	15	+	{"Schema", reflect.TypeFor[uint16]()},
	16	+	{"ProcessID", reflect.TypeFor[string]()},
	17	+	{"Component", reflect.TypeFor[string]()},
	18	+	{"Scenario", reflect.TypeFor[string]()},
	19	+	{"Run", reflect.TypeFor[uint64]()},
	20	+	{"Sequence", reflect.TypeFor[uint64]()},
	21	+	{"RequestID", reflect.TypeFor[uint64]()},
	22	+	{"Epoch", reflect.TypeFor[uint64]()},
	23	+	{"Kind", reflect.TypeFor[RuntimeMarkKind]()},
	24	+	{"Tick", reflect.TypeFor[int64]()},
	25	+	{"Bytes", reflect.TypeFor[uint64]()},
	26	+	{"Fragments", reflect.TypeFor[uint64]()},
	27	+	{"Retransmits", reflect.TypeFor[uint64]()},
	28	+	{"Pending", reflect.TypeFor[uint64]()},
	29	+	{"AckRTTNanos", reflect.TypeFor[int64]()},
	30	+	{"Valid", reflect.TypeFor[bool]()},
	31	+	}
	32	+
	33	+	typ := reflect.TypeFor[RuntimeMark]()
	34	+	if typ.NumField() != len(wantFields) {
	35	+	t.Fatalf("RuntimeMark has %d fields, want exactly %d", typ.NumField(), len(wantFields))
	36	+	}
	37	+	for i, want := range wantFields {
	38	+	field := typ.Field(i)
	39	+	if field.Name != want.name \|\| field.Type != want.typ {
	40	+	t.Fatalf("RuntimeMark field %d = %s %v, want %s %v", i, field.Name, field.Type, want.name, want.typ)
	41	+	}
	42	+	}
	43	+
	44	+	var _ RuntimeObserver = runtimeObserverFunc(func(RuntimeMark) {})
	45	+	}
	46	+
	47	+	func TestSerializedRuntimeObserverFlushesInOrderAndCloses(t *testing.T) {
	48	+	var mu sync.Mutex
	49	+	var got []RuntimeMark
	50	+	observer := NewSerializedRuntimeObserver(runtimeObserverFunc(func(mark RuntimeMark) {
	51	+	mu.Lock()
	52	+	got = append(got, mark)
	53	+	mu.Unlock()
	54	+	}), 2)
	55	+	observer.ObserveRuntime(NewRuntimeMark("client", RuntimeEmitStart, 1, true))
	56	+	observer.ObserveRuntime(NewRuntimeMark("client", RuntimeEmitEnd, 1, true))
	57	+	observer.Flush()
	58	+	observer.Close()
	59	+	observer.Close()
	60	+
	61	+	mu.Lock()
	62	+	defer mu.Unlock()
	63	+	if len(got) != 2 \|\| got[0].Kind != RuntimeEmitStart \|\| got[1].Kind != RuntimeEmitEnd {
	64	+	t.Fatalf("serialized marks = %#v", got)
	65	+	}
	66	+	}
	67	+
	68	+	func TestEnsureSerializedRuntimeObserverReusesExistingOwner(t *testing.T) {
	69	+	existing := NewSerializedRuntimeObserver(runtimeObserverFunc(func(RuntimeMark) {}), 1)
	70	+	reporter, owned := EnsureSerializedRuntimeObserver(existing, 1)
	71	+	if owned \|\| reporter != existing {
	72	+	t.Fatalf("EnsureSerializedRuntimeObserver() = (%T, %t), want existing unowned reporter", reporter, owned)
	73	+	}
	74	+	existing.Close()
	75	+	}
	76	+
	77	+	func TestSerializedRuntimeObserverReportsBoundedQueueLoss(t *testing.T) {
	78	+	entered := make(chan struct{})
	79	+	release := make(chan struct{})
	80	+	var mu sync.Mutex
	81	+	var got []RuntimeMark
	82	+	observer := NewSerializedRuntimeObserver(runtimeObserverFunc(func(mark RuntimeMark) {
	83	+	if mark.Kind == RuntimeEmitStart {
	84	+	close(entered)
	85	+	<-release
	86	+	}
	87	+	mu.Lock()
	88	+	got = append(got, mark)
	89	+	mu.Unlock()
	90	+	}), 2)
	91	+	observer.ObserveRuntime(NewRuntimeMark("client", RuntimeEmitStart, 1, true))
	92	+	select {
	93	+	case <-entered:
	94	+	case <-time.After(time.Second):
	95	+	t.Fatal("observer did not start")
	96	+	}
	97	+	observer.ObserveRuntime(NewRuntimeMark("client", RuntimeComposeStart, 1, true))
	98	+	observer.ObserveRuntime(NewRuntimeMark("client", RuntimeComposeEnd, 1, true))
	99	+	observer.ObserveRuntime(NewRuntimeMark("client", RuntimeEmitEnd, 1, true)) // bounded loss
	100	+	close(release)
	101	+	observer.Flush()
	102	+	observer.Close()
	103	+
	104	+	mu.Lock()
	105	+	defer mu.Unlock()
	106	+	if len(got) != 4 \|\| got[0].Kind != RuntimeEmitStart \|\| got[1].Kind != RuntimeComposeStart \|\| got[2].Kind != RuntimeComposeEnd \|\| got[3].Kind != RuntimeTransportDiagnostic \|\| got[3].Bytes != 1 \|\| got[3].Valid {
	107	+	t.Fatalf("bounded-loss marks = %#v", got)
	108	+	}
	109	+	}
	110	+
	111	+	func TestRuntimeObserverRequiredSpanKinds(t *testing.T) {
	112	+	pairs := []struct {
	113	+	start RuntimeMarkKind
	114	+	end RuntimeMarkKind
	115	+	}{
	116	+	{RuntimeDiffStart, RuntimeDiffEnd},
	117	+	{RuntimeQueueEnqueued, RuntimeQueueDequeued},
	118	+	{RuntimeACKBlockedStart, RuntimeACKBlockedEnd},
	119	+	{RuntimeAdapterSendStart, RuntimeAdapterSendEnd},
	120	+	{RuntimeAdapterReceiveStart, RuntimeAdapterReceiveEnd},
	121	+	}
	122	+	for _, pair := range pairs {
	123	+	if pair.start == "" \|\| pair.end == "" \|\| pair.start == pair.end {
	124	+	t.Fatalf("invalid required span mapping %#v", pair)
	125	+	}
	126	+	}
	127	+	}
	128	+
	129	+	func TestRuntimeCorrelationObserverUsesHarnessProcessInputs(t *testing.T) {
	130	+	var got []RuntimeMark
	131	+	observer, err := NewRuntimeCorrelationObserver(runtimeObserverFunc(func(mark RuntimeMark) {
	132	+	got = append(got, mark)
	133	+	}), RuntimeCorrelationInputs{Scenario: "1x4-idle-local", Run: 7})
	134	+	if err != nil {
	135	+	t.Fatal(err)
	136	+	}
	137	+
	138	+	observer.ObserveRuntime(RuntimeMark{Schema: RuntimeMarkSchema, Component: "daemon", Scenario: "runtime", Run: 1, Kind: RuntimeEmitEnd, Valid: true})
	139	+	observer.ObserveRuntime(RuntimeMark{Schema: RuntimeMarkSchema, Component: "daemon", Scenario: "runtime", Run: 1, Sequence: 9, RequestID: 10, Epoch: 11, Kind: RuntimeEmitStart, Valid: true})
	140	+
	141	+	if len(got) != 2 {
	142	+	t.Fatalf("marks=%+v", got)
	143	+	}
	144	+	for _, mark := range got {
	145	+	if mark.Scenario != "1x4-idle-local" \|\| mark.Run != 7 {
	146	+	t.Fatalf("harness correlation was not applied: %+v", mark)
	147	+	}
	148	+	if mark.Sequence == 0 \|\| mark.RequestID == 0 \|\| mark.Epoch == 0 {
	149	+	t.Fatalf("missing deterministic operation identity: %+v", mark)
	150	+	}
	151	+	}
	152	+	if got[1].Sequence != 9 \|\| got[1].RequestID != 10 \|\| got[1].Epoch != 11 {
	153	+	t.Fatalf("explicit operation identity changed: %+v", got[1])
	154	+	}
	155	+	}
	156	+
	157	+	func TestRuntimeCorrelationObserverAllocatesMissingIDsProcessWide(t *testing.T) {
	158	+	var got []RuntimeMark
	159	+	collect := runtimeObserverFunc(func(mark RuntimeMark) { got = append(got, mark) })
	160	+	first, err := NewRuntimeCorrelationObserver(collect, RuntimeCorrelationInputs{Scenario: "scenario", Run: 1})
	161	+	if err != nil {
	162	+	t.Fatal(err)
	163	+	}
	164	+	second, err := NewRuntimeCorrelationObserver(collect, RuntimeCorrelationInputs{Scenario: "scenario", Run: 1})
	165	+	if err != nil {
	166	+	t.Fatal(err)
	167	+	}
	168	+
	169	+	first.ObserveRuntime(RuntimeMark{Schema: RuntimeMarkSchema, Component: "first", Kind: RuntimeEmitStart, Valid: true})
	170	+	second.ObserveRuntime(RuntimeMark{Schema: RuntimeMarkSchema, Component: "second", Kind: RuntimeEmitEnd, Valid: true})
	171	+
	172	+	if len(got) != 2 {
	173	+	t.Fatalf("marks=%+v", got)
	174	+	}
	175	+	if got[0].Sequence == got[1].Sequence {
	176	+	t.Fatalf("wrapper-local sequence collision: %+v", got)
	177	+	}
	178	+	for _, mark := range got {
	179	+	if mark.RequestID != mark.Sequence \|\| mark.Epoch != mark.RequestID {
	180	+	t.Fatalf("missing-ID fallback changed: %+v", mark)
	181	+	}
	182	+	}
	183	+	}
	184	+
	185	+	func TestRuntimeCorrelationObserverRejectsInvalidHarnessInputs(t *testing.T) {
	186	+	for _, in := range []RuntimeCorrelationInputs{{}, {Scenario: "scenario"}} {
	187	+	if _, err := NewRuntimeCorrelationObserver(runtimeObserverFunc(func(RuntimeMark) {}), in); err == nil {
	188	+	t.Fatalf("invalid inputs accepted: %+v", in)
	189	+	}
	190	+	}
	191	+	}
	192	+
	193	+	func TestRuntimeMarkHasNoRenderingPayloadOrPolicyFields(t *testing.T) {
	194	+	typ := reflect.TypeFor[RuntimeMark]()
	195	+	for _, forbidden := range []string{"Cell", "Cells", "Payload", "Frame", "Renderer", "Policy", "Transport"} {
	196	+	if _, ok := typ.FieldByName(forbidden); ok {
	197	+	t.Fatalf("RuntimeMark must not carry %s", forbidden)
	198	+	}
	199	+	}
	200	+	}
	201	+
	202	+	type runtimeObserverFunc func(RuntimeMark)
	203	+
	204	+	func (f runtimeObserverFunc) ObserveRuntime(mark RuntimeMark) { f(mark) }

A internal/testutil/replaytest/replay.go +64 -0

@@ -0,0 +1,64 @@
	1	+	// Package replaytest owns the protocol-neutral transport replay fixture and
	2	+	// assertions shared by transport adapter tests.
	3	+	package replaytest
	4	+
	5	+	import (
	6	+	"bytes"
	7	+	"reflect"
	8	+	"testing"
	9	+
	10	+	"github.com/bnema/vev/internal/ports"
	11	+	)
	12	+
	13	+	var transcript = []ports.Frame{
	14	+	{Type: ports.MsgOutput, Payload: ports.MarshalOutput(ports.Output{BaseStateNum: 0, NewStateNum: 1, Data: []byte("\x1b[2J\x1b[Hone\r\ntwo")})},
	15	+	{Type: ports.MsgOutput, Payload: ports.MarshalOutput(ports.Output{BaseStateNum: 1, NewStateNum: 2, EchoAck: 7, Data: []byte("\x1b[2;1HTWO")})},
	16	+	}
	17	+
	18	+	// Transcript returns a deep copy of the canonical replay transcript so an
	19	+	// adapter cannot mutate the fixture observed by another adapter.
	20	+	func Transcript() []ports.Frame {
	21	+	return cloneFrames(transcript)
	22	+	}
	23	+
	24	+	// Exchange carries already-composed frames through one transport adapter.
	25	+	type Exchange func(t *testing.T, frames []ports.Frame) []ports.Frame
	26	+
	27	+	// Run verifies that an adapter preserves the canonical transcript exactly.
	28	+	func Run(t *testing.T, exchange Exchange) {
	29	+	t.Helper()
	30	+	want := Transcript()
	31	+	got := exchange(t, cloneFrames(want))
	32	+	if len(got) != len(want) {
	33	+	t.Fatalf("replayed frame count = %d, want %d", len(got), len(want))
	34	+	}
	35	+	for i := range want {
	36	+	if got[i].Type != want[i].Type {
	37	+	t.Errorf("frame %d type = %d, want %d", i, got[i].Type, want[i].Type)
	38	+	}
	39	+	if !bytes.Equal(got[i].Payload, want[i].Payload) {
	40	+	t.Errorf("frame %d payload = %x, want byte-for-byte %x", i, got[i].Payload, want[i].Payload)
	41	+	continue
	42	+	}
	43	+	gotOutput, err := ports.UnmarshalOutput(got[i].Payload)
	44	+	if err != nil {
	45	+	t.Errorf("frame %d output decode: %v", i, err)
	46	+	continue
	47	+	}
	48	+	wantOutput, err := ports.UnmarshalOutput(want[i].Payload)
	49	+	if err != nil {
	50	+	t.Fatalf("canonical frame %d output decode: %v", i, err)
	51	+	}
	52	+	if !reflect.DeepEqual(gotOutput, wantOutput) {
	53	+	t.Errorf("frame %d output = %#v, want %#v", i, gotOutput, wantOutput)
	54	+	}
	55	+	}
	56	+	}
	57	+
	58	+	func cloneFrames(frames []ports.Frame) []ports.Frame {
	59	+	cloned := make([]ports.Frame, len(frames))
	60	+	for i, frame := range frames {
	61	+	cloned[i] = ports.Frame{Type: frame.Type, Payload: append([]byte(nil), frame.Payload...)}
	62	+	}
	63	+	return cloned
	64	+	}

M internal/usecase/client/client.go +105 -14


@@ -102,6 +102,10 @@ const stdinBufSize = 4096
102		// goroutine, decoupling the input/resize pumps from transport back-pressure.	102		// goroutine, decoupling the input/resize pumps from transport back-pressure.
103		const sendQueueDepth = 64	103		const sendQueueDepth = 64
104			104
			105	+	// preWelcomeTimeout bounds each blocking operation before the daemon has
			106	+	// accepted the attach. Mosh uses the same 15-second startup budget.
			107	+	const preWelcomeTimeout = 15 * time.Second
			108	+
105		var reconnectSleep = sleepReconnect	109		var reconnectSleep = sleepReconnect
106		var reconnectSleepWithResize = sleepReconnectWithResizeEvents	110		var reconnectSleepWithResize = sleepReconnectWithResizeEvents
107			111

@@ -120,7 +124,29 @@ const (
120		// ClipboardReader configured (clipboard non-nil) — local attaches forward	124		// ClipboardReader configured (clipboard non-nil) — local attaches forward
121		// 0x16 untouched so the locally running agent's own clipboard handling	125		// 0x16 untouched so the locally running agent's own clipboard handling
122		// applies.	126		// applies.
123	-	func Run(ctx context.Context, dialer ports.Dialer, term ports.Terminal, clk ports.Clock, intent uint8, name string, remote bool, clipboard ports.ClipboardReader, log *slog.Logger) (retErr error) {	127	+	// Option configures opt-in client runtime observation.
			128	+	type Option func(*runtimeOptions)
			129	+
			130	+	type runtimeOptions struct {
			131	+	observer ports.SerializedRuntimeObserver
			132	+	}
			133	+
			134	+	// WithRuntimeObserver accepts only a composition-root serialized observer, so
			135	+	// terminal and transport progress can never wait for a raw trace sink.
			136	+	func WithRuntimeObserver(observer ports.SerializedRuntimeObserver) Option {
			137	+	return func(opts *runtimeOptions) { opts.observer = observer }
			138	+	}
			139	+
			140	+	func Run(ctx context.Context, dialer ports.Dialer, term ports.Terminal, clk ports.Clock, intent uint8, name string, remote bool, clipboard ports.ClipboardReader, log *slog.Logger) error {
			141	+	return run(ctx, dialer, term, clk, intent, name, remote, clipboard, log, nil)
			142	+	}
			143	+
			144	+	// RunWithRuntimeObserver is the application wiring entry point.
			145	+	func RunWithRuntimeObserver(ctx context.Context, dialer ports.Dialer, term ports.Terminal, clk ports.Clock, intent uint8, name string, remote bool, clipboard ports.ClipboardReader, log *slog.Logger, observer ports.SerializedRuntimeObserver) error {
			146	+	return run(ctx, dialer, term, clk, intent, name, remote, clipboard, log, observer)
			147	+	}
			148	+
			149	+	func run(ctx context.Context, dialer ports.Dialer, term ports.Terminal, clk ports.Clock, intent uint8, name string, remote bool, clipboard ports.ClipboardReader, log *slog.Logger, observer ports.SerializedRuntimeObserver) (retErr error) {
124		if log == nil {	150		if log == nil {
125		log = slog.Default()	151		log = slog.Default()
126		}	152		}

@@ -244,11 +270,13 @@ func Run(ctx context.Context, dialer ports.Dialer, term ports.Terminal, clk port
244		if reporter, ok := transport.(ports.LinkStateReporter); ok {	270		if reporter, ok := transport.(ports.LinkStateReporter); ok {
245		linkEvents = reporter.LinkEvents()	271		linkEvents = reporter.LinkEvents()
246		}	272		}
247	-	result := attachOnce(ctx, transport, term, clk, attemptIntent, attemptName, resumeToken, processClientID, &ms, enterRaw, clearStatus, drawStatusStage, linkEvents, remote, clipboard, log)	273	+	result := attachOnce(ctx, transport, term, clk, attemptIntent, attemptName, resumeToken, processClientID, &ms, enterRaw, clearStatus, drawStatusStage, linkEvents, remote, clipboard, log, observer)
248		if result.welcomed {	274		if result.welcomed {
249		backoff = defaultReconnectBackoff.initial	275		backoff = defaultReconnectBackoff.initial
250		}	276		}
251	-	_ = transport.Close()	277	+	if !result.transportClosed {
			278	+	_ = transport.Close()
			279	+	}
252		if result.resumeToken != 0 {	280		if result.resumeToken != 0 {
253		resumeToken = result.resumeToken	281		resumeToken = result.resumeToken
254		}	282		}

@@ -311,8 +339,14 @@ func sleepReconnectWithResizeEvents(ctx context.Context, clk ports.Clock, d time
311		// disappears, or the context is cancelled.	339		// disappears, or the context is cancelled.
312		//	340		//
313		// It is kept as a compatibility wrapper for callers that still own dialing.	341		// It is kept as a compatibility wrapper for callers that still own dialing.
314	-	func Attach(ctx context.Context, transport ports.Transport, term ports.Terminal, clk ports.Clock, intent uint8, name string) error {	342	+	func Attach(ctx context.Context, transport ports.Transport, term ports.Terminal, clk ports.Clock, intent uint8, name string, options ...Option) error {
315	-	return Run(ctx, singleTransportDialer{transport: transport}, term, clk, intent, name, false, nil, slog.Default())	343	+	var opts runtimeOptions
			344	+	for _, option := range options {
			345	+	if option != nil {
			346	+	option(&opts)
			347	+	}
			348	+	}
			349	+	return run(ctx, singleTransportDialer{transport: transport}, term, clk, intent, name, false, nil, slog.Default(), opts.observer)
316		}	350		}
317			351
318		type singleTransportDialer struct{ transport ports.Transport }	352		type singleTransportDialer struct{ transport ports.Transport }

@@ -322,10 +356,52 @@ func (d singleTransportDialer) Dial(context.Context) (ports.Transport, error) {
322		}	356		}
323			357
324		type attachResult struct {	358		type attachResult struct {
325	-	resumeToken uint64	359	+	resumeToken uint64
326	-	sessionName string	360	+	sessionName string
327	-	welcomed bool	361	+	welcomed bool
328	-	err error	362	+	transportClosed bool
			363	+	err error
			364	+	}
			365	+
			366	+	// boundedPreWelcome runs one blocking handshake operation. Transport methods
			367	+	// do not accept a context, so cancellation and expiry close this attempt's
			368	+	// transport to interrupt the call, then wait for its goroutine to exit.
			369	+	func boundedPreWelcome(ctx context.Context, clk ports.Clock, transport ports.Transport, operation func() error) (bool, error) {
			370	+	if err := ctx.Err(); err != nil {
			371	+	_ = transport.Close()
			372	+	return true, err
			373	+	}
			374	+
			375	+	timer := clk.NewTimer(preWelcomeTimeout)
			376	+	defer timer.Stop()
			377	+	completed := make(chan error, 1)
			378	+	go func() { completed <- operation() }()
			379	+
			380	+	select {
			381	+	case err := <-completed:
			382	+	// Prefer a cancellation/expiry observed concurrently with completion.
			383	+	// That event owns the transport lifetime even though the operation has
			384	+	// already returned by the time we observe it here.
			385	+	if ctx.Err() != nil {
			386	+	_ = transport.Close()
			387	+	return true, ctx.Err()
			388	+	}
			389	+	select {
			390	+	case <-timer.C():
			391	+	_ = transport.Close()
			392	+	return true, context.DeadlineExceeded
			393	+	default:
			394	+	return false, err
			395	+	}
			396	+	case <-ctx.Done():
			397	+	_ = transport.Close()
			398	+	<-completed
			399	+	return true, ctx.Err()
			400	+	case <-timer.C():
			401	+	_ = transport.Close()
			402	+	<-completed
			403	+	return true, context.DeadlineExceeded
			404	+	}
329		}	405		}
330			406
331		// DetectTrueColor reports whether TERM/COLORTERM advertise direct color support.	407		// DetectTrueColor reports whether TERM/COLORTERM advertise direct color support.

@@ -346,7 +422,11 @@ func requestedOutputWindow(transport ports.Transport) uint8 {
346		return 8	422		return 8
347		}	423		}
348			424
349	-	func attachOnce(ctx context.Context, transport ports.Transport, term ports.Terminal, clk ports.Clock, intent uint8, name string, resumeToken uint64, clientID [16]byte, ms milestones, enterRaw func() error, clearStatus func(), drawStatusStage func(reconnectStage), linkEvents <-chan ports.LinkEvent, remote bool, clipboard ports.ClipboardReader, log slog.Logger) attachResult {	425	+	func attachOnce(ctx context.Context, transport ports.Transport, term ports.Terminal, clk ports.Clock, intent uint8, name string, resumeToken uint64, clientID [16]byte, ms milestones, enterRaw func() error, clearStatus func(), drawStatusStage func(reconnectStage), linkEvents <-chan ports.LinkEvent, remote bool, clipboard ports.ClipboardReader, log slog.Logger, observers ...ports.RuntimeObserver) attachResult {
			426	+	var observer ports.RuntimeObserver
			427	+	if len(observers) != 0 {
			428	+	observer = observers[0]
			429	+	}
350		// 1. Handshake: send Hello with our size and TERM.	430		// 1. Handshake: send Hello with our size and TERM.
351		size, err := term.Size()	431		size, err := term.Size()
352		if err != nil {	432		if err != nil {

@@ -371,15 +451,22 @@ func attachOnce(ctx context.Context, transport ports.Transport, term ports.Termi
371		TrueColor: trueColor,	451		TrueColor: trueColor,
372		MaxOutputInFlight: requestedOutputWindow(transport),	452		MaxOutputInFlight: requestedOutputWindow(transport),
373		}	453		}
374	-	if err := transport.Send(ports.Frame{Type: ports.MsgHello, Payload: ports.MarshalHello(hello)}); err != nil {	454	+	if closed, err := boundedPreWelcome(ctx, clk, transport, func() error {
375	-	return attachResult{err: fmt.Errorf("vev: sending hello: %w", err)}	455	+	return transport.Send(ports.Frame{Type: ports.MsgHello, Payload: ports.MarshalHello(hello)})
			456	+	}); err != nil {
			457	+	return attachResult{transportClosed: closed, err: fmt.Errorf("vev: sending hello: %w", err)}
376		}	458		}
377		ms.helloSent = true	459		ms.helloSent = true
378			460
379		// 2. Await Welcome or a typed rejection.	461		// 2. Await Welcome or a typed rejection.
380	-	reply, err := transport.Recv()	462	+	var reply ports.Frame
			463	+	closed, err := boundedPreWelcome(ctx, clk, transport, func() error {
			464	+	var recvErr error
			465	+	reply, recvErr = transport.Recv()
			466	+	return recvErr
			467	+	})
381		if err != nil {	468		if err != nil {
382	-	return attachResult{err: fmt.Errorf("vev: awaiting welcome: %w", err)}	469	+	return attachResult{transportClosed: closed, err: fmt.Errorf("vev: awaiting welcome: %w", err)}
383		}	470		}
384		result := func(welcomed bool, err error) attachResult {	471		result := func(welcomed bool, err error) attachResult {
385		return attachResult{resumeToken: resumeToken, sessionName: name, welcomed: welcomed, err: err}	472		return attachResult{resumeToken: resumeToken, sessionName: name, welcomed: welcomed, err: err}

@@ -501,6 +588,10 @@ func attachOnce(ctx context.Context, transport ports.Transport, term ports.Termi
501		if ferr := term.Flush(); ferr != nil {	588		if ferr := term.Flush(); ferr != nil {
502		return welcomedResult(fmt.Errorf("vev: flushing terminal: %w", ferr))	589		return welcomedResult(fmt.Errorf("vev: flushing terminal: %w", ferr))
503		}	590		}
			591	+	// The terminal boundary is after a successful flush and before ACK.
			592	+	if observer != nil {
			593	+	observer.ObserveRuntime(ports.NewRuntimeMark("client", ports.RuntimeTerminalFlushed, uint64(len(o.Data)), true))
			594	+	}
504		if o.NewStateNum != 0 {	595		if o.NewStateNum != 0 {
505		ackQueue.offer(o.NewStateNum)	596		ackQueue.offer(o.NewStateNum)
506		}	597		}

M internal/usecase/client/client_test.go +114 -0


@@ -155,6 +155,120 @@ func isType(typ ports.MsgType) any {
155	return mock.MatchedBy(func(f ports.Frame) bool { return f.Type == typ })	155		return mock.MatchedBy(func(f ports.Frame) bool { return f.Type == typ })
156	}	156		}
157		157
		158	+	const testPreWelcomeTimeout = 15 * time.Second
		159	+
		160	+	func newHandshakeClock(t testing.T, timerCount int) (portsmocks.MockClock, <-chan chan time.Time) {
		161	+	t.Helper()
		162	+	created := make(chan chan time.Time, timerCount)
		163	+	clk := portsmocks.NewMockClock(t)
		164	+	clk.EXPECT().NewTimer(testPreWelcomeTimeout).RunAndReturn(func(time.Duration) ports.Timer {
		165	+	timerC := make(chan time.Time, 1)
		166	+	timer := portsmocks.NewMockTimer(t)
		167	+	timer.EXPECT().C().Return((<-chan time.Time)(timerC)).Maybe()
		168	+	timer.EXPECT().Stop().Return(true).Once()
		169	+	created <- timerC
		170	+	return timer
		171	+	}).Times(timerCount)
		172	+	return clk, created
		173	+	}
		174	+
		175	+	func TestRunBoundsPreWelcomeOperations(t *testing.T) {
		176	+	tests := []struct {
		177	+	name string
		178	+	phase string
		179	+	end string
		180	+	want error
		181	+	}{
		182	+	{name: "cancel while sending hello", phase: "send", end: "cancel", want: context.Canceled},
		183	+	{name: "timeout while sending hello", phase: "send", end: "timeout", want: context.DeadlineExceeded},
		184	+	{name: "cancel while receiving welcome", phase: "recv", end: "cancel", want: context.Canceled},
		185	+	{name: "timeout while receiving welcome", phase: "recv", end: "timeout", want: context.DeadlineExceeded},
		186	+	}
		187	+
		188	+	for _, tt := range tests {
		189	+	t.Run(tt.name, func(t *testing.T) {
		190	+	ctx, cancel := context.WithCancel(context.Background())
		191	+	defer cancel()
		192	+	timerCount := 1
		193	+	if tt.phase == "recv" {
		194	+	timerCount = 2
		195	+	}
		196	+	clk, createdTimers := newHandshakeClock(t, timerCount)
		197	+
		198	+	term := portsmocks.NewMockTerminal(t)
		199	+	term.EXPECT().Size().Return(domain.Size{Cols: 80, Rows: 24}, nil).Once()
		200	+	// No EnterRaw expectation: it must not run before Welcome.
		201	+
		202	+	started := make(chan struct{})
		203	+	closed := make(chan struct{})
		204	+	operationExited := make(chan struct{})
		205	+	tr := portsmocks.NewMockTransport(t)
		206	+	tr.EXPECT().Close().Run(func() { close(closed) }).Return(nil).Once()
		207	+
		208	+	block := func() error {
		209	+	close(started)
		210	+	<-closed
		211	+	close(operationExited)
		212	+	return io.ErrClosedPipe
		213	+	}
		214	+	if tt.phase == "send" {
		215	+	tr.EXPECT().Send(isType(ports.MsgHello)).RunAndReturn(func(ports.Frame) error {
		216	+	return block()
		217	+	}).Once()
		218	+	} else {
		219	+	tr.EXPECT().Send(isType(ports.MsgHello)).Return(nil).Once()
		220	+	tr.EXPECT().Recv().RunAndReturn(func() (ports.Frame, error) {
		221	+	return ports.Frame{}, block()
		222	+	}).Once()
		223	+	}
		224	+	dialer := portsmocks.NewMockDialer(t)
		225	+	dialer.EXPECT().Dial(mock.Anything).Return(tr, nil).Once()
		226	+
		227	+	errCh := make(chan error, 1)
		228	+	go func() {
		229	+	errCh <- client.Run(ctx, dialer, term, clk, ports.IntentAttach, "main", false, nil, slog.New(slog.DiscardHandler))
		230	+	}()
		231	+
		232	+	var timerC chan time.Time
		233	+	select {
		234	+	case timerC = <-createdTimers:
		235	+	case <-time.After(time.Second):
		236	+	t.Fatal("pre-Welcome timer was not created")
		237	+	}
		238	+	select {
		239	+	case <-started:
		240	+	case <-time.After(time.Second):
		241	+	t.Fatal("pre-Welcome operation did not start")
		242	+	}
		243	+	if tt.phase == "recv" {
		244	+	select {
		245	+	case timerC = <-createdTimers:
		246	+	case <-time.After(time.Second):
		247	+	t.Fatal("Welcome timer was not created")
		248	+	}
		249	+	}
		250	+
		251	+	if tt.end == "cancel" {
		252	+	cancel()
		253	+	} else {
		254	+	timerC <- time.Time{}
		255	+	}
		256	+
		257	+	select {
		258	+	case err := <-errCh:
		259	+	require.ErrorIs(t, err, tt.want)
		260	+	case <-time.After(time.Second):
		261	+	t.Fatal("Run did not return after pre-Welcome cancellation")
		262	+	}
		263	+	select {
		264	+	case <-operationExited:
		265	+	case <-time.After(time.Second):
		266	+	t.Fatal("pre-Welcome operation leaked after Run returned")
		267	+	}
		268	+	})
		269	+	}
		270	+	}
		271	+
158	func TestAttachHelloIncludesTrueColor(t *testing.T) {	272		func TestAttachHelloIncludesTrueColor(t *testing.T) {
159	t.Setenv("TERM", "xterm-256color")	273		t.Setenv("TERM", "xterm-256color")
160	t.Setenv("COLORTERM", "truecolor")	274		t.Setenv("COLORTERM", "truecolor")

M internal/usecase/client/reconnect_toast_test.go +18 -6

 	return hello
 }
+// newReconnectHandshakeClock supplies the non-firing timer needed by the
+// bounded pre-Welcome handshake; reconnect sleeps are stubbed by these tests.
+func newReconnectHandshakeClock(t *testing.T) *portsmocks.MockClock {
+	t.Helper()
+	clk := portsmocks.NewMockClock(t)
+	timer := portsmocks.NewMockTimer(t)
+	timer.EXPECT().C().Return((<-chan time.Time)(make(chan time.Time))).Maybe()
+	timer.EXPECT().Stop().Return(true).Maybe()
+	clk.EXPECT().NewTimer(preWelcomeTimeout).Return(timer).Maybe()
+	return clk
+}
 func TestReconnectToastDegradedClearsModalAndProbingIsVisible(t *testing.T) {
 	term := newReconnectToastTerminalHarness(t)
 	defer term.closeInput()
 	resultCh := make(chan attachResult, 1)
 	ms := milestones{}
 	go func() {
-		resultCh <- attachOnce(context.Background(), tr, term.term, portsmocks.NewMockClock(t), ports.IntentAttach, "main", 0, [16]byte{1}, &ms, func() error {
+		resultCh <- attachOnce(context.Background(), tr, term.term, newReconnectHandshakeClock(t), ports.IntentAttach, "main", 0, [16]byte{1}, &ms, func() error {
 			_, err := term.term.EnterRaw()
 			return err
 		}, func() { clears <- struct{}{} }, func(stage reconnectStage) {
 	resultCh := make(chan attachResult, 1)
 	ms := milestones{}
 	go func() {
-		resultCh <- attachOnce(context.Background(), tr, term.term, portsmocks.NewMockClock(t), ports.IntentAttach, "main", 0, [16]byte{1}, &ms, func() error {
+		resultCh <- attachOnce(context.Background(), tr, term.term, newReconnectHandshakeClock(t), ports.IntentAttach, "main", 0, [16]byte{1}, &ms, func() error {
 			_, err := term.term.EnterRaw()
 			return err
 		}, func() {}, func(stage reconnectStage) {
 	}}
 	dialer := &reconnectToastSequenceDialer{transports: []ports.Transport{tr}}
-	err := Run(ctx, dialer, term.term, portsmocks.NewMockClock(t), ports.IntentAttach, "main", true, nil, slog.New(slog.DiscardHandler))
+	err := Run(ctx, dialer, term.term, newReconnectHandshakeClock(t), ports.IntentAttach, "main", true, nil, slog.New(slog.DiscardHandler))
 	require.ErrorIs(t, err, context.Canceled)
 	require.True(t, out.failed)
 	assertReconnectToastClearCoversBounds(t, out.String(), reconnectToastBoundsFor(term.size, reconnectStageMessage(reconnectStageSSH)))
 	}}
 	dialer := &reconnectToastSequenceDialer{transports: []ports.Transport{tr1, tr2}}
-	err := Run(context.Background(), dialer, term.term, portsmocks.NewMockClock(t), ports.IntentAttach, "main", true, nil, slog.New(slog.DiscardHandler))
+	err := Run(context.Background(), dialer, term.term, newReconnectHandshakeClock(t), ports.IntentAttach, "main", true, nil, slog.New(slog.DiscardHandler))
 	require.NoError(t, err)
 	require.True(t, tr1.closed)
 	require.True(t, tr2.closed)
 	}}
 	dialer := &reconnectToastSequenceDialer{transports: []ports.Transport{tr1, tr2}}
-	err := Run(context.Background(), dialer, term.term, portsmocks.NewMockClock(t), ports.IntentEphemeral, "", true, nil, slog.New(slog.DiscardHandler))
+	err := Run(context.Background(), dialer, term.term, newReconnectHandshakeClock(t), ports.IntentEphemeral, "", true, nil, slog.New(slog.DiscardHandler))
 	require.NoError(t, err)
 	require.Equal(t, 2, dialer.calls)
 			dialer := portsmocks.NewMockDialer(t)
 			tt.configure(t, dialer)
-			err := Run(ctx, dialer, term.term, portsmocks.NewMockClock(t), ports.IntentAttach, "main", true, nil, slog.New(slog.DiscardHandler))
+			err := Run(ctx, dialer, term.term, newReconnectHandshakeClock(t), ports.IntentAttach, "main", true, nil, slog.New(slog.DiscardHandler))
 			tt.wantErr(t, err)
 			out := term.out.String()
 			require.Contains(t, out, reconnectToastMessage)

A internal/usecase/client/transport_observability_test.go +154 -0

@@ -0,0 +1,154 @@
	1	+	package client_test
	2	+
	3	+	import (
	4	+	"bytes"
	5	+	"context"
	6	+	"reflect"
	7	+	"sync"
	8	+	"sync/atomic"
	9	+	"testing"
	10	+	"time"
	11	+
	12	+	"github.com/stretchr/testify/require"
	13	+
	14	+	"github.com/bnema/vev/internal/domain"
	15	+	"github.com/bnema/vev/internal/ports"
	16	+	portsmocks "github.com/bnema/vev/internal/ports/mocks"
	17	+	"github.com/bnema/vev/internal/usecase/client"
	18	+	)
	19	+
	20	+	func TestRuntimeObserverOptionRequiresSerializedContract(t *testing.T) {
	21	+	got := reflect.TypeOf(client.WithRuntimeObserver).In(0)
	22	+	want := reflect.TypeFor[ports.SerializedRuntimeObserver]()
	23	+	if got != want {
	24	+	t.Fatalf("WithRuntimeObserver accepts %v, want %v; raw blocking sinks must not enter the hot path", got, want)
	25	+	}
	26	+	}
	27	+
	28	+	func TestBlockingRuntimeObserverDoesNotDelayTerminalFlushOrACK(t *testing.T) {
	29	+	var out bytes.Buffer
	30	+	var restores atomic.Int32
	31	+	resizeCh := make(chan domain.Size)
	32	+	flushed := make(chan struct{})
	33	+	term := portsmocks.NewMockTerminal(t)
	34	+	term.EXPECT().Size().Return(domain.Size{Cols: 80, Rows: 24}, nil).Once()
	35	+	term.EXPECT().EnterRaw().Return(func() error { restores.Add(1); return nil }, nil).Once()
	36	+	in := newBlockingReader()
	37	+	defer in.unblock()
	38	+	term.EXPECT().In().Return(in).Maybe()
	39	+	term.EXPECT().Out().Return(&out).Maybe()
	40	+	term.EXPECT().Flush().Run(func() { close(flushed) }).Return(nil).Once()
	41	+	term.EXPECT().ResizeEvents().Return(resizeCh).Maybe()
	42	+
	43	+	observer := &blockingClientRuntimeObserver{entered: make(chan struct{}), release: make(chan struct{})}
	44	+	reporter := ports.NewSerializedRuntimeObserver(observer, 1)
	45	+	defer reporter.Close()
	46	+	acked := make(chan struct{})
	47	+	transport := portsmocks.NewMockTransport(t)
	48	+	transport.EXPECT().Send(isType(ports.MsgHello)).Return(nil).Once()
	49	+	transport.EXPECT().Send(isType(ports.MsgAck)).Run(func(ports.Frame) { close(acked) }).Return(nil).Once()
	50	+	unblock := scriptRecv(transport,
	51	+	recvItem{f: frameOf(ports.MsgWelcome, ports.MarshalWelcome(ports.Welcome{SessionID: "s1"}))},
	52	+	recvItem{f: frameOf(ports.MsgOutput, ports.MarshalOutput(ports.Output{Data: []byte("flush-before-observe"), NewStateNum: 3}))},
	53	+	)
	54	+	defer unblock()
	55	+	transport.EXPECT().Close().Return(nil).Once()
	56	+
	57	+	ctx, cancel := context.WithCancel(context.Background())
	58	+	result := make(chan error, 1)
	59	+	go func() {
	60	+	result <- client.Attach(ctx, transport, term, realClock{}, ports.IntentEphemeral, "", client.WithRuntimeObserver(reporter))
	61	+	}()
	62	+	awaitClientRuntime(t, observer.entered, "blocking observer")
	63	+	awaitClientRuntime(t, flushed, "terminal flush")
	64	+	awaitClientRuntime(t, acked, "ACK")
	65	+	cancel()
	66	+	close(observer.release)
	67	+	select {
	68	+	case err := <-result:
	69	+	require.NoError(t, err)
	70	+	case <-time.After(time.Second):
	71	+	t.Fatal("Attach did not complete after observer release")
	72	+	}
	73	+	require.Equal(t, "flush-before-observe", out.String())
	74	+	require.Equal(t, int32(1), restores.Load())
	75	+	}
	76	+
	77	+	func TestTerminalFlushBoundaryTransportObservability(t *testing.T) {
	78	+	var out bytes.Buffer
	79	+	var restores atomic.Int32
	80	+	resizeCh := make(chan domain.Size)
	81	+	term, in := newHappyTerminal(t, &out, &restores, resizeCh)
	82	+	defer in.unblock()
	83	+
	84	+	observer := &clientRuntimeObserver{}
	85	+	reporter := ports.NewSerializedRuntimeObserver(observer, 64)
	86	+	defer reporter.Close()
	87	+	transport := portsmocks.NewMockTransport(t)
	88	+	transport.EXPECT().Send(isType(ports.MsgHello)).Return(nil).Once()
	89	+	unblock := scriptRecv(transport,
	90	+	recvItem{f: frameOf(ports.MsgWelcome, ports.MarshalWelcome(ports.Welcome{SessionID: "s1"}))},
	91	+	recvItem{f: frameOf(ports.MsgOutput, ports.MarshalOutput(ports.Output{Data: []byte("unchanged-by-observer"), NewStateNum: 3}))},
	92	+	recvItem{f: frameOf(ports.MsgDetached, ports.MarshalDetached(ports.Detached{Reason: ports.ReasonDetach}))},
	93	+	)
	94	+	defer unblock()
	95	+	transport.EXPECT().Send(isType(ports.MsgAck)).Return(nil).Maybe()
	96	+	transport.EXPECT().Close().Return(nil).Once()
	97	+
	98	+	err := client.Attach(context.Background(), transport, term, realClock{}, ports.IntentEphemeral, "", client.WithRuntimeObserver(reporter))
	99	+	require.NoError(t, err)
	100	+	reporter.Flush()
	101	+	require.Equal(t, "unchanged-by-observer", out.String(), "observer must not alter terminal bytes")
	102	+	require.Equal(t, int32(1), restores.Load())
	103	+	// Carriage spans belong only to concrete adapters. The client owns the
	104	+	// post-successful-flush terminal boundary, before its ACK is queued.
	105	+	observer.requireOrdered(t, ports.RuntimeTerminalFlushed)
	106	+	}
	107	+
	108	+	type blockingClientRuntimeObserver struct {
	109	+	entered chan struct{}
	110	+	release chan struct{}
	111	+	once sync.Once
	112	+	}
	113	+
	114	+	func (o *blockingClientRuntimeObserver) ObserveRuntime(ports.RuntimeMark) {
	115	+	o.once.Do(func() {
	116	+	close(o.entered)
	117	+	<-o.release
	118	+	})
	119	+	}
	120	+
	121	+	func awaitClientRuntime(t *testing.T, ch <-chan struct{}, what string) {
	122	+	t.Helper()
	123	+	select {
	124	+	case <-ch:
	125	+	case <-time.After(time.Second):
	126	+	t.Fatalf("timed out waiting for %s", what)
	127	+	}
	128	+	}
	129	+
	130	+	type clientRuntimeObserver struct {
	131	+	mu sync.Mutex
	132	+	marks []ports.RuntimeMark
	133	+	}
	134	+
	135	+	func (o *clientRuntimeObserver) ObserveRuntime(mark ports.RuntimeMark) {
	136	+	o.mu.Lock()
	137	+	defer o.mu.Unlock()
	138	+	o.marks = append(o.marks, mark)
	139	+	}
	140	+
	141	+	func (o clientRuntimeObserver) requireOrdered(t testing.T, kinds ...ports.RuntimeMarkKind) {
	142	+	t.Helper()
	143	+	o.mu.Lock()
	144	+	defer o.mu.Unlock()
	145	+	next := 0
	146	+	for _, mark := range o.marks {
	147	+	if next < len(kinds) && mark.Kind == kinds[next] {
	148	+	next++
	149	+	}
	150	+	}
	151	+	if next != len(kinds) {
	152	+	t.Fatalf("runtime marks %#v do not contain ordered %#v", o.marks, kinds)
	153	+	}
	154	+	}

M internal/usecase/copy/mode.go +44 -18

 	"unicode"
 	"github.com/bnema/vev/pkg/renderer"
+	"github.com/bnema/vev/pkg/vt"
 )
 // OSC52MaxPayloadBytes caps clipboard payloads while vev intentionally emits
 const OSC52MaxPayloadBytes = 75_000
 // Snapshot is the immutable scrollback-mode document: history followed by the
-// visible screen. History rows are immutable by Scrollback's append contract;
+// visible screen. VT HistoryView retains immutable chunks; the visible frame is
-// the visible frame is cloned once when the snapshot is constructed.
+// cloned once when the snapshot is constructed.
 type Snapshot struct {
-	history HistoryView
+	history vt.HistoryView
 	screen  renderer.Frame
 	Width   int
 	Height  int
 }
-// NewSnapshot freezes the current scrollback view and clones the visible screen.
+// NewSnapshot seals the current VT history tail, then freezes that immutable
-func NewSnapshot(sb *Scrollback, screen renderer.Frame) Snapshot {
+// view and clones the visible screen. Sealing prevents repeated copy entry from
-	var history HistoryView
+// cloning every row in a partial tail.
-	if sb != nil {
+func NewSnapshot(historySource *vt.History, screen renderer.Frame) Snapshot {
-		history = sb.View()
+	var history vt.HistoryView
+	if historySource != nil {
+		history = historySource.SealAndView()
 	}
 	return Snapshot{history: history, screen: screen.Clone(), Width: screen.Width, Height: screen.Height}
 }
 // NewSnapshotFromRows constructs a snapshot that owns explicit caller rows.
 // It is intended for callers that already have a complete document rather than
-// a scrollback and visible frame.
+// a scrollback and visible frame. Height is viewport geometry, not a request to
+// append blank document rows.
 func NewSnapshotFromRows(rows [][]renderer.Cell, width, height int) Snapshot {
-	owned := make([][]renderer.Cell, len(rows))
+	history := vt.NewHistory(vt.HistoryConfig{MaxRows: len(rows), ChunkRows: 256})
-	for i, row := range rows {
+	for _, row := range rows {
-		owned[i] = append([]renderer.Cell(nil), row...)
+		history.Append(row)
 	}
-	return Snapshot{history: HistoryView{rows: owned}, Width: width, Height: height}
+	return Snapshot{history: history.SealAndView(), Width: width, Height: height}
 }
 // Len returns the number of document rows.
 func (s Snapshot) Len() int { return s.history.Len() + s.screen.Height }
-// Row returns document row i, or nil when i is out of range.
+// Row returns document row i, or nil when i is out of range. History rows
+// borrow immutable history storage and must be treated as read-only.
 func (s Snapshot) Row(i int) []renderer.Cell {
 	if i < 0 {
 		return nil
 	}
 	if i < s.history.Len() {
-		return s.history.Row(i)
+		return s.history.BorrowedRow(i)
 	}
 	i -= s.history.Len()
 	if i >= s.screen.Height {
 	return s.screen.Row(i)
 }
+func (s Snapshot) rangeRows(yield func(int, []renderer.Cell) bool) {
+	rowIndex := 0
+	stopped := false
+	s.history.Range(func(row []renderer.Cell) bool {
+		if !yield(rowIndex, row) {
+			stopped = true
+			return false
+		}
+		rowIndex++
+		return true
+	})
+	if stopped {
+		return
+	}
+	for y := range s.screen.Height {
+		if !yield(rowIndex+y, s.screen.Row(y)) {
+			return
+		}
+	}
+}
 // Mode stores per-client scrollback viewport and line-selection state.
 type SearchMatch struct {
 	Row   int
 	}
 	needle := lowerRunes(query)
 	matches := make([]SearchMatch, 0)
-	for row := range s.Len() {
+	s.rangeRows(func(row int, cells []renderer.Cell) bool {
-		cells := s.Row(row)
 		haystack, cellIndexes := searchableCells(cells)
 		var text string
 		for start := 0; start+len(needle) <= len(haystack); {
 			matches = append(matches, SearchMatch{Row: row, Start: cellIndexes[start], End: cellEnd, Text: text})
 			start += len(needle)
 		}
-	}
+		return true
+	})
 	return matches
 }

M internal/usecase/copy/mode_test.go +128 -3


@@ -7,11 +7,30 @@ import (
7		"testing"	7		"testing"
8			8
9		"github.com/bnema/vev/pkg/renderer"	9		"github.com/bnema/vev/pkg/renderer"
			10	+	"github.com/bnema/vev/pkg/vt"
10		"github.com/stretchr/testify/require"	11		"github.com/stretchr/testify/require"
11		)	12		)
12			13
13		var benchmarkSnapshotSink Snapshot	14		var benchmarkSnapshotSink Snapshot
14			15
			16	+	func row(text string) []renderer.Cell {
			17	+	result := make([]renderer.Cell, 0, len([]rune(text)))
			18	+	for _, r := range text {
			19	+	result = append(result, renderer.Cell{Rune: r, Style: renderer.DefaultStyle()})
			20	+	}
			21	+	return result
			22	+	}
			23	+
			24	+	func newHistory(rows int) *vt.History { return vt.NewHistory(vt.HistoryConfig{MaxRows: rows}) }
			25	+
			26	+	func rowText(cells []renderer.Cell) string {
			27	+	runes := make([]rune, len(cells))
			28	+	for i, cell := range cells {
			29	+	runes[i] = cell.Rune
			30	+	}
			31	+	return string(runes)
			32	+	}
			33	+
15		func snapshot(lines []string, height int) Snapshot {	34		func snapshot(lines []string, height int) Snapshot {
16		rows := make([][]renderer.Cell, len(lines))	35		rows := make([][]renderer.Cell, len(lines))
17		for i, line := range lines {	36		for i, line := range lines {

@@ -355,6 +374,24 @@ func TestFindMatchesKeepsRuneAndCellIndexesAlignedWhenLowercasing(t *testing.T)
355		require.Equal(t, 2, matches[0].End)	374		require.Equal(t, 2, matches[0].End)
356		}	375		}
357			376
			377	+	func TestFindMatchesDoesNotCopyEveryHistoryRow(t *testing.T) {
			378	+	const rows = 100
			379	+	history := vt.NewHistory(vt.HistoryConfig{MaxRows: rows, ChunkRows: 10})
			380	+	for range rows {
			381	+	history.Append(row(strings.Repeat("x", 120)))
			382	+	}
			383	+	s := NewSnapshot(history, renderer.NewFrame(120, 1))
			384	+
			385	+	allocs := testing.AllocsPerRun(10, func() {
			386	+	if matches := FindMatches(s, "missing"); len(matches) != 0 {
			387	+	t.Fatalf("unexpected matches: %d", len(matches))
			388	+	}
			389	+	})
			390	+	if allocs >= 250 {
			391	+	t.Fatalf("FindMatches allocations = %.0f, want < 250 (no row copy per history row)", allocs)
			392	+	}
			393	+	}
			394	+
358		func TestCopyModeSearchUsesDisplayCellOffsetsForWideRows(t *testing.T) {	395		func TestCopyModeSearchUsesDisplayCellOffsetsForWideRows(t *testing.T) {
359		row := []renderer.Cell{	396		row := []renderer.Cell{
360		{Rune: '界'},	397		{Rune: '界'},

@@ -478,6 +515,94 @@ func TestOSC52FromBase64(t *testing.T) {
478		}	515		}
479		}	516		}
480			517
			518	+	func TestNewSnapshotFreezesHistoryAndVisibleScreen(t *testing.T) {
			519	+	history := newHistory(1)
			520	+	history.Append(row("old"))
			521	+	screen := renderer.NewFrame(3, 1)
			522	+	copy(screen.Row(0), row("one"))
			523	+
			524	+	document := NewSnapshot(history, screen)
			525	+	history.Append(row("new"))
			526	+	screen.Set(0, 0, renderer.Cell{Rune: 'x'})
			527	+
			528	+	require.Equal(t, 2, document.Len())
			529	+	require.Equal(t, "old", rowText(document.Row(0)))
			530	+	require.Equal(t, "one", rowText(document.Row(1)))
			531	+	require.Nil(t, document.Row(-1))
			532	+	require.Nil(t, document.Row(2))
			533	+	}
			534	+
			535	+	func TestNewSnapshotFromRowsOwnsCallerRows(t *testing.T) {
			536	+	rows := [][]renderer.Cell{row("one")}
			537	+	document := NewSnapshotFromRows(rows, 3, 1)
			538	+	rows[0][0].Rune = 'x'
			539	+	require.Equal(t, "one", rowText(document.Row(0)))
			540	+	}
			541	+
			542	+	func TestNewSnapshotFromRowsKeepsDocumentAndViewportDimensionsIndependent(t *testing.T) {
			543	+	document := NewSnapshotFromRows([][]renderer.Cell{row("one"), row("two")}, 7, 5)
			544	+
			545	+	require.Equal(t, 2, document.Len())
			546	+	require.Equal(t, 7, document.Width)
			547	+	require.Equal(t, 5, document.Height)
			548	+	require.Equal(t, "one", rowText(document.Row(0)))
			549	+	require.Equal(t, "two", rowText(document.Row(1)))
			550	+	require.Nil(t, document.Row(2))
			551	+	}
			552	+
			553	+	// This scaling gate deliberately compares equal row counts rather than an
			554	+	// absolute budget, which varies between Go releases. A HistoryView shares
			555	+	// sealed chunks, so its allocation must not grow with history-row width.
			556	+	func TestNewSnapshotAllocationIsWidthIndependent(t *testing.T) {
			557	+	if testing.Short() {
			558	+	t.Skip("skipping benchmark-driven allocation gate in short mode")
			559	+	}
			560	+	narrow := benchmarkNewSnapshotBytes(16)
			561	+	wide := benchmarkNewSnapshotBytes(512)
			562	+	const conservativeWidthTolerance = 2
			563	+	require.LessOrEqual(t, wide, narrow*conservativeWidthTolerance,
			564	+	"NewSnapshot B/op scaled with row width: narrow=%d wide=%d", narrow, wide)
			565	+	}
			566	+
			567	+	func benchmarkNewSnapshotBytes(width int) int64 {
			568	+	const historyRows = 256
			569	+	history := newHistory(historyRows)
			570	+	for range historyRows {
			571	+	history.Append(make([]renderer.Cell, width))
			572	+	}
			573	+	// Keep visible-frame work constant so the comparison isolates history rows.
			574	+	screen := renderer.NewFrame(8, 4)
			575	+	result := testing.Benchmark(func(b *testing.B) {
			576	+	b.ReportAllocs()
			577	+	b.ResetTimer()
			578	+	for b.Loop() {
			579	+	benchmarkSnapshotSink = NewSnapshot(history, screen)
			580	+	}
			581	+	})
			582	+	return result.AllocedBytesPerOp()
			583	+	}
			584	+
			585	+	// This regression replaces the removed Scrollback ring tests: the VT history
			586	+	// remains bounded oldest-first and a captured copy document survives eviction.
			587	+	func TestNewSnapshotRetainsBoundedHistoryAcrossLaterEviction(t *testing.T) {
			588	+	history := newHistory(2)
			589	+	history.Append(row("one"))
			590	+	history.Append(row("two"))
			591	+	document := NewSnapshot(history, renderer.NewFrame(3, 1))
			592	+	history.Append(row("three"))
			593	+	history.Append(row("four"))
			594	+
			595	+	// A copy document is history followed by the visible frame. Its total is
			596	+	// therefore the bounded retained history plus the one visible row.
			597	+	require.Equal(t, 3, document.Len())
			598	+	require.Equal(t, "one", rowText(document.Row(0)))
			599	+	require.Equal(t, "two", rowText(document.Row(1)))
			600	+	require.Equal(t, " ", rowText(document.Row(2)), "the visible frame remains part of the document")
			601	+	view := history.View()
			602	+	require.Equal(t, "three", rowText(view.Row(0)))
			603	+	require.Equal(t, "four", rowText(view.Row(1)))
			604	+	}
			605	+
481		func BenchmarkNewSnapshot10KRows(b *testing.B) {	606		func BenchmarkNewSnapshot10KRows(b *testing.B) {
482		const (	607		const (
483		width = 120	608		width = 120

@@ -485,9 +610,9 @@ func BenchmarkNewSnapshot10KRows(b *testing.B) {
485		historyRows = 10_000	610		historyRows = 10_000
486		)	611		)
487			612
488	-	scrollback := NewScrollback(historyRows)	613	+	history := newHistory(historyRows)
489		for range historyRows {	614		for range historyRows {
490	-	scrollback.Append(row(strings.Repeat("x", width)))	615	+	history.Append(row(strings.Repeat("x", width)))
491		}	616		}
492		screen := renderer.NewFrame(width, height)	617		screen := renderer.NewFrame(width, height)
493		for y := range height {	618		for y := range height {

@@ -497,7 +622,7 @@ func BenchmarkNewSnapshot10KRows(b *testing.B) {
497		b.ReportAllocs()	622		b.ReportAllocs()
498		b.ResetTimer()	623		b.ResetTimer()
499		for b.Loop() {	624		for b.Loop() {
500	-	benchmarkSnapshotSink = NewSnapshot(scrollback, screen)	625	+	benchmarkSnapshotSink = NewSnapshot(history, screen)
501		}	626		}
502		}	627		}
503			628

D internal/usecase/copy/scrollback.go +0 -83

@@ -1,83 +0,0 @@
1	-	package copy
2	-
3	-	import "github.com/bnema/vev/pkg/renderer"
4	-
5	-	// Scrollback stores evicted terminal rows in a fixed-size ring buffer.
6	-	type Scrollback struct {
7	-	rows [][]renderer.Cell
8	-	head int
9	-	len int
10	-	}
11	-
12	-	// NewScrollback constructs a scrollback ring capped at capacity rows.
13	-	func NewScrollback(capacity int) *Scrollback {
14	-	if capacity < 0 {
15	-	capacity = 0
16	-	}
17	-	return &Scrollback{rows: make([][]renderer.Cell, capacity)}
18	-	}
19	-
20	-	// Cap returns the maximum number of rows retained.
21	-	func (s *Scrollback) Cap() int { return len(s.rows) }
22	-
23	-	// Len returns the number of retained rows.
24	-	func (s *Scrollback) Len() int { return s.len }
25	-
26	-	// Head returns the physical ring slot holding the oldest retained row.
27	-	func (s *Scrollback) Head() int { return s.head }
28	-
29	-	// Append stores a copy of row, evicting the oldest row when the ring is full.
30	-	func (s *Scrollback) Append(row []renderer.Cell) {
31	-	if len(s.rows) == 0 {
32	-	return
33	-	}
34	-	cp := append([]renderer.Cell(nil), row...)
35	-	if s.len < len(s.rows) {
36	-	s.rows[(s.head+s.len)%len(s.rows)] = cp
37	-	s.len++
38	-	return
39	-	}
40	-	s.rows[s.head] = cp
41	-	s.head = (s.head + 1) % len(s.rows)
42	-	}
43	-
44	-	// Snapshot returns retained rows oldest-first.
45	-	//
46	-	// The returned outer slice is a copy, but row slices are shared with the
47	-	// scrollback storage. Callers must treat the returned rows as read-only.
48	-	func (s *Scrollback) Snapshot() [][]renderer.Cell {
49	-	if s.len == 0 \|\| len(s.rows) == 0 {
50	-	return nil
51	-	}
52	-	rows := make([][]renderer.Cell, s.len)
53	-	for i := range s.len {
54	-	rows[i] = s.rows[(s.head+i)%len(s.rows)]
55	-	}
56	-	return rows
57	-	}
58	-
59	-	// HistoryView is a frozen oldest-first view of scrollback rows.
60	-	//
61	-	// Its rows share cell slices with the scrollback at the time View is called.
62	-	// Callers must treat rows returned by Row as read-only.
63	-	type HistoryView struct {
64	-	rows [][]renderer.Cell
65	-	}
66	-
67	-	// View returns a frozen scrollback view. Appending to the scrollback cannot
68	-	// alter the view because Append replaces ring slots rather than mutating rows.
69	-	func (s *Scrollback) View() HistoryView {
70	-	return HistoryView{rows: s.Snapshot()}
71	-	}
72	-
73	-	// Len returns the number of rows in the view.
74	-	func (v HistoryView) Len() int { return len(v.rows) }
75	-
76	-	// Row returns the row at logical index i, oldest first, or nil when i is out
77	-	// of range. Returned rows are shared with the view and must be read-only.
78	-	func (v HistoryView) Row(i int) []renderer.Cell {
79	-	if i < 0 \|\| i >= len(v.rows) {
80	-	return nil
81	-	}
82	-	return v.rows[i]
83	-	}

D internal/usecase/copy/scrollback_test.go +0 -209

@@ -1,209 +0,0 @@
1	-	package copy
2	-
3	-	import (
4	-	"testing"
5	-
6	-	"github.com/bnema/vev/pkg/renderer"
7	-	)
8	-
9	-	func row(text string) []renderer.Cell {
10	-	runes := []rune(text)
11	-	cells := make([]renderer.Cell, len(runes))
12	-	for i, r := range runes {
13	-	cells[i] = renderer.Cell{Rune: r, Style: renderer.DefaultStyle()}
14	-	}
15	-	return cells
16	-	}
17	-
18	-	func rowText(cells []renderer.Cell) string {
19	-	runes := make([]rune, len(cells))
20	-	for i, c := range cells {
21	-	runes[i] = c.Rune
22	-	}
23	-	return string(runes)
24	-	}
25	-
26	-	func TestScrollbackRing(t *testing.T) {
27	-	tests := []struct {
28	-	name string
29	-	cap int
30	-	append []string
31	-	wantHead int
32	-	wantLen int
33	-	wantRows []string
34	-	}{
35	-	{
36	-	name: "empty ring drops rows",
37	-	cap: 0,
38	-	append: []string{"aa", "bb"},
39	-	wantHead: 0,
40	-	wantLen: 0,
41	-	},
42	-	{
43	-	name: "below capacity keeps head at zero",
44	-	cap: 3,
45	-	append: []string{"aa", "bb"},
46	-	wantHead: 0,
47	-	wantLen: 2,
48	-	wantRows: []string{"aa", "bb"},
49	-	},
50	-	{
51	-	name: "at capacity fills without rotating head",
52	-	cap: 3,
53	-	append: []string{"aa", "bb", "cc"},
54	-	wantHead: 0,
55	-	wantLen: 3,
56	-	wantRows: []string{"aa", "bb", "cc"},
57	-	},
58	-	{
59	-	name: "over capacity evicts oldest and rotates head",
60	-	cap: 3,
61	-	append: []string{"aa", "bb", "cc", "dd", "ee"},
62	-	wantHead: 2,
63	-	wantLen: 3,
64	-	wantRows: []string{"cc", "dd", "ee"},
65	-	},
66	-	}
67	-
68	-	for _, tt := range tests {
69	-	t.Run(tt.name, func(t *testing.T) {
70	-	sb := NewScrollback(tt.cap)
71	-	for _, r := range tt.append {
72	-	sb.Append(row(r))
73	-	}
74	-
75	-	if got := sb.Cap(); got != tt.cap {
76	-	t.Fatalf("Cap() = %d, want %d", got, tt.cap)
77	-	}
78	-	if got := sb.Head(); got != tt.wantHead {
79	-	t.Fatalf("Head() = %d, want %d", got, tt.wantHead)
80	-	}
81	-	if got := sb.Len(); got != tt.wantLen {
82	-	t.Fatalf("Len() = %d, want %d", got, tt.wantLen)
83	-	}
84	-	for i, want := range tt.wantRows {
85	-	if got := rowText(sb.View().Row(i)); got != want {
86	-	t.Fatalf("Row(%d) = %q, want %q", i, got, want)
87	-	}
88	-	}
89	-	})
90	-	}
91	-	}
92	-
93	-	func TestScrollbackCopiesRows(t *testing.T) {
94	-	tests := []struct {
95	-	name string
96	-	}{
97	-	{name: "append copies caller row"},
98	-	}
99	-	for _, tt := range tests {
100	-	t.Run(tt.name, func(t *testing.T) {
101	-	sb := NewScrollback(1)
102	-	r := row("ab")
103	-	sb.Append(r)
104	-	r[0].Rune = 'z'
105	-
106	-	if text := rowText(sb.View().Row(0)); text != "ab" {
107	-	t.Fatalf("stored row = %q, want copy unaffected by caller mutation", text)
108	-	}
109	-	})
110	-	}
111	-	}
112	-
113	-	func TestScrollbackSnapshot(t *testing.T) {
114	-	tests := []struct {
115	-	name string
116	-	cap int
117	-	append []string
118	-	wantRows []string
119	-	}{
120	-	{
121	-	name: "empty capacity snapshots no rows",
122	-	cap: 0,
123	-	append: []string{"aa", "bb"},
124	-	wantRows: nil,
125	-	},
126	-	{
127	-	name: "oldest first after ring wrap",
128	-	cap: 3,
129	-	append: []string{"aa", "bb", "cc", "dd"},
130	-	wantRows: []string{"bb", "cc", "dd"},
131	-	},
132	-	}
133	-
134	-	for _, tt := range tests {
135	-	t.Run(tt.name, func(t *testing.T) {
136	-	sb := NewScrollback(tt.cap)
137	-	for _, r := range tt.append {
138	-	sb.Append(row(r))
139	-	}
140	-
141	-	got := sb.Snapshot()
142	-	if len(got) != len(tt.wantRows) {
143	-	t.Fatalf("Snapshot() len = %d, want %d", len(got), len(tt.wantRows))
144	-	}
145	-	for i, want := range tt.wantRows {
146	-	if text := rowText(got[i]); text != want {
147	-	t.Fatalf("Snapshot()[%d] = %q, want %q", i, text, want)
148	-	}
149	-	}
150	-	if len(got) > 0 {
151	-	got[0] = row("zz")
152	-	if text := rowText(sb.View().Row(0)); text != tt.wantRows[0] {
153	-	t.Fatalf("Snapshot() outer slice mutation changed storage: Row(0) = %q, want %q", text, tt.wantRows[0])
154	-	}
155	-	}
156	-	})
157	-	}
158	-	}
159	-
160	-	func TestScrollbackAppendDoesNotMutateSnapshottedRows(t *testing.T) {
161	-	sb := NewScrollback(2)
162	-	sb.Append(row("aa"))
163	-	sb.Append(row("bb"))
164	-	snapshot := sb.Snapshot()
165	-
166	-	sb.Append(row("cc"))
167	-	sb.Append(row("dd"))
168	-
169	-	if text := rowText(snapshot[0]); text != "aa" {
170	-	t.Fatalf("snapshot[0] after append = %q, want original row", text)
171	-	}
172	-	if text := rowText(snapshot[1]); text != "bb" {
173	-	t.Fatalf("snapshot[1] after append = %q, want original row", text)
174	-	}
175	-	}
176	-
177	-	func TestScrollbackViewSurvivesCompleteRingOverwrite(t *testing.T) {
178	-	sb := NewScrollback(2)
179	-	sb.Append(row("aa"))
180	-	sb.Append(row("bb"))
181	-	view := sb.View()
182	-	if &view.Row(0)[0] != &sb.rows[0][0] {
183	-	t.Fatal("View() deep-copied cells, want shared row storage")
184	-	}
185	-
186	-	sb.Append(row("cc"))
187	-	sb.Append(row("dd"))
188	-
189	-	if got := view.Len(); got != 2 {
190	-	t.Fatalf("View().Len() = %d, want 2", got)
191	-	}
192	-	for i, want := range []string{"aa", "bb"} {
193	-	if got := rowText(view.Row(i)); got != want {
194	-	t.Fatalf("View().Row(%d) = %q, want %q", i, got, want)
195	-	}
196	-	}
197	-	}
198	-
199	-	func TestHistoryViewRowBounds(t *testing.T) {
200	-	sb := NewScrollback(1)
201	-	sb.Append(row("aa"))
202	-	view := sb.View()
203	-
204	-	for _, i := range []int{-1, 1} {
205	-	if got := view.Row(i); got != nil {
206	-	t.Errorf("View().Row(%d) = %v, want nil", i, got)
207	-	}
208	-	}
209	-	}

D internal/usecase/copy/snapshot_test.go +0 -85

@@ -1,85 +0,0 @@
1	-	package copy
2	-
3	-	import (
4	-	"testing"
5	-
6	-	"github.com/bnema/vev/pkg/renderer"
7	-	)
8	-
9	-	func TestNewSnapshotFreezesHistoryAndVisibleScreen(t *testing.T) {
10	-	sb := NewScrollback(1)
11	-	sb.Append(row("old"))
12	-	screen := renderer.NewFrame(3, 1)
13	-	copy(screen.Row(0), row("one"))
14	-
15	-	snapshot := NewSnapshot(sb, screen)
16	-	if &snapshot.Row(0)[0] != &sb.View().Row(0)[0] {
17	-	t.Fatal("NewSnapshot() deep-copied immutable history")
18	-	}
19	-	sb.Append(row("new"))
20	-	screen.Set(0, 0, renderer.Cell{Rune: 'x'})
21	-
22	-	if got := snapshot.Len(); got != 2 {
23	-	t.Fatalf("Len() = %d, want 2", got)
24	-	}
25	-	for i, want := range []string{"old", "one"} {
26	-	if got := rowText(snapshot.Row(i)); got != want {
27	-	t.Errorf("Row(%d) = %q, want %q", i, got, want)
28	-	}
29	-	}
30	-	for _, i := range []int{-1, 2} {
31	-	if got := snapshot.Row(i); got != nil {
32	-	t.Errorf("Row(%d) = %v, want nil", i, got)
33	-	}
34	-	}
35	-	}
36	-
37	-	func TestNewSnapshotFromRowsOwnsCallerRows(t *testing.T) {
38	-	rows := [][]renderer.Cell{row("one")}
39	-	snapshot := NewSnapshotFromRows(rows, 3, 1)
40	-	rows[0][0].Rune = 'x'
41	-	if got := rowText(snapshot.Row(0)); got != "one" {
42	-	t.Fatalf("Row(0) = %q, want deep-owned caller row", got)
43	-	}
44	-	}
45	-
46	-	// This scaling gate deliberately compares equal row counts rather than an
47	-	// absolute budget, which varies between Go releases. A full history-row clone
48	-	// grows with row width; the immutable HistoryView allocation does not.
49	-	func TestNewSnapshotAllocationIsWidthIndependent(t *testing.T) {
50	-	if testing.Short() {
51	-	t.Skip("skipping benchmark-driven allocation gate in short mode")
52	-	}
53	-	narrow := benchmarkNewSnapshotBytes(16)
54	-	wide := benchmarkNewSnapshotBytes(512)
55	-	assertWidthIndependentAllocations(t, "NewSnapshot", narrow, wide)
56	-	}
57	-
58	-	func benchmarkNewSnapshotBytes(width int) int64 {
59	-	const historyRows = 256
60	-
61	-	scrollback := NewScrollback(historyRows)
62	-	for range historyRows {
63	-	scrollback.Append(make([]renderer.Cell, width))
64	-	}
65	-	// Keep visible-frame work constant so the comparison isolates history rows.
66	-	screen := renderer.NewFrame(8, 4)
67	-	result := testing.Benchmark(func(b *testing.B) {
68	-	b.ReportAllocs()
69	-	b.ResetTimer()
70	-	for b.Loop() {
71	-	benchmarkSnapshotSink = NewSnapshot(scrollback, screen)
72	-	}
73	-	})
74	-	return result.AllocedBytesPerOp()
75	-	}
76	-
77	-	func assertWidthIndependentAllocations(t *testing.T, operation string, narrow, wide int64) {
78	-	t.Helper()
79	-	// Two times permits allocator noise while decisively rejecting a cell-row
80	-	// copy, whose allocation grows by 32x for the widths above.
81	-	const conservativeWidthTolerance = 2
82	-	if wide > narrow*conservativeWidthTolerance {
83	-	t.Fatalf("%s B/op scaled with row width: narrow=%d wide=%d (limit=%d)", operation, narrow, wide, narrow*conservativeWidthTolerance)
84	-	}
85	-	}

M internal/usecase/daemon/attention.go +2 -2


@@ -49,7 +49,7 @@ func (d Daemon) jumpAttention(sess session, ac *attachedClient) {
49		if idx, ok := oldestAttentionTab(sess); ok {	49		if idx, ok := oldestAttentionTab(sess); ok {
50		if sess.switchTab(idx) {	50		if sess.switchTab(idx) {
51		d.activateTab(sess, sess.activeTab())	51		d.activateTab(sess, sess.activeTab())
52	-	d.paint(sess, ac, true)	52	+	d.invalidateRender(sess, ac, true, "attention.go")
53		}	53		}
54		return	54		return
55		}	55		}

@@ -119,7 +119,7 @@ func (d Daemon) repaintAttachedClients(sess session) {
119		ac := sess.client	119		ac := sess.client
120		sess.mu.Unlock()	120		sess.mu.Unlock()
121		if ac != nil {	121		if ac != nil {
122	-	d.paint(sess, ac, false)	122	+	d.invalidateRender(sess, ac, false, "attention.go")
123		}	123		}
124		}	124		}
125			125

M internal/usecase/daemon/attention_test.go +9 -9


@@ -53,7 +53,7 @@ func TestPaintDoesNotAckActiveAttentionOnBlankPulseFrame(t *testing.T) {
53		d.setAttentionFrame(0)	53		d.setAttentionFrame(0)
54		d.noteAttention(sess, sess.tabs[0])	54		d.noteAttention(sess, sess.tabs[0])
55			55
56	-	d.paint(sess, ac, true)	56	+	d.paint(sess, ac, true, nil)
57		data := mustOutputData(t, sends)	57		data := mustOutputData(t, sends)
58		require.NotContains(t, string(data), string(ui.AttentionGlyph))	58		require.NotContains(t, string(data), string(ui.AttentionGlyph))
59			59

@@ -62,7 +62,7 @@ func TestPaintDoesNotAckActiveAttentionOnBlankPulseFrame(t *testing.T) {
62		sess.mu.Unlock()	62		sess.mu.Unlock()
63			63
64		d.setAttentionFrame(1)	64		d.setAttentionFrame(1)
65	-	d.paint(sess, ac, true)	65	+	d.paint(sess, ac, true, nil)
66		data = mustOutputData(t, sends)	66		data = mustOutputData(t, sends)
67		require.Contains(t, string(data), string(ui.AttentionGlyph))	67		require.Contains(t, string(data), string(ui.AttentionGlyph))
68			68

@@ -97,7 +97,7 @@ func TestPTYReaderActiveVisibleAgentNotificationRendersBellBeforeAck(t *testing.
97		return sess.tabs[0].attention && !sess.tabs[0].attentionAt.IsZero()	97		return sess.tabs[0].attention && !sess.tabs[0].attentionAt.IsZero()
98		}, 2time.Second, 5time.Millisecond)	98		}, 2time.Second, 5time.Millisecond)
99			99
100	-	d.paint(sess, ac, true)	100	+	d.paint(sess, ac, true, nil)
101		data := mustOutputData(t, sends)	101		data := mustOutputData(t, sends)
102		require.Contains(t, string(data), string(ui.AttentionGlyph))	102		require.Contains(t, string(data), string(ui.AttentionGlyph))
103			103

@@ -242,7 +242,7 @@ func TestAckAttentionClearsOnlyPaintedVisibleTab(t *testing.T) {
242		sess.tabs[1].attentionAt = time.Unix(2, 0)	242		sess.tabs[1].attentionAt = time.Unix(2, 0)
243		sess.mu.Unlock()	243		sess.mu.Unlock()
244			244
245	-	d.paint(sess, ac, true)	245	+	d.paint(sess, ac, true, nil)
246		_ = mustOutputData(t, sends)	246		_ = mustOutputData(t, sends)
247		// The deferred repaint (from ackAttention) targets all attached clients,	247		// The deferred repaint (from ackAttention) targets all attached clients,
248		// including this one, but the bars it would draw are already exactly what	248		// including this one, but the bars it would draw are already exactly what

@@ -272,7 +272,7 @@ func TestPaintPreservesBackgroundAttention(t *testing.T) {
272		sess.mu.Unlock()	272		sess.mu.Unlock()
273		d.setAttentionFrame(1)	273		d.setAttentionFrame(1)
274			274
275	-	d.paint(sess, ac, true)	275	+	d.paint(sess, ac, true, nil)
276		data := mustOutputData(t, sends)	276		data := mustOutputData(t, sends)
277			277
278		require.Contains(t, string(data), "")	278		require.Contains(t, string(data), "")

@@ -292,7 +292,7 @@ func TestSwitchTabClearsAttentionEndToEnd(t *testing.T) {
292		sess.mu.Unlock()	292		sess.mu.Unlock()
293			293
294		require.True(t, sess.switchTab(1))	294		require.True(t, sess.switchTab(1))
295	-	d.paint(sess, ac, true)	295	+	d.paint(sess, ac, true, nil)
296		data := mustOutputData(t, sends)	296		data := mustOutputData(t, sends)
297		// As above: the deferred all-clients repaint has nothing new to say about	297		// As above: the deferred all-clients repaint has nothing new to say about
298		// this client's bars, so no second frame follows.	298		// this client's bars, so no second frame follows.

@@ -321,7 +321,7 @@ func TestCloseRingingTabClearsClientBar(t *testing.T) {
321		sess.mu.Unlock()	321		sess.mu.Unlock()
322		d.setAttentionFrame(1)	322		d.setAttentionFrame(1)
323			323
324	-	d.paint(sess, ac, true)	324	+	d.paint(sess, ac, true, nil)
325		before := mustOutputData(t, sends)	325		before := mustOutputData(t, sends)
326		require.Contains(t, string(before), "")	326		require.Contains(t, string(before), "")
327			327

@@ -346,7 +346,7 @@ func TestAnimationRepaintConfinedToBarRows(t *testing.T) {
346		sess.tabs[1].attentionAt = time.Unix(1, 0)	346		sess.tabs[1].attentionAt = time.Unix(1, 0)
347		sess.mu.Unlock()	347		sess.mu.Unlock()
348			348
349	-	d.paint(sess, ac, true)	349	+	d.paint(sess, ac, true, nil)
350		full := mustOutputData(t, sends)	350		full := mustOutputData(t, sends)
351		require.Contains(t, string(full), "MIDSCREENMARKER")	351		require.Contains(t, string(full), "MIDSCREENMARKER")
352			352

@@ -428,7 +428,7 @@ func TestCloseRingingTabRefreshesOtherSessionBottomBar(t *testing.T) {
428		acB.keys = keys.NewRouter(d.clock, daemonKeyHandler{d: d, ac: acB}, nil)	428		acB.keys = keys.NewRouter(d.clock, daemonKeyHandler{d: d, ac: acB}, nil)
429		d.sessions[sessB.id] = sessB	429		d.sessions[sessB.id] = sessB
430			430
431	-	d.paint(sessB, acB, true)	431	+	d.paint(sessB, acB, true, nil)
432		before := mustOutputData(t, sendsB)	432		before := mustOutputData(t, sendsB)
433		require.Contains(t, string(before), string(ui.AttentionGlyph))	433		require.Contains(t, string(before), string(ui.AttentionGlyph))
434			434

M internal/usecase/daemon/bar_refresh.go +7 -8


@@ -306,12 +306,11 @@ func (d Daemon) pokeSessionRender(sess session) {
306		if sess == nil {	306		if sess == nil {
307		return	307		return
308		}	308		}
309	-	if tb := sess.activeTab(); tb != nil {	309	+	sess.mu.Lock()
310	-	tb.mu.Lock()	310	+	ac := sess.client
311	-	p := tb.focusedPane()	311	+	sess.mu.Unlock()
312	-	tb.mu.Unlock()	312	+	// A headless session can still supply a picker preview to an attachment
313	-	if p != nil {	313	+	// owned by another session. Its coordinator fans this wake out to those
314	-	signal(p.dirty)	314	+	// preview subscribers even though no primary attachment is bound.
315	-	}	315	+	d.invalidateRender(sess, ac, false, "bar_refresh.go")
316	-	}
317		}	316		}

M internal/usecase/daemon/bar_refresh_test.go +12 -0


@@ -277,6 +277,18 @@ func TestBarScriptRunIgnoresStaleConfigVersion(t *testing.T) {
277	require.Empty(t, state.bottomRight)	277		require.Empty(t, state.bottomRight)
278	}	278		}
279		279
		280	+	func TestPokeSessionRenderWakesHeadlessPreviewCoordinator(t *testing.T) {
		281	+	d := newBarRefreshTestDaemon(nil, time.Second)
		282	+	sess := newBarRefreshTestSession()
		283	+	var invalidations int
		284	+	rc := newRenderCoordinator(renderCoordinatorOptions{onInvalidate: func(renderInvalidation) { invalidations++ }})
		285	+	sess.installRenderCoordinator(rc)
		286	+
		287	+	d.pokeSessionRender(sess)
		288	+
		289	+	require.Equal(t, 1, invalidations)
		290	+	}
		291	+
280	func TestBarScriptRefreshIsPerSession(t *testing.T) {	292		func TestBarScriptRefreshIsPerSession(t *testing.T) {
281	r := &fakeBarRunner{outs: []string{"top-a", "bottom-a", "top-b", "bottom-b"}}	293		r := &fakeBarRunner{outs: []string{"top-a", "bottom-a", "top-b", "bottom-b"}}
282	d := newBarRefreshTestDaemon(r, time.Second)	294		d := newBarRefreshTestDaemon(r, time.Second)

M internal/usecase/daemon/client.go +273 -177

-// Package daemon holds vev's server-side session multiplexer use case: the
+// Package daemon holds vev's server-side session multiplexer use case.
-// accept loop, the ephemeral/named session registry, the per-tab PTY reader
-// and VT screen, and the per-client debounced render scheduler.
 //
-// Concurrency model (sessions own one or more PTY-backed tabs):
+// Lock ordering: never acquire sendMu or session/daemon locks while holding a
-//
+// pane lock. Coordinator callbacks and timer methods run without coordinator.mu;
-//   - Serve runs the accept loop. Each accepted connection is handled by its
+// see the lock-specific comments on attachedClient and renderCoordinator.
-//     own goroutine (handleConn): it reads the first frame and routes it to a
-//     session create/attach, a list, or a kill.
-//   - Per session there are exactly two long-lived goroutines: the PTY reader
-//     (drains child output into the VT screen and pokes a cap-1 dirty channel)
-//     and the render scheduler (debounces dirties and paints the attached
-//     client). Both are tied to the session context and unwind when the
-//     session is killed (pty.Close unblocks the reader; ctx cancel stops the
-//     scheduler).
-//   - The daemon exits (Serve returns) when the last session is removed, or
-//     when the parent context is cancelled (graceful shutdown notifies any
-//     attached clients with ReasonServerShutdown).
-//
-// Locking: a pane's screen/scrollback and per-client renderer shadow are
-// guarded by pane.mu/tab.mu as appropriate; the attached-client pointer by
-// session.mu; the registry by Daemon.mu. When more than one is held the order
-// is always attachedClient.sendMu > Daemon.mu > session.mu > tab.mu > pane.mu.
-// The PTY reader only ever takes pane.mu, so it never blocks on a slow client.
 package daemon
 import (
 	themeui "github.com/bnema/vev/internal/usecase/theme"
 )
+type transportSnapshot struct {
+	transport   ports.Transport
+	incarnation uint64
+}
 type attachedClient struct {
-	tr                    ports.Transport
+	tr                   ports.Transport
-	output                *outputStateStream
+	transportIncarnation uint64
-	overlays              *overlayRuntime
+	output               *outputStateStream
-	clientID              [16]byte
+	overlays             *overlayRuntime
-	resumeCapable         bool
+	overlayOnce          sync.Once
-	resumeToken           uint64
+	clientID             [16]byte
-	parked                bool
+	resumeCapable        bool
-	echoAck               atomic.Uint64
+	resumeToken          uint64
-	bars                  barCache           // only touched while sendMu is held
+	parked               bool
-	composed              composedFrameCache // only touched while sendMu is held
+	echoAck              atomic.Uint64
-	resizePaint           pendingByteTimer   // guarded by sendMu
+	// pipelineCache is the last successfully emitted composition. pipelineScratch
-	resizePaintGeneration uint64             // guarded by sendMu
+	// is its attachment-owned alternate buffer; both are only touched under
-	resizePaintPending    bool               // guarded by sendMu
+	// sendMu and must never share mutable backing storage.
-	size                  domain.Size
+	pipelineCache   composeCacheInput
-	keys                  *keys.Router
+	pipelineScratch composeCacheInput
-	sess                  Guarded[*session]
+	renderScratch   renderCaptureScratch // only touched while sendMu is held
-	mouseScan             mouse.Scanner
+	// captureFrames is keyed by pane ownership, not the tab-local PaneID, so
-	themeMu               sync.Mutex
+	// snapshots cannot leak when an attachment switches tabs or sessions.
-	theme                 themeui.Theme
+	captureFrames map[*pane]capturedPaneRenderState // only touched while sendMu is held
-	clientTheme           themeui.Theme
+	size          domain.Size
-	lastCursor            cursorOut
+	keys          *keys.Router
+	sess          Guarded[*session]
+	mouseScan     mouse.Scanner
+	themeMu       sync.Mutex
+	theme         themeui.Theme
+	clientTheme   themeui.Theme
+	lastCursor    cursorOut
+	renderStages  renderStageHooks // invoked at real pipeline boundaries while sendMu is held
 	// previousSession is guarded independently. It is retained through temporary
 	// setSession(nil) hand-offs and cleared only on terminal teardown.
 	previousSession Guarded[*session]
 	hasStyle bool
 }
+// pendingByteTimer is retained for overlay input timers. Resize timing belongs
+// exclusively to renderCoordinator.
 type pendingByteTimer struct {
 	timer ports.Timer
 	done  chan struct{}
 	}
 }
-// cancelResizePaint invalidates callbacks before detaching or replacing an
-// attachment. It deliberately takes sendMu before any session/daemon lock.
-func (ac *attachedClient) cancelResizePaint() {
-	ac.sendMu.Lock()
-	ac.resizePaintGeneration++
-	ac.resizePaintPending = false
-	ac.resizePaint.stop()
-	ac.sendMu.Unlock()
-}
 func (ac *attachedClient) initOverlays() {
-	if ac.overlays == nil {
+	ac.overlayOnce.Do(func() { ac.overlays = newOverlayRuntime(ac) })
-		ac.overlays = newOverlayRuntime(ac)
-	}
 }
 func (ac *attachedClient) currentSession() *session { return ac.sess.Get() }
 	}
 }
+// pruneCaptureFrames releases snapshots for panes that have left their owner.
+// Callers must not hold daemon, session, tab, or pane locks.
+func (ac *attachedClient) pruneCaptureFrames(panes ...*pane) {
+	if ac == nil || len(panes) == 0 {
+		return
+	}
+	ac.sendMu.Lock()
+	for _, p := range panes {
+		delete(ac.captureFrames, p)
+	}
+	ac.sendMu.Unlock()
+}
+// clearCaptureFrames releases all pane snapshots when the attachment no longer
+// owns a session.
+func (ac *attachedClient) clearCaptureFrames() {
+	if ac == nil {
+		return
+	}
+	ac.sendMu.Lock()
+	ac.captureFrames = nil
+	ac.sendMu.Unlock()
+}
 func (ac *attachedClient) getTheme() themeui.Theme {
 	ac.themeMu.Lock()
 	defer ac.themeMu.Unlock()
 func (ac *attachedClient) replaceTransport(tr ports.Transport) {
 	ac.linkMu.Lock()
 	ac.tr = tr
+	ac.transportIncarnation++
 	ac.linkMu.Unlock()
 }
+// revokeTransport removes tr only if it is still this attachment's current
+// link. It returns the revoked transport, so asynchronous retirement can
+// close exactly that link without touching a later resume/replacement.
+func (ac *attachedClient) revokeTransport(tr ports.Transport) ports.Transport {
+	if tr == nil {
+		return nil
+	}
+	ac.linkMu.Lock()
+	defer ac.linkMu.Unlock()
+	if ac.tr != tr {
+		return nil
+	}
+	ac.tr = nil
+	ac.transportIncarnation++
+	return tr
+}
+// transportSnapshot binds a send to one concrete link lifetime. Callers must
+// revalidate it after acquiring sendMu before writing to that link.
+func (ac *attachedClient) transportSnapshot() transportSnapshot {
+	ac.linkMu.Lock()
+	defer ac.linkMu.Unlock()
+	return transportSnapshot{transport: ac.tr, incarnation: ac.transportIncarnation}
+}
+func (ac *attachedClient) transportSnapshotCurrent(expected transportSnapshot) bool {
+	ac.linkMu.Lock()
+	defer ac.linkMu.Unlock()
+	return ac.tr == expected.transport && ac.transportIncarnation == expected.incarnation
+}
 func (ac *attachedClient) closeTransport() error {
 	tr := ac.transport()
 	if tr == nil {
 	ac.output.ack(state)
 }
-// takeDeferredPaint reports whether a deferred diff paint is now flushable
-// because acks have caught up, clearing the flag when it is. The caller must
-// then paint outside sendMu (paint re-acquires it).
-func (ac *attachedClient) takeDeferredPaint() (reset bool, ok bool) {
-	ac.sendMu.Lock()
-	defer ac.sendMu.Unlock()
-	return ac.output.takeDeferred()
-}
-// clearPaintDeferred drops any pending deferred-paint flag. Used on park so a
-// resumed client does not inherit a stale deferral.
-func (ac *attachedClient) clearPaintDeferred() {
-	ac.sendMu.Lock()
-	ac.output.clearDeferred()
-	ac.sendMu.Unlock()
-}
 // send serialises a frame onto the client's transport.
 func (ac *attachedClient) send(f ports.Frame) error {
 	_, err := ac.sendTransport(f)
 	return tr, tr.Send(f)
 }
+var errTransportReplaced = errors.New("client transport was replaced")
+// sendExpectedTransport writes only when expected is still the attachment's
+// current transport incarnation. It preserves sendMu -> linkMu lock ordering.
+func (ac *attachedClient) sendExpectedTransport(expected transportSnapshot, f ports.Frame) error {
+	ac.sendMu.Lock()
+	defer ac.sendMu.Unlock()
+	if !ac.transportSnapshotCurrent(expected) {
+		return errTransportReplaced
+	}
+	if expected.transport == nil {
+		return errors.New("client transport is nil")
+	}
+	return expected.transport.Send(f)
+}
 // boundedSend sends f to ac with a deadline watchdog: if the send (including
 // waiting on sendMu behind a wedged paint) does not complete within
 // detachNotifyTimeout, the transport is force-closed, failing the in-flight
 }
 func (d *Daemon) boundedSendErr(ac *attachedClient, f ports.Frame) error {
-	tr, err := d.boundedSendWith(func(capture func(ports.Transport)) error {
+	expected := ac.transportSnapshot()
-		tr, err := ac.sendTransport(f)
+	if expected.transport == nil {
-		capture(tr)
+		return errors.New("client transport is nil")
-		return err
+	}
+	tr, err := d.boundedSendWith(expected.transport, func() error {
+		return ac.sendExpectedTransport(expected, f)
 	})
 	if err != nil && errors.Is(err, errSendTimedOut) {
 		_ = ac.closeCapturedTransport(tr)
 func (d *Daemon) boundedSendOutputErrTransport(ac *attachedClient, b []byte) (ports.Transport, error) {
 	frame := ac.output.sideEffect(b, ac.echoAck.Load())
+	expected := ac.transportSnapshot()
+	if expected.transport == nil {
+		return nil, errors.New("client transport is nil")
+	}
 	ac.sendMu.Lock()
-	tr := ac.transport()
+	if !ac.transportSnapshotCurrent(expected) {
-	if tr == nil {
 		ac.sendMu.Unlock()
-		return nil, errors.New("client transport is nil")
+		return expected.transport, errTransportReplaced
 	}
-	if owned, ok := tr.(ports.OwnedSynchronousTransport); ok {
+	if owned, ok := expected.transport.(ports.OwnedSynchronousTransport); ok {
 		err := owned.SendSynchronous(frame)
 		ac.sendMu.Unlock()
-		return tr, err
+		return expected.transport, err
 	}
 	ac.sendMu.Unlock()
-	return d.boundedSendWith(func(capture func(ports.Transport)) error {
+	return d.boundedSendWith(expected.transport, func() error {
-		ac.sendMu.Lock()
+		return ac.sendExpectedTransport(expected, frame)
-		defer ac.sendMu.Unlock()
-		tr := ac.transport()
-		capture(tr)
-		if tr == nil {
-			return errors.New("client transport is nil")
-		}
-		return tr.Send(frame)
 	})
 }
 var errSendTimedOut = errors.New("send timed out")
-func (d *Daemon) boundedSendWith(send func(capture func(ports.Transport)) error) (ports.Transport, error) {
+func (d *Daemon) boundedSendWith(tr ports.Transport, send func() error) (ports.Transport, error) {
-	return d.boundedSendWithTimeout(detachNotifyTimeout, send)
+	return d.boundedSendWithTimeout(detachNotifyTimeout, tr, send)
 }
-func (d *Daemon) boundedSendWithTimeout(timeout time.Duration, send func(capture func(ports.Transport)) error) (ports.Transport, error) {
+func (d *Daemon) boundedSendWithTimeout(timeout time.Duration, tr ports.Transport, send func() error) (ports.Transport, error) {
 	timer := d.clock.NewTimer(timeout)
 	result := make(chan error, 1)
-	var (
-		capturedMu sync.Mutex
-		captured   ports.Transport
-	)
-	capture := func(tr ports.Transport) {
-		capturedMu.Lock()
-		captured = tr
-		capturedMu.Unlock()
-	}
-	capturedTransport := func() ports.Transport {
-		capturedMu.Lock()
-		defer capturedMu.Unlock()
-		return captured
-	}
 	go func() {
-		result <- send(capture)
+		result <- send()
 	}()
 	select {
 	case err := <-result:
 		timer.Stop()
-		return capturedTransport(), err
+		return tr, err
 	case <-timer.C():
 		select {
 		case err := <-result:
-			return capturedTransport(), err
+			return tr, err
 		default:
 		}
 		d.log.Warn("bounded send timed out; force closing client transport")
-		return capturedTransport(), errSendTimedOut
+		return tr, errSendTimedOut
 	}
 }
 }
 func (d *Daemon) attachClient(sess *session, tr ports.Transport, sz domain.Size, opts attachClientOptions) (*attachedClient, *attachedClient) {
+	ac, old, cleanup := d.attachClientDeferred(sess, tr, sz, opts)
+	cleanup.finish()
+	return ac, old
+}
+func (d *Daemon) attachClientDeferred(sess *session, tr ports.Transport, sz domain.Size, opts attachClientOptions) (*attachedClient, *attachedClient, renderLifecycleCleanup) {
 	resumeToken := uint64(0)
 	if opts.resumeCapable {
 		resumeToken = d.nextResumeTokenLocked()
 	ac.initOverlays()
 	ac.setSession(sess)
 	ac.keys = keys.NewRouter(d.clock, daemonKeyHandler{d: d, ac: ac}, &d.bindings)
+	// Daemon attachment callers hold d.mu, serialising this preparation with
+	// other publications. Bind coordinator ownership before sess.client becomes
+	// visible: an old deadline can then never target this new output chain.
 	sess.mu.Lock()
 	old := sess.client
-	sess.client = ac
 	name := sess.name
 	sess.mu.Unlock()
+	_, cleanup := d.attachCoordinatorDeferred(sess, old, ac, false)
+	sess.mu.Lock()
+	sess.client = ac
+	sess.mu.Unlock()
 	d.touchMRU(sess)
 	d.log.Info("client attached", "session", name, "resume", opts.resumeCapable)
 	d.applyHostTheme(sess, ac, d.effectiveTheme(themeui.Theme{}), true)
-	return ac, old
+	return ac, old, cleanup
+}
+// handoffCoordinator prepares a cross-session ownership transfer before the
+// destination publishes sess.client. It never takes daemon/session locks while
+// taking sendMu, avoiding a d.mu/sendMu cycle with transport callbacks.
+func (d *Daemon) handoffCoordinator(from, target *session, old, current *attachedClient) []renderLifecycleCleanup {
+	cleanups := make([]renderLifecycleCleanup, 0, 2)
+	if rc := from.renderCoordinator(); rc != nil {
+		cleanups = append(cleanups, rc.beginDetach(current))
+	}
+	current.sendMu.Lock()
+	current.output.rebase()
+	// Capture snapshots are attachment-owned. Discard them while the attachment
+	// is exclusively held so the transfer releases panes no longer reachable
+	// from the target session.
+	current.captureFrames = nil
+	current.sendMu.Unlock()
+	_, cleanup := d.attachCoordinatorDeferred(target, old, current, true)
+	return append(cleanups, cleanup)
+}
+func finishRenderLifecycleCleanups(cleanups []renderLifecycleCleanup) {
+	for _, cleanup := range cleanups {
+		cleanup.finish()
+	}
+}
+// attachCoordinator is the sole direct attachment handoff. It creates at
+// most one coordinator for sess and changes its identity before any caller
+// can publish resize or render state for the new client.
+func (d *Daemon) attachCoordinator(sess *session, old, current *attachedClient, ready bool) *renderCoordinator {
+	rc, cleanup := d.attachCoordinatorDeferred(sess, old, current, ready)
+	cleanup.finish()
+	return rc
+}
+func (d *Daemon) attachCoordinatorDeferred(sess *session, old, current *attachedClient, ready bool) (*renderCoordinator, renderLifecycleCleanup) {
+	sess.mu.Lock()
+	rc := sess.renderCoordinator()
+	if rc == nil {
+		rc = newRenderCoordinator(renderCoordinatorOptions{
+			clock:    d.clock,
+			observer: d.runtimeObserver,
+			wake: func(w renderWake) {
+				// Never reread sess.client here: a wake is bound to both its
+				// attachment and its coordinator incarnation. The second check in
+				// paint occurs under sendMu, closing a park/resume race after this
+				// unlocked coordinator validation.
+				if w.lease != nil && rc.wakeCurrent(w) {
+					d.paint(sess, w.lease.attachment, w.reset, w.lease)
+				}
+			},
+			ackReady: func() bool {
+				sess.mu.Lock()
+				attached := sess.client
+				sess.mu.Unlock()
+				if attached == nil {
+					return false
+				}
+				attached.sendMu.Lock()
+				ready := !attached.output.atCapacity()
+				attached.sendMu.Unlock()
+				return ready
+			},
+		})
+		sess.installRenderCoordinator(rc)
+	}
+	sess.mu.Unlock()
+	var cleanup renderLifecycleCleanup
+	if old != nil {
+		cleanup = rc.beginReplace(old, current, ready)
+	} else if current != nil {
+		rc.attachWithReadiness(current, ready)
+	}
+	return rc, cleanup
 }
 // resetScreenDefaultColors clears the known default foreground/background
 	d.applyHostTheme(sess, nil, d.effectiveTheme(themeui.Theme{}), true)
 }
-func (d *Daemon) detachReplacedClient(old *attachedClient) {
+// retireReplacedClient owns old-link retirement and obsolete render-worker
+// joins outside the replacement handshake. If a render already owns sendMu,
+// close immediately interrupts it; no Detached frame can safely overtake that
+// render. Otherwise the old link gets its bounded protocol notice first.
+func (d *Daemon) retireReplacedClient(old *attachedClient, cleanup renderLifecycleCleanup) {
 	if old == nil {
+		d.attachmentCleanupWg.Go(cleanup.finish)
 		return
 	}
-	old.cancelResizePaint()
+	oldTransport := old.transportSnapshot().transport
+	blockedRender := !old.sendMu.TryLock()
+	if !blockedRender {
+		old.sendMu.Unlock()
+	}
-	d.unregisterPreview(old)
+	d.attachmentCleanupWg.Go(func() {
-	old.clearPreviousSession()
+		if !blockedRender {
-	old.setSession(nil)
+			// A healthy idle client observes the required replacement notice. Its
-	d.log.Info("client displaced")
+			// watchdog still bounds a peer that stops draining after this check.
-	// Async + bounded: a dead or wedged old client must not stall the new
+			d.boundedSend(old, frameDetached(ports.ReasonDetach))
-	// client's handshake.
+		}
-	d.notifyDetachedAsync(old, ports.ReasonDetach)
+		_ = old.closeCapturedTransport(old.revokeTransport(oldTransport))
+		d.unregisterPreview(old)
+		old.clearPreviousSession()
+		old.setSession(nil)
+		d.log.Info("client displaced")
+	})
+	// Join independently: the close above releases a blocked render while the
+	// new attachment is already fully published.
+	d.attachmentCleanupWg.Go(cleanup.finish)
 }
 // firstPaint guarantees the freshly attached client sees the full screen: if
 	wsz := tb.size
 	tb.mu.Unlock()
+	outerResizeAccepted := false
 	if clientSize.Valid() && wsz != tabSize(clientSize) {
-		d.resizeForFirstPaint(sess, ac, clientSize)
+		outerResizeAccepted = d.resizeForFirstPaint(sess, ac, clientSize)
 	}
 	d.refreshBarScriptsIfDue(sess, d.clock.Now(), true)
-	d.paint(sess, ac, true)
+	// Activation can synchronously resize a retained floating pane. An accepted
-	d.activateTab(sess, tb)
+	// synchronous outer request already includes that pane and emits the reset,
+	// but activation still performs its warmup work.
+	activationResized := d.activateTabAfterResize(sess, tb, outerResizeAccepted)
+	if !outerResizeAccepted && !activationResized {
+		d.invalidateRenderNow(sess, ac, true, "client.go")
+	}
 }
 // runConnLoop is the per-connection input router: it pumps client messages
 		case ports.MsgAck:
 			if ack, derr := ports.UnmarshalAck(f.Payload); derr == nil {
 				ac.ackOutputState(ack.AckedStateNum)
-				if reset, ok := ac.takeDeferredPaint(); ok {
+				if rc := sess.renderCoordinator(); rc != nil {
-					d.paint(sess, ac, reset)
+					rc.notifyAck()
 				}
 			}
 		case ports.MsgPing:
 	if !sess.detachIfCurrent(ac) {
 		return // already displaced by a newer client; nothing to do
 	}
-	ac.cancelResizePaint()
+	if rc := sess.renderCoordinator(); rc != nil {
+		rc.noteDetach(ac)
+	}
 	d.unregisterPreview(ac)
 	sess.mu.Lock()
 	ephemeral := sess.ephemeral
 	if !d.applyHostTheme(sess, ac, t, false) {
 		return
 	}
-	d.paint(sess, ac, true)
+	d.invalidateRender(sess, ac, true, "client.go")
 }
 func (d *Daemon) resize(sess *session, ac *attachedClient, sz domain.Size) {
-	d.resizeWithPaint(sess, ac, sz, true)
+	d.requestTransactionalResize(sess, ac, sz, false)
 }
-// resizeForFirstPaint applies attach geometry without installing an idle timer:
+// resizeForFirstPaint retains attach's synchronous geometry guarantee. The
-// firstPaint immediately emits the required full frame itself.
+// returned value reports whether the synchronous request was accepted.
-func (d *Daemon) resizeForFirstPaint(sess *session, ac *attachedClient, sz domain.Size) {
+func (d *Daemon) resizeForFirstPaint(sess *session, ac *attachedClient, sz domain.Size) bool {
-	d.resizeWithPaint(sess, ac, sz, false)
+	return d.requestTransactionalResize(sess, ac, sz, true)
-}
-func (d *Daemon) resizeWithPaint(sess *session, ac *attachedClient, sz domain.Size, schedulePaint bool) {
-	if !sz.Valid() {
-		return
-	}
-	d.exitCopyMode(ac)
-	if ac != nil {
-		// Suspend emission until every pane has adopted this geometry and the
-		// bounded resize paint is armed. PTY readers remain independent.
-		ac.sendMu.Lock()
-		defer ac.sendMu.Unlock()
-		if ac.currentSession() != sess {
-			return
-		}
-		ac.size = sz
-	}
-	tbSize := tabSize(sz)
-	sess.mu.Lock()
-	tabs := append([]*tab(nil), sess.tabs...)
-	sess.mu.Unlock()
-	if len(tabs) == 0 {
-		return
-	}
-	for _, tb := range tabs {
-		tb.mu.Lock()
-		tb.size = tbSize
-		d.applyLayoutLocked(tb)
-		tb.mu.Unlock()
-	}
-	// Only the shown tab's floating terminal tracks the client size. This is
-	// outside tab.mu so its PTY resize cannot block tab state.
-	d.resizeActiveFloating(sess.activeTab())
-	markSnapshotDirty(sess)
-	if ac != nil && schedulePaint {
-		d.refreshBarScriptsIfDue(sess, d.clock.Now(), true)
-		d.scheduleResizePaintLocked(sess, ac)
-	}
 }
 // detachOnSendError drops a client whose transport failed, leaving the session
 		return
 	}
 	if sess.detachIfCurrent(ac) {
-		ac.cancelResizePaint()
+		if rc := sess.renderCoordinator(); rc != nil {
+			rc.noteDetach(ac)
+		}
 		d.unregisterPreview(ac)
 		if d.parkAttachment(sess, ac) {
 			_ = ac.closeCapturedTransport(failed)

M internal/usecase/daemon/clipboard.go +1 -1

@@ -1,6 +1,6 @@
1		// Package daemon holds vev's server-side session multiplexer use case: the	1		// Package daemon holds vev's server-side session multiplexer use case: the
2		// accept loop, the ephemeral/named session registry, the per-tab PTY reader	2		// accept loop, the ephemeral/named session registry, the per-tab PTY reader
3	-	// and VT screen, and the per-client debounced render scheduler.	3	+	// and VT screen, and the per-client debounced render coordinator.
4		package daemon	4		package daemon
5			5
6		import (	6		import (

A internal/usecase/daemon/copy_enter_allocation_budget_race_test.go +8 -0

@@ -0,0 +1,8 @@
	1	+	//go:build race
	2	+
	3	+	package daemon
	4	+
	5	+	// Race instrumentation allocations are not part of the parent benchmark's
	6	+	// allocation baseline. TestCopyEnterAllocationBudget still exercises copy
	7	+	// entry and validates its result under -race.
	8	+	const copyEnterAllocationBudgetEnabled = false

A internal/usecase/daemon/copy_enter_allocation_budget_test.go +7 -0

@@ -0,0 +1,7 @@
	1	+	//go:build !race
	2	+
	3	+	package daemon
	4	+
	5	+	// copyEnterAllocationBudgetEnabled keeps the allocation gate aligned with the
	6	+	// non-race parent benchmark, whose allocation baseline defines the budget.
	7	+	const copyEnterAllocationBudgetEnabled = true

M internal/usecase/daemon/copymode.go +10 -38

-// Package daemon holds vev's server-side session multiplexer use case: the
+// Package daemon holds vev's server-side session multiplexer use case.
-// accept loop, the ephemeral/named session registry, the per-tab PTY reader
-// and VT screen, and the per-client debounced render scheduler.
-//
-// Concurrency model (sessions own one or more PTY-backed tabs):
-//
-//   - Serve runs the accept loop. Each accepted connection is handled by its
-//     own goroutine (handleConn): it reads the first frame and routes it to a
-//     session create/attach, a list, or a kill.
-//   - Per session there are exactly two long-lived goroutines: the PTY reader
-//     (drains child output into the VT screen and pokes a cap-1 dirty channel)
-//     and the render scheduler (debounces dirties and paints the attached
-//     client). Both are tied to the session context and unwind when the
-//     session is killed (pty.Close unblocks the reader; ctx cancel stops the
-//     scheduler).
-//   - The daemon exits (Serve returns) when the last session is removed, or
-//     when the parent context is cancelled (graceful shutdown notifies any
-//     attached clients with ReasonServerShutdown).
-//
-// Locking: a pane's screen/scrollback and per-client renderer shadow are
-// guarded by pane.mu/tab.mu as appropriate; the attached-client pointer by
-// session.mu; the registry by Daemon.mu. When more than one is held the order
-// is always attachedClient.sendMu > Daemon.mu > session.mu > tab.mu > pane.mu.
-// The PTY reader only ever takes pane.mu, so it never blocks on a slow client.
 package daemon
 import (
 	if delta > 0 && rt.copyMode.AtBottom(*document) {
 		rt.clearCopyModeLocked()
 		rt.copyMu.Unlock()
-		d.paint(sess, ac, true)
+		d.invalidateRender(sess, ac, true, "copymode.go")
 		return
 	}
 	rt.copyMode.Move(*document, delta)
 	}
 	rt.copyMu.Unlock()
-	d.paint(sess, ac, true)
+	d.invalidateRender(sess, ac, true, "copymode.go")
 }
 func (d *Daemon) enterCopyMode(sess *session, ac *attachedClient) {
 	// Capture the live pane under pane.mu; all subsequent copy interaction uses
 	// the immutable document while holding only copyMu.
 	p.mu.Lock()
-	document := scopy.NewSnapshot(p.scrollback, p.screen.Frame)
+	document := scopy.NewSnapshot(p.history, p.screen.Frame)
 	p.mu.Unlock()
 	if !d.publishCopyMode(sess, ac, tb, p, document, nil) {
 		return
 	}
-	d.paint(sess, ac, true)
+	d.invalidateRender(sess, ac, true, "copymode.go")
 }
 // publishCopyMode installs a non-renderable candidate, validates its captured
 	rt.copyMu.Unlock()
 	if changed {
-		d.paint(sess, ac, true)
+		d.invalidateRender(sess, ac, true, "copymode.go")
 	}
 }
 		changed, closeSearch, accepted := d.routeCopySearchInputLocked(rt, *document, data)
 		rt.copyMu.Unlock()
 		if changed || closeSearch || accepted {
-			d.paint(sess, ac, true)
+			d.invalidateRender(sess, ac, true, "copymode.go")
 		}
 		return
 	}
 		rt.copyMu.Unlock()
 	}
 	if exit {
-		d.paint(sess, ac, true)
+		d.invalidateRender(sess, ac, true, "copymode.go")
 		return
 	}
 	if changed {
-		d.paint(sess, ac, true)
+		d.invalidateRender(sess, ac, true, "copymode.go")
 	}
 }
 		rt.copyMu.Unlock()
 		if sess := ac.currentSession(); sess != nil {
-			d.paint(sess, ac, true)
+			d.invalidateRender(sess, ac, true, "copymode.go")
 		}
 	})
 }
 		len(data) == 3 && data[0] == 0x1b && data[1] == '[' && (data[2] == '5' || data[2] == '6')
 }
-func composeCopySearchClientFrame(model *visualsearch.Model, base renderer.Frame, styles ...themeStyles) (renderer.Frame, []renderer.Damage) {
-	styleSet := resolveThemeStyles(styles)
-	return composeModalClientFrame(base, copySearchModal, styleSet, styleSet.selection, model.Render)
-}
 func composeCopyClientFrame(mode *scopy.Mode, document *scopy.Snapshot, target domain.Rect, frame renderer.Frame, bars barState) (renderer.Frame, []renderer.Damage) {
 	if mode == nil || document == nil || target.Width <= 0 || target.Height <= 0 || frame.Width <= 0 || frame.Height <= 0 {
 		return frame, nil

M internal/usecase/daemon/copymode_lifecycle_test.go +3 -3


@@ -54,7 +54,7 @@ func TestCopyModeLifecycleClosePaneClearsRecoveredClientState(t *testing.T) {
54		sess.client = ac	54		sess.client = ac
55		ac.setSession(sess)	55		ac.setSession(sess)
56		closing.mu.Lock()	56		closing.mu.Lock()
57	-	document := scopy.NewSnapshot(closing.scrollback, closing.screen.Frame)	57	+	document := scopy.NewSnapshot(closing.history, closing.screen.Frame)
58		closing.mu.Unlock()	58		closing.mu.Unlock()
59		ac.overlays.copyMu.Lock()	59		ac.overlays.copyMu.Lock()
60		ac.overlays.copyPane = closing	60		ac.overlays.copyPane = closing

@@ -94,7 +94,7 @@ func TestCopyModeLifecycleDoesNotRenderCandidateBeforeValidation(t *testing.T) {
94		tb := sess.activeTab()	94		tb := sess.activeTab()
95		p := tb.focusedPane()	95		p := tb.focusedPane()
96		p.mu.Lock()	96		p.mu.Lock()
97	-	document := scopy.NewSnapshot(p.scrollback, p.screen.Frame)	97	+	document := scopy.NewSnapshot(p.history, p.screen.Frame)
98		p.mu.Unlock()	98		p.mu.Unlock()
99			99
100		// Hold session validation after publication and make the captured target	100		// Hold session validation after publication and make the captured target

@@ -157,7 +157,7 @@ func TestCopyModeLifecycleRejectsPublicationForInactiveOrRemovedPane(t *testing.
157		tb := sess.activeTab()	157		tb := sess.activeTab()
158		p := tb.focusedPane()	158		p := tb.focusedPane()
159		p.mu.Lock()	159		p.mu.Lock()
160	-	document := scopy.NewSnapshot(p.scrollback, p.screen.Frame)	160	+	document := scopy.NewSnapshot(p.history, p.screen.Frame)
161		p.mu.Unlock()	161		p.mu.Unlock()
162		tc.invalidate(sess, tb, p)	162		tc.invalidate(sess, tb, p)
163			163

M internal/usecase/daemon/copymode_test.go +43 -28

 	tb := sess.activeTab()
 	pane := tb.focusedPane()
 	pane.mu.Lock()
-	snap := scopy.NewSnapshot(pane.scrollback, pane.screen.Frame)
+	snap := scopy.NewSnapshot(pane.history, pane.screen.Frame)
 	pane.mu.Unlock()
 	mode := scopy.NewMode(snap)
 	bars := barState{status: sess.statusSegments(true)}
 	tb := sess.activeTab()
 	tb.focusedPane().screen = vt.NewScreen(12, 3)
 	tb.focusedPane().screen.Write([]byte("live"))
-	snap := scopy.NewSnapshot(tb.focusedPane().scrollback, tb.focusedPane().screen.Frame)
+	snap := scopy.NewSnapshot(tb.focusedPane().history, tb.focusedPane().screen.Frame)
 	mode := scopy.NewMode(snap)
 	bars := barState{status: sess.statusSegments(true)}
 	writes := make(chan []byte, 1)
 	p, _ := newBlockingPTYWithWrites(t, writes)
 	d, sess, ac, sends := newManualSessionWithPTYs(t, p)
-	sess.tabs[0].focusedPane().scrollback = scopy.NewScrollback(4)
+	installTestHistory(sess.tabs[0].focusedPane(), vt.HistoryConfig{MaxRows: 4})
 	sess.tabs[0].focusedPane().screen.Write([]byte("live"))
 	d.handleInput(sess, ac, []byte("\x1b "))
 	p, _ := newBlockingPTY(t)
 	d, sess, ac, sends := newManualSessionWithPTYs(t, p)
 	pane := sess.tabs[0].focusedPane()
-	pane.scrollback = scopy.NewScrollback(4)
+	installTestHistory(pane, vt.HistoryConfig{MaxRows: 4})
-	pane.scrollback.Append(testRow("old alpha"))
+	pane.history.Append(testRow("old alpha"))
 	copy(pane.screen.Frame.Row(0), testRow("live alpha"))
 	d.enterCopyMode(sess, ac)
 	writes := make(chan []byte, 1)
 	p, _ := newBlockingPTYWithWrites(t, writes)
 	d, sess, ac, sends := newManualSessionWithPTYs(t, p)
-	sess.tabs[0].focusedPane().scrollback = scopy.NewScrollback(4)
+	installTestHistory(sess.tabs[0].focusedPane(), vt.HistoryConfig{MaxRows: 4})
-	sess.tabs[0].focusedPane().scrollback.Append(testRow("old1    "))
+	sess.tabs[0].focusedPane().history.Append(testRow("old1    "))
-	sess.tabs[0].focusedPane().scrollback.Append(testRow("old2    "))
+	sess.tabs[0].focusedPane().history.Append(testRow("old2    "))
 	sess.tabs[0].focusedPane().screen.Write([]byte("live"))
 	d.enterCopyMode(sess, ac)
 		t.Fatalf("live repaint = %q, want live screen", string(liveMsg.Data))
 	}
-	d.paint(sess, ac, true)
+	d.paint(sess, ac, true, nil)
 	followup := awaitFrame(t, sends, ports.MsgOutput)
 	followupMsg, err := ports.UnmarshalOutput(followup.Payload)
 	require.NoError(t, err)
 	p.EXPECT().Close().RunAndReturn(func() error { closeOnce.Do(func() { close(readDone) }); return nil }).Maybe()
 	p.EXPECT().Pid().Return(4242).Maybe()
-	d := newTestDaemon(t, newFactory(t, p), stubClock{})
+	// Rendering is coordinator-driven: advance a controllable clock through the
+	// retained resize timer and its resulting coordinator wake rather than
+	// relying on wall-clock debounce delivery.
+	clk := newCoordinatorMockClock(t, 16)
+	d := newTestDaemon(t, newFactory(t, p), clk.clock)
 	tr, sends, releaseConn := newConn(t, mustHello(ports.IntentNew, "work", domain.Size{Cols: 16, Rows: 5}))
+	advanceRender := func() {
+		deadline := time.NewTimer(2 * time.Second)
+		defer deadline.Stop()
+		select {
+		case timer := <-clk.timers:
+			timer.ch <- time.Time{}
+		case <-deadline.C:
+			t.Fatal("coordinator did not arm a controllable render timer")
+		}
+	}
 	var hg sync.WaitGroup
 	hg.Go(func() { d.handleConn(tr) })
 	awaitFrame(t, sends, ports.MsgWelcome)
+	advanceRender() // initial coordinator invalidation
 	awaitFrame(t, sends, ports.MsgOutput)
 	for i := range 12 {
 		}
 		p.mu.Lock()
 		defer p.mu.Unlock()
-		return scopy.NewSnapshot(p.scrollback, p.screen.Frame).Len() >= 12
+		return scopy.NewSnapshot(p.history, p.screen.Frame).Len() >= 12
 	}, 2*time.Second, 5*time.Millisecond)
 	sess := firstSession(d)
 	ac := sess.client
 	require.NotNil(t, ac)
 	d.handleInput(sess, ac, []byte("\x1b "))
-	awaitFrame(t, sends, ports.MsgOutput)
+	awaitCoordinatorOutput(t, sends, clk.timers, "while advancing render clock", "controllable timers did not produce an output frame")
 	d.handleInput(sess, ac, []byte("VIS\r"))
-	awaitFrame(t, sends, ports.MsgOutput)
+	awaitCoordinatorOutput(t, sends, clk.timers, "while advancing render clock", "controllable timers did not produce an output frame")
 	d.handleInput(sess, ac, []byte{'g', ' ', 'G', 'y'})
 	var payload string
 func TestCopyModeEscapeRestoresLiveFullRepaint(t *testing.T) {
 	p, _ := newBlockingPTY(t)
 	d, sess, ac, sends := newManualSessionWithPTYs(t, p)
-	sess.tabs[0].focusedPane().scrollback = scopy.NewScrollback(4)
+	installTestHistory(sess.tabs[0].focusedPane(), vt.HistoryConfig{MaxRows: 4})
 	sess.tabs[0].focusedPane().screen.Write([]byte("live"))
 	d.enterCopyMode(sess, ac)
 		t.Run(tc.name, func(t *testing.T) {
 			p, _ := newBlockingPTY(t)
 			d, sess, ac, sends := newManualSessionWithPTYs(t, p)
-			sess.tabs[0].focusedPane().scrollback = scopy.NewScrollback(4)
+			installTestHistory(sess.tabs[0].focusedPane(), vt.HistoryConfig{MaxRows: 4})
-			sess.tabs[0].focusedPane().scrollback.Append(testRow("old1    "))
+			sess.tabs[0].focusedPane().history.Append(testRow("old1    "))
-			sess.tabs[0].focusedPane().scrollback.Append(testRow("old2    "))
+			sess.tabs[0].focusedPane().history.Append(testRow("old2    "))
 			sess.tabs[0].focusedPane().screen.Write([]byte("live"))
 			d.enterCopyMode(sess, ac)
 func TestCopyModeOversizedYankShowsTooLargeFeedback(t *testing.T) {
 	p, _ := newBlockingPTY(t)
 	d, sess, ac, sends := newManualSessionWithPTYs(t, p)
-	sess.tabs[0].focusedPane().scrollback = scopy.NewScrollback(1)
+	installTestHistory(sess.tabs[0].focusedPane(), vt.HistoryConfig{MaxRows: 1})
 	longLine := strings.Repeat("x", scopy.OSC52MaxPayloadBytes+1)
 	sess.tabs[0].focusedPane().screen.Frame = renderer.NewFrame(len(longLine), 1)
 	copy(sess.tabs[0].focusedPane().screen.Frame.Row(0), testRow(longLine))
 	d, sess, ac, sends := newManualSessionWithPTYs(t, p)
 	clk := &signalClock{timers: make(chan *signalTimer, 1)}
 	d.clock = clk
-	sess.tabs[0].focusedPane().scrollback = scopy.NewScrollback(4)
+	installTestHistory(sess.tabs[0].focusedPane(), vt.HistoryConfig{MaxRows: 4})
 	sess.tabs[0].focusedPane().screen.Write([]byte("live"))
 	d.enterCopyMode(sess, ac)
 	d, sess, ac, sends := newManualSessionWithPTYs(t, p)
 	clk := &signalClock{timers: make(chan *signalTimer, 1)}
 	d.clock = clk
-	sess.tabs[0].focusedPane().scrollback = scopy.NewScrollback(4)
+	installTestHistory(sess.tabs[0].focusedPane(), vt.HistoryConfig{MaxRows: 4})
 	sess.tabs[0].focusedPane().screen.Write([]byte("live"))
 	d.enterCopyMode(sess, ac)
 		t.Run(tc.name, func(t *testing.T) {
 			p, _ := newBlockingPTY(t)
 			d, sess, ac, sends := newManualSessionWithPTYs(t, p)
-			sess.tabs[0].focusedPane().scrollback = scopy.NewScrollback(4)
+			installTestHistory(sess.tabs[0].focusedPane(), vt.HistoryConfig{MaxRows: 4})
 			sess.tabs[0].focusedPane().screen.Write([]byte("live"))
 			d.enterCopyMode(sess, ac)
 func TestCopyModeEnterExitConcurrentWithPaintRace(t *testing.T) {
 	p, _ := newBlockingPTY(t)
 	d, sess, ac, sends := newManualSessionWithPTYs(t, p)
-	sess.tabs[0].focusedPane().scrollback = scopy.NewScrollback(4)
+	installTestHistory(sess.tabs[0].focusedPane(), vt.HistoryConfig{MaxRows: 4})
 	sess.tabs[0].focusedPane().screen.Write([]byte("live"))
 	done := make(chan struct{})
 			d.handleInput(sess, ac, []byte("q"))
 		})
 		wg.Go(func() {
-			d.paint(sess, ac, true)
+			d.paint(sess, ac, true, nil)
 		})
 	}
 	wg.Wait()
 			var want *pane
 			if tc.useFloating {
 				want = installFloatingCopyFixture(t, sess, domain.Size{Cols: 20, Rows: 3})
-				want.scrollback.Append(testRow("flt-old"))
+				want.history.Append(testRow("flt-old"))
 				want.screen.Write([]byte("flt-live"))
 			} else {
 				want = main
-				want.scrollback.Append(testRow("one-old"))
+				want.history.Append(testRow("one-old"))
 				want.screen.Write([]byte("one-live"))
 			}
 	d, sess, ac, sends := newManualSessionWithPTYs(t, normal)
 	fp := installFloatingCopyFixture(t, sess, domain.Size{Cols: 20, Rows: 3})
 	for i := range 30 {
-		fp.scrollback.Append(testRow(fmt.Sprintf("old-%02d", i)))
+		fp.history.Append(testRow(fmt.Sprintf("old-%02d", i)))
 	}
 	fp.screen.Write([]byte("live"))
 	fp.mu.Lock()
-	total := scopy.NewSnapshot(fp.scrollback, fp.screen.Frame).Len()
+	total := scopy.NewSnapshot(fp.history, fp.screen.Frame).Len()
 	fp.mu.Unlock()
 	d.enterCopyMode(sess, ac)
 	d, sess, ac, sends := newManualSessionWithPTYs(t, normal)
 	fp := installFloatingCopyFixture(t, sess, domain.Size{Cols: 20, Rows: 3})
 	for i := range 5 {
-		fp.scrollback.Append(testRow(fmt.Sprintf("old-%d", i)))
+		fp.history.Append(testRow(fmt.Sprintf("old-%d", i)))
 	}
 	fp.screen.Write([]byte("live"))

M internal/usecase/daemon/daemon.go +93 -71

@@ -1,29 +1,4 @@
1	-	// Package daemon holds vev's server-side session multiplexer use case: the	1	+	// Package daemon holds vev's server-side session multiplexer use case.
2	-	// accept loop, the ephemeral/named session registry, the per-tab PTY reader
3	-	// and VT screen, and the per-client debounced render scheduler.
4	-	//
5	-	// Concurrency model (sessions own one or more PTY-backed tabs):
6	-	//
7	-	// - Serve runs the accept loop. Each accepted connection is handled by its
8	-	// own goroutine (handleConn): it reads the first frame and routes it to a
9	-	// session create/attach, a list, or a kill.
10	-	// - Per session there are exactly two long-lived goroutines: the PTY reader
11	-	// (drains child output into the VT screen and pokes a cap-1 dirty channel)
12	-	// and the render scheduler (debounces dirties and paints the attached
13	-	// client). Both are tied to the session context and unwind when the
14	-	// session is killed (pty.Close unblocks the reader; ctx cancel stops the
15	-	// scheduler).
16	-	// - The daemon exits (Serve returns) when the last session is removed, or
17	-	// when the parent context is cancelled (graceful shutdown notifies any
18	-	// attached clients with ReasonServerShutdown). Client detach does not end
19	-	// a session: ephemeral sessions stay in memory until explicitly killed or
20	-	// until the daemon exits, and named sessions can also persist on disk.
21	-	//
22	-	// Locking: a pane's screen/scrollback and per-client renderer shadow are
23	-	// guarded by pane.mu/tab.mu as appropriate; the attached-client pointer by
24	-	// session.mu; the registry by Daemon.mu. When more than one is held the order
25	-	// is always attachedClient.sendMu > Daemon.mu > session.mu > tab.mu > pane.mu.
26	-	// The PTY reader only ever takes pane.mu, so it never blocks on a slow client.
27		package daemon	2		package daemon
28			3
29		import (	4		import (

@@ -41,14 +16,13 @@ import (
41		"github.com/bnema/vev/internal/persist"	16		"github.com/bnema/vev/internal/persist"
42		"github.com/bnema/vev/internal/ports"	17		"github.com/bnema/vev/internal/ports"
43		"github.com/bnema/vev/internal/usecase/keys"	18		"github.com/bnema/vev/internal/usecase/keys"
			19	+	snapcodec "github.com/bnema/vev/internal/usecase/snapshot"
44		)	20		)
45			21
46		// Scheduler debounce bounds. Idle updates use the minimum for low latency;	22		// Scheduler debounce bounds. Idle updates use the minimum for low latency;
47		// sustained floods step toward the maximum to reduce frame/syscall pressure.	23		// sustained floods step toward the maximum to reduce frame/syscall pressure.
48		const (	24		const (
49		minDebounceInterval = 2 * time.Millisecond	25		minDebounceInterval = 2 * time.Millisecond
50	-	maxDebounceInterval = 16 * time.Millisecond
51	-	debounceStep = 2 * time.Millisecond
52		maxSyncUpdateDuration = 500 * time.Millisecond	26		maxSyncUpdateDuration = 500 * time.Millisecond
53		)	27		)
54			28

@@ -117,17 +91,33 @@ type Daemon struct {
117		// beforeRecentSessionHandoff is a deterministic test seam for the narrow	91		// beforeRecentSessionHandoff is a deterministic test seam for the narrow
118		// interval between JRS validation and its committed hand-off.	92		// interval between JRS validation and its committed hand-off.
119		beforeRecentSessionHandoff func()	93		beforeRecentSessionHandoff func()
120	-		94	+	ptys ports.PTYFactory
121	-	ptys ports.PTYFactory	95	+	clock ports.Clock
122	-	clock ports.Clock	96	+	log *slog.Logger
123	-	log *slog.Logger	97	+	runtimeObserver ports.RuntimeObserver
124	-	baseEnv []string	98	+	baseEnv []string
125	-	shell string	99	+	shell string
126	-	shellArgs []string	100	+	shellArgs []string
127	-	persist *persist.Persister	101	+	persist *persist.Persister
128	-	persistEnabled bool	102	+	persistEnabled bool
129	-	snaps ports.SnapshotStore	103	+	snaps ports.SnapshotStore
130	-	snapsEnabled bool	104	+	snapsEnabled bool
			105	+	snapshotMarshal func(snapcodec.Session) ([]byte, error)
			106	+	snapshotJobs chan *snapshotCapture
			107	+	snapshotWorkerMu sync.Mutex
			108	+	snapshotWorkerID uint64
			109	+	snapshotWorkerCtx context.Context
			110	+	snapshotWorkerCancel context.CancelFunc
			111	+	snapshotWorkerDone chan struct{}
			112	+	snapshotWorkerFlush chan struct{}
			113	+	snapshotWorkerFinalWake chan struct{}
			114	+	// snapshotFinalJobs coalesces terminal captures by session when the bounded
			115	+	// regular queue is full. It retains at most snapshotFinalQueueCapacity named
			116	+	// sessions, each with only its newest terminal state while the worker blocks.
			117	+	snapshotFinalJobs map[session]snapshotCapture
			118	+	snapshotFinalOrder []*session
			119	+	snapshotWorkerClosing bool
			120	+	snapshotWorkerInFlight *snapshotCapture
131		restoreDone chan struct{}	121		restoreDone chan struct{}
132		restoreOnce sync.Once	122		restoreOnce sync.Once
133		procCwd func(int) (string, error)	123		procCwd func(int) (string, error)

@@ -160,8 +150,14 @@ type Daemon struct {
160		done chan struct{}	150		done chan struct{}
161		doneOnce sync.Once	151		doneOnce sync.Once
162			152
163	-	sessWg sync.WaitGroup // PTY reader + scheduler goroutines	153	+	// sessWg owns attention animation, bar-script polling, CWD sampling,
			154	+	// snapshot save/restore, and floating-session launch goroutines.
			155	+	sessWg sync.WaitGroup
164		connWg sync.WaitGroup // per-connection handler goroutines	156		connWg sync.WaitGroup // per-connection handler goroutines
			157	+	// attachmentCleanupWg owns replacement transport closes and retired render
			158	+	// worker joins. Only connection handlers add work; Serve waits for those
			159	+	// handlers before joining this group, so no Add races its terminal Wait.
			160	+	attachmentCleanupWg sync.WaitGroup
165		}	161		}
166			162
167		type parkedAttachment struct {	163		type parkedAttachment struct {

@@ -194,6 +190,13 @@ func (s stoppedSession) same(other stoppedSession) bool {
194			190
195		type Option func(*Daemon)	191		type Option func(*Daemon)
196			192
			193	+	// WithRuntimeObserver accepts only a composition-root serialized observer.
			194	+	// The application owns its lifecycle; the daemon never creates or closes a
			195	+	// second reporting worker around it.
			196	+	func WithRuntimeObserver(observer ports.SerializedRuntimeObserver) Option {
			197	+	return func(d *Daemon) { d.runtimeObserver = observer }
			198	+	}
			199	+
197		// WithShell overrides the command (and its args) each session spawns. The	200		// WithShell overrides the command (and its args) each session spawns. The
198		// default is $SHELL (or /bin/sh) with no arguments; tests use this to run a	201		// default is $SHELL (or /bin/sh) with no arguments; tests use this to run a
199		// deterministic program.	202		// deterministic program.

@@ -312,6 +315,8 @@ func New(ptys ports.PTYFactory, clock ports.Clock, log *slog.Logger, opts ...Opt
312		done: make(chan struct{}),	315		done: make(chan struct{}),
313		restoreDone: make(chan struct{}),	316		restoreDone: make(chan struct{}),
314		animWake: make(chan struct{}, 1),	317		animWake: make(chan struct{}, 1),
			318	+	snapshotMarshal: snapcodec.Marshal,
			319	+	snapshotJobs: make(chan *snapshotCapture, snapshotQueueCapacity),
315		resumeParkGrace: defaultResumeParkGrace,	320		resumeParkGrace: defaultResumeParkGrace,
316		barScripts: &barScriptState{	321		barScripts: &barScriptState{
317		cfg: barConfigFromDomain(domain.Defaults().Bar),	322		cfg: barConfigFromDomain(domain.Defaults().Bar),

@@ -382,6 +387,7 @@ func (d *Daemon) Serve(ctx context.Context, l ports.Listener) error {
382		})	387		})
383		}	388		}
384		if d.snapsEnabled {	389		if d.snapsEnabled {
			390	+	d.startSnapshotEncodeWorker()
385		d.sessWg.Go(func() {	391		d.sessWg.Go(func() {
386		d.snapshotSaver(d.serveCtx)	392		d.snapshotSaver(d.serveCtx)
387		})	393		})

@@ -425,12 +431,17 @@ func (d *Daemon) Serve(ctx context.Context, l ports.Listener) error {
425		// remaining parked connections, so graceful notices are never raced by the	431		// remaining parked connections, so graceful notices are never raced by the
426		// force-close. Finally drain the conn handlers, run one defensive sweep in	432		// force-close. Finally drain the conn handlers, run one defensive sweep in
427		// case any ordering ever leaves a session behind, and join the session	433		// case any ordering ever leaves a session behind, and join the session
428	-	// goroutines (readers unblock via pty.Close, schedulers via ctx cancel).	434	+	// goroutines (readers unblock via pty.Close, coordinators via ctx cancel).
429		d.shutdownAll(ports.ReasonServerShutdown)	435		d.shutdownAll(ports.ReasonServerShutdown)
430		d.waitNotifies()	436		d.waitNotifies()
431		d.hardCancel()	437		d.hardCancel()
432		d.serveCancel()	438		d.serveCancel()
			439	+	// Wait for handlers before flushing snapshots: a handler that already
			440	+	// entered killSession may still submit its terminal capture after the
			441	+	// registry snapshot has been removed.
433		d.connWg.Wait()	442		d.connWg.Wait()
			443	+	d.attachmentCleanupWg.Wait()
			444	+	d.stopSnapshotEncodeWorker()
434		d.shutdownAll(ports.ReasonServerShutdown)	445		d.shutdownAll(ports.ReasonServerShutdown)
435		d.sessWg.Wait()	446		d.sessWg.Wait()
436		d.waitNotifies()	447		d.waitNotifies()

@@ -459,7 +470,9 @@ func (d *Daemon) shutdownAll(reason uint8) {
459		return	470		return
460		}	471		}
461		for _, s := range snapshot {	472		for _, s := range snapshot {
462	-	_ = d.killSession(s, reason, false)	473	+	if err := d.killSession(s, reason, false); err != nil {
			474	+	d.log.Error("closing session with unpersisted terminal state", "err", err)
			475	+	}
463		}	476		}
464		}	477		}
465			478

@@ -620,7 +633,23 @@ func (d *Daemon) handleHello(tr ports.Transport, f ports.Frame) {
620		return	633		return
621		}	634		}
622			635
623	-	if err := ac.send(frameWelcome(sess, ac)); err != nil {	636	+	rc := sess.renderCoordinator()
			637	+	lease := (*attachmentLease)(nil)
			638	+	if rc != nil {
			639	+	lease = rc.attachmentLease(ac)
			640	+	}
			641	+	expected := ac.transportSnapshot()
			642	+	if lease == nil \|\| expected.transport != tr {
			643	+	d.clientGone(sess, ac, tr, false)
			644	+	return
			645	+	}
			646	+	if err := ac.sendExpectedTransport(expected, frameWelcome(sess, ac)); err != nil {
			647	+	d.clientGone(sess, ac, tr, false)
			648	+	return
			649	+	}
			650	+	if !rc.markAttachmentReady(lease) {
			651	+	// The attachment was displaced or detached while Welcome was in flight;
			652	+	// never let this stale handshake emit an Output frame.
624		d.clientGone(sess, ac, tr, false)	653		d.clientGone(sess, ac, tr, false)
625		return	654		return
626		}	655		}

@@ -637,6 +666,23 @@ type protoErr struct {
637			666
638		func (e *protoErr) Error() string { return e.text }	667		func (e *protoErr) Error() string { return e.text }
639			668
			669	+	// finishAttach completes an attachment prepared while d.mu is held. It
			670	+	// publishes the terminal state before releasing d.mu, then queues replacement
			671	+	// teardown so obsolete worker joins never delay the new handshake.
			672	+	func (d Daemon) finishAttach(sess session, tr ports.Transport, sz domain.Size, term terminalEnv, h ports.Hello) *attachedClient {
			673	+	ac, old, cleanup := d.attachClientDeferred(sess, tr, sz, attachClientOptions{
			674	+	clientID: h.ClientID,
			675	+	resumeCapable: true,
			676	+	maxOutputInFlight: normalizeOutputWindow(h.MaxOutputInFlight),
			677	+	})
			678	+	sess.mu.Lock()
			679	+	sess.terminal = term
			680	+	sess.mu.Unlock()
			681	+	d.mu.Unlock()
			682	+	d.retireReplacedClient(old, cleanup)
			683	+	return ac
			684	+	}
			685	+
640		// route resolves a Hello to a session and a freshly attached client, creating	686		// route resolves a Hello to a session and a freshly attached client, creating
641		// the session for ephemeral/new intents.	687		// the session for ephemeral/new intents.
642		func (d Daemon) route(h ports.Hello, tr ports.Transport) (session, *attachedClient, error) {	688		func (d Daemon) route(h ports.Hello, tr ports.Transport) (session, *attachedClient, error) {

@@ -687,13 +733,7 @@ func (d Daemon) route(h ports.Hello, tr ports.Transport) (session, *attachedCl
687		}	733		}
688		}	734		}
689		d.purgeParkedForSessionLocked(sess)	735		d.purgeParkedForSessionLocked(sess)
690	-	ac, old := d.attachClient(sess, tr, sz, attachClientOptions{clientID: h.ClientID, resumeCapable: true, maxOutputInFlight: normalizeOutputWindow(h.MaxOutputInFlight)})	736	+	return sess, d.finishAttach(sess, tr, sz, term, h), nil
691	-	sess.mu.Lock()
692	-	sess.terminal = term
693	-	sess.mu.Unlock()
694	-	d.mu.Unlock()
695	-	d.detachReplacedClient(old)
696	-	return sess, ac, nil
697			737
698		case ports.IntentEphemeral:	738		case ports.IntentEphemeral:
699		name := d.allocEphemeralNameLocked()	739		name := d.allocEphemeralNameLocked()

@@ -703,13 +743,7 @@ func (d Daemon) route(h ports.Hello, tr ports.Transport) (session, *attachedCl
703		return nil, nil, err	743		return nil, nil, err
704		}	744		}
705		d.purgeParkedForSessionLocked(sess)	745		d.purgeParkedForSessionLocked(sess)
706	-	ac, old := d.attachClient(sess, tr, sz, attachClientOptions{clientID: h.ClientID, resumeCapable: true, maxOutputInFlight: normalizeOutputWindow(h.MaxOutputInFlight)})	746	+	return sess, d.finishAttach(sess, tr, sz, term, h), nil
707	-	sess.mu.Lock()
708	-	sess.terminal = term
709	-	sess.mu.Unlock()
710	-	d.mu.Unlock()
711	-	d.detachReplacedClient(old)
712	-	return sess, ac, nil
713			747
714		case ports.IntentNew:	748		case ports.IntentNew:
715		if h.Name == "" {	749		if h.Name == "" {

@@ -730,13 +764,7 @@ func (d Daemon) route(h ports.Hello, tr ports.Transport) (session, *attachedCl
730		return nil, nil, err	764		return nil, nil, err
731		}	765		}
732		d.purgeParkedForSessionLocked(sess)	766		d.purgeParkedForSessionLocked(sess)
733	-	ac, old := d.attachClient(sess, tr, sz, attachClientOptions{clientID: h.ClientID, resumeCapable: true, maxOutputInFlight: normalizeOutputWindow(h.MaxOutputInFlight)})	767	+	return sess, d.finishAttach(sess, tr, sz, term, h), nil
734	-	sess.mu.Lock()
735	-	sess.terminal = term
736	-	sess.mu.Unlock()
737	-	d.mu.Unlock()
738	-	d.detachReplacedClient(old)
739	-	return sess, ac, nil
740			768
741		case ports.IntentAttach:	769		case ports.IntentAttach:
742		sess := d.findByNameLocked(h.Name)	770		sess := d.findByNameLocked(h.Name)

@@ -755,13 +783,7 @@ func (d Daemon) route(h ports.Hello, tr ports.Transport) (session, *attachedCl
755		}	783		}
756		}	784		}
757		d.purgeParkedForSessionLocked(sess)	785		d.purgeParkedForSessionLocked(sess)
758	-	ac, old := d.attachClient(sess, tr, sz, attachClientOptions{clientID: h.ClientID, resumeCapable: true, maxOutputInFlight: normalizeOutputWindow(h.MaxOutputInFlight)})	786	+	return sess, d.finishAttach(sess, tr, sz, term, h), nil
759	-	sess.mu.Lock()
760	-	sess.terminal = term
761	-	sess.mu.Unlock()
762	-	d.mu.Unlock()
763	-	d.detachReplacedClient(old)
764	-	return sess, ac, nil
765			787
766		default:	788		default:
767		d.mu.Unlock()	789		d.mu.Unlock()

M internal/usecase/daemon/daemon_test.go +99 -5


@@ -142,7 +142,7 @@ func newBlockingOpenFactory(t testing.T, d Daemon) *blockingOpenFactory {
142		return f	142		return f
143		}	143		}
144			144
145	-	func (f *blockingOpenFactory) Open(string, []string, []string, string, domain.Size) (ports.PTY, error) {	145	+	func (f *blockingOpenFactory) Open(_ context.Context, _ string, _ []string, _ []string, _ string, _ domain.Size) (ports.PTY, error) {
146		<-f.release	146		<-f.release
147		return nil, io.ErrClosedPipe	147		return nil, io.ErrClosedPipe
148		}	148		}

@@ -154,8 +154,8 @@ func newFactory(t testing.T, p ports.PTY) portsmocks.MockPTYFactory {
154		t.Helper()	154		t.Helper()
155		f := portsmocks.NewMockPTYFactory(t)	155		f := portsmocks.NewMockPTYFactory(t)
156		var normal domain.Size	156		var normal domain.Size
157	-	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything).RunAndReturn(	157	+	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything).RunAndReturn(
158	-	func(_ string, _ []string, _ []string, _ string, sz domain.Size) (ports.PTY, error) {	158	+	func(_ context.Context, _ string, _ []string, _ []string, _ string, sz domain.Size) (ports.PTY, error) {
159		if !normal.Valid() {	159		if !normal.Valid() {
160		normal = sz	160		normal = sz
161		}	161		}

@@ -174,8 +174,8 @@ func newFactorySeq(t testing.T, ptys ...ports.PTY) portsmocks.MockPTYFactory {
174		var mu sync.Mutex	174		var mu sync.Mutex
175		var normal domain.Size	175		var normal domain.Size
176		next := 0	176		next := 0
177	-	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything).RunAndReturn(	177	+	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything).RunAndReturn(
178	-	func(_ string, _ []string, _ []string, _ string, sz domain.Size) (ports.PTY, error) {	178	+	func(_ context.Context, _ string, _ []string, _ []string, _ string, sz domain.Size) (ports.PTY, error) {
179		mu.Lock()	179		mu.Lock()
180		defer mu.Unlock()	180		defer mu.Unlock()
181		if !normal.Valid() {	181		if !normal.Valid() {

@@ -286,6 +286,54 @@ func listSessions(t testing.T, d Daemon) ports.Sessions {
286		return sessions	286		return sessions
287		}	287		}
288			288
			289	+	// welcomeBlockingTransport holds the route-level handshake precisely between
			290	+	// attachment publication and Welcome acceptance. It uses the generated
			291	+	// transport mock while channels make both sides of the wire deterministic.
			292	+	type welcomeBlockingTransport struct {
			293	+	tr *portsmocks.MockTransport
			294	+	sends chan ports.Frame
			295	+	welcomeEntered chan struct{}
			296	+	releaseWelcome chan struct{}
			297	+	recvDone chan struct{}
			298	+	releaseOnce sync.Once
			299	+	closeOnce sync.Once
			300	+	}
			301	+
			302	+	func newWelcomeBlockingTransport(t testing.T) welcomeBlockingTransport {
			303	+	t.Helper()
			304	+	b := &welcomeBlockingTransport{
			305	+	sends: make(chan ports.Frame, 8),
			306	+	welcomeEntered: make(chan struct{}),
			307	+	releaseWelcome: make(chan struct{}),
			308	+	recvDone: make(chan struct{}),
			309	+	}
			310	+	b.tr = portsmocks.NewMockTransport(t)
			311	+	b.tr.EXPECT().Send(mock.Anything).RunAndReturn(func(f ports.Frame) error {
			312	+	b.sends <- f
			313	+	if f.Type == ports.MsgWelcome {
			314	+	close(b.welcomeEntered)
			315	+	<-b.releaseWelcome
			316	+	}
			317	+	return nil
			318	+	}).Maybe()
			319	+	b.tr.EXPECT().Recv().RunAndReturn(func() (ports.Frame, error) {
			320	+	<-b.recvDone
			321	+	return ports.Frame{}, io.EOF
			322	+	}).Maybe()
			323	+	b.tr.EXPECT().Close().Return(nil).Maybe()
			324	+	t.Cleanup(b.finish)
			325	+	return b
			326	+	}
			327	+
			328	+	func (b *welcomeBlockingTransport) release() {
			329	+	b.releaseOnce.Do(func() { close(b.releaseWelcome) })
			330	+	}
			331	+
			332	+	func (b *welcomeBlockingTransport) finish() {
			333	+	b.release()
			334	+	b.closeOnce.Do(func() { close(b.recvDone) })
			335	+	}
			336	+
289		func newCapturingTransport(t testing.TB) (*portsmocks.MockTransport, chan ports.Frame) {	337		func newCapturingTransport(t testing.TB) (*portsmocks.MockTransport, chan ports.Frame) {
290		t.Helper()	338		t.Helper()
291		tr := portsmocks.NewMockTransport(t)	339		tr := portsmocks.NewMockTransport(t)

@@ -395,6 +443,52 @@ func (r realTimer) C() <-chan time.Time { return r.t.C }
395		func (r realTimer) Reset(d time.Duration) bool { return r.t.Reset(d) }	443		func (r realTimer) Reset(d time.Duration) bool { return r.t.Reset(d) }
396		func (r realTimer) Stop() bool { return r.t.Stop() }	444		func (r realTimer) Stop() bool { return r.t.Stop() }
397			445
			446	+	func TestHandleHelloDefersFreshOutputUntilWelcome(t *testing.T) {
			447	+	pty, releasePTY := newBlockingPTY(t)
			448	+	defer releasePTY()
			449	+	clock := newCoordinatorMockClock(t, 4)
			450	+	d := newTestDaemon(t, newFactory(t, pty), clock.clock)
			451	+	tr := newWelcomeBlockingTransport(t)
			452	+	done := make(chan struct{})
			453	+	go func() {
			454	+	d.handleHello(tr.tr, mustHello(ports.IntentNew, "welcome-gate", domain.Size{Cols: 80, Rows: 24}))
			455	+	close(done)
			456	+	}()
			457	+
			458	+	awaitTestCompletion(t, tr.welcomeEntered, "timed out waiting for Welcome send")
			459	+	welcome := awaitFrame(t, tr.sends, ports.MsgWelcome)
			460	+	require.Equal(t, ports.MsgWelcome, welcome.Type)
			461	+	sess := firstSession(d)
			462	+	sess.mu.Lock()
			463	+	ac := sess.client
			464	+	sess.mu.Unlock()
			465	+	require.NotNil(t, ac, "route must publish the attachment before Welcome returns")
			466	+
			467	+	// Exercise the real producer fan-in after publication. The deadline worker
			468	+	// has completed before asserting the gate, so this is not a timing race.
			469	+	d.invalidateRender(sess, ac, true, "welcome-gate-test")
			470	+	// Attach-time theme setup may have armed an older bulk timer. The fresh
			471	+	// urgent producer replaces it, so advance the newest published deadline.
			472	+	timer := awaitLatestCoordinatorTimer(t, clock)
			473	+	rc := sess.renderCoordinator()
			474	+	rc.mu.Lock()
			475	+	workerDone := rc.normalLane.token.done
			476	+	rc.mu.Unlock()
			477	+	require.NotNil(t, workerDone)
			478	+	timer.ch <- time.Time{}
			479	+	awaitTestCompletion(t, workerDone, "coordinator deadline worker did not complete")
			480	+	requireNoCoordinatorOutputFrame(t, tr.sends)
			481	+
			482	+	tr.release()
			483	+	output := awaitFrame(t, tr.sends, ports.MsgOutput)
			484	+	first, err := ports.UnmarshalOutput(output.Payload)
			485	+	require.NoError(t, err)
			486	+	require.Zero(t, first.BaseStateNum, "the first post-Welcome frame is full")
			487	+	tr.finish()
			488	+	awaitTestCompletion(t, done, "timed out waiting for Welcome handler completion")
			489	+	requireNoCoordinatorOutputFrame(t, tr.sends)
			490	+	}
			491	+
398		// TestServeReturnsDespiteWedgedClientOnShutdown: killSession's teardown must	492		// TestServeReturnsDespiteWedgedClientOnShutdown: killSession's teardown must
399		// not be gated behind the Detached notice. The client transport wedges every	493		// not be gated behind the Detached notice. The client transport wedges every
400		// MsgDetached send until the transport is closed (mirroring a full kernel	494		// MsgDetached send until the transport is closed (mirroring a full kernel

M internal/usecase/daemon/floating.go +109 -19

 // activateTab performs the work associated with making a tab the destination.
 // It is deliberately idempotent: every tab transition can use this single
 // hook, while inactive restored tabs remain cold until actually selected.
-func (d *Daemon) activateTab(sess *session, tb *tab) {
+func (d *Daemon) activateTab(sess *session, tb *tab) bool {
+	return d.activateTabAfterResize(sess, tb, false)
+}
+// activateTabAfterResize retains tab activation's warmup work while avoiding a
+// second resize when a synchronous outer resize request was already accepted.
+func (d *Daemon) activateTabAfterResize(sess *session, tb *tab, outerResizeAccepted bool) bool {
 	if d == nil || sess == nil || tb == nil {
-		return
+		return false
 	}
 	// A headless session has no actual terminal destination. Deferring warmup
 	// until firstPaint keeps restored tabs cold and avoids launching children
 	sess.mu.Unlock()
 	d.exitCopyMode(ac)
 	if ac == nil {
-		return
+		return false
 	}
 	d.ensureFloatingWarm(sess, tb)
-	d.resizeInstalledFloating(tb)
+	if outerResizeAccepted {
+		return false
+	}
+	tb.mu.Lock()
+	hasFloating := tb.floating.pane != nil
+	size := domain.Size{Cols: tb.size.Cols, Rows: tb.size.Rows + 2}
+	tb.mu.Unlock()
+	if hasFloating {
+		return d.requestTransactionalResize(sess, ac, size, true)
+	}
+	return false
 }
 // toggleFloating changes only this tab's slot. Opening an uninitialized slot
 	}
 	cfg := d.currentFloatingConfig()
 	tb.mu.Lock()
+	wasVisible := tb.floating.state == floatingVisible
 	start, generation := tb.toggleFloatingLocked()
 	// desiredVisible is meaningful only while a launch is warming. Installed
 	// slots derive their resize and paint work from the resulting slot state.
 	visible := tb.floating.state == floatingVisible
+	hidden := wasVisible && tb.floating.state == floatingHidden
 	tb.mu.Unlock()
 	if start {
 		d.launchFloating(sess, tb, cfg, generation, true)
 		if ac != nil {
-			d.paint(sess, ac, true)
+			d.invalidateRender(sess, ac, true, "floating.go")
 		}
 		return nil
 	}
-	if visible {
+	if hidden {
-		d.resizeActiveFloating(tb)
+		d.invalidateRender(sess, ac, true, "floating.go")
+		return nil
 	}
-	if ac != nil {
+	if visible {
-		d.paint(sess, ac, true)
+		tb.mu.Lock()
+		size := domain.Size{Cols: tb.size.Cols, Rows: tb.size.Rows + 2}
+		tb.mu.Unlock()
+		d.requestTransactionalResize(sess, ac, size, true)
 	}
 	return nil
 }
 // floatingLaunchSpec is an immutable snapshot of one launch. It is built
 // before the worker starts so config reloads cannot affect that launch.
+// floatingLaunch owns one worker from queueing through installation or stale
+// result cleanup. Its done channel lets session teardown reap an in-flight
+// Open before returning.
+type floatingLaunch struct {
+	done chan struct{}
+}
+func (sess *session) registerFloatingLaunch() (*floatingLaunch, bool) {
+	launch := &floatingLaunch{done: make(chan struct{})}
+	sess.floatingLaunchMu.Lock()
+	defer sess.floatingLaunchMu.Unlock()
+	if sess.floatingLaunchStopping {
+		return nil, false
+	}
+	if sess.floatingLaunches == nil {
+		sess.floatingLaunches = make(map[*floatingLaunch]struct{})
+	}
+	sess.floatingLaunches[launch] = struct{}{}
+	return launch, true
+}
+func (sess *session) finishFloatingLaunch(launch *floatingLaunch) {
+	sess.floatingLaunchMu.Lock()
+	delete(sess.floatingLaunches, launch)
+	close(launch.done)
+	sess.floatingLaunchMu.Unlock()
+}
+func (sess *session) stopFloatingLaunches() {
+	sess.floatingLaunchMu.Lock()
+	sess.floatingLaunchStopping = true
+	sess.floatingLaunchMu.Unlock()
+}
 type floatingLaunchSpec struct {
 	sessionName  string
 	cwd          string
 }
 // launchFloating snapshots launch inputs, then accounts for a worker through
-// Open, pane initialization, publication, and reader/scheduler startup.
+// Open, pane initialization, publication, and reader/coordinator startup.
 func (d *Daemon) launchFloating(sess *session, tb *tab, cfg domain.FloatingConfig, generation uint64, userOpen bool) {
 	spec, err := d.newFloatingLaunchSpec(sess, tb, cfg, userOpen)
 	if err != nil {
 		d.failFloatingLaunch(tb, generation, userOpen, spec.sessionName, err)
 		return
 	}
+	launch, ok := sess.registerFloatingLaunch()
+	if !ok {
+		d.failFloatingLaunch(tb, generation, userOpen, spec.sessionName, context.Canceled)
+		return
+	}
 	// Count the launch before its goroutine starts. It remains counted through
-	// install, including the reader/scheduler Adds, so shutdown cannot return
+	// install, including the reader/coordinator Adds, so shutdown cannot return
 	// before a late Open completion has either been rejected or fully joined.
-	d.sessWg.Go(func() { d.openAndInstallFloating(sess, tb, spec, generation) })
+	d.sessWg.Go(func() {
+		defer sess.finishFloatingLaunch(launch)
+		d.openAndInstallFloating(sess, tb, spec, generation)
+	})
 }
 func (d *Daemon) newFloatingLaunchSpec(sess *session, tb *tab, cfg domain.FloatingConfig, userOpen bool) (floatingLaunchSpec, error) {
 // openAndInstallFloating runs entirely in the launch worker. No PTY operation
 // occurs under tb.mu; the pane is fully initialized before publication.
 func (d *Daemon) openAndInstallFloating(sess *session, tb *tab, spec floatingLaunchSpec, generation uint64) {
-	pty, err := d.ptys.Open(spec.command, spec.args, spec.env, spec.cwd, spec.size)
+	// Open is context-aware and is intentionally called without any daemon,
+	// session, tab, or launch-ownership lock. Cancellation makes teardown
+	// bounded even if the adapter is waiting to create the child.
+	if err := spec.parentCtx.Err(); err != nil {
+		return
+	}
+	openCtx, cancelOpen := context.WithCancel(spec.parentCtx)
+	stopSessionCancel := context.AfterFunc(sess.ctx, cancelOpen)
+	ownedByPane := false
+	defer func() {
+		if !ownedByPane {
+			stopSessionCancel()
+			cancelOpen()
+		}
+	}()
+	pty, err := d.ptys.Open(openCtx, spec.command, spec.args, spec.env, spec.cwd, spec.size)
 	if err != nil {
 		d.failFloatingLaunch(tb, generation, spec.userOpen, spec.sessionName, err)
 		return
 	p.rect = spec.geometry.Inner
 	p.popupGeometry = spec.geometry
 	p.title.displayFallback = spec.fallback
-	p.ctx, p.cancel = context.WithCancel(spec.parentCtx)
+	// CommandContext owns the child through openCtx, so retain both this context
+	// and its cancellation until the installed pane is reaped or torn down.
+	p.ctx = openCtx
+	p.cancel = func() {
+		stopSessionCancel()
+		cancelOpen()
+	}
+	ownedByPane = true
 	// The reader may run as soon as install returns; make its exit policy
 	// immutable before publishing the pane to the slot.
 	p.onExit = func() { d.reapFloating(sess, tb, p, generation) }
 	d.reapplyThemeSession(sess)
 	d.startPaneGoroutines(sess, tb, p)
 	if visible {
-		d.resizeActiveFloating(tb)
 		sess.mu.Lock()
 		ac := sess.client
 		sess.mu.Unlock()
-		if ac != nil {
+		tb.mu.Lock()
-			d.paint(sess, ac, true)
+		size := domain.Size{Cols: tb.size.Cols, Rows: tb.size.Rows + 2}
-		}
+		tb.mu.Unlock()
+		d.requestTransactionalResize(sess, ac, size, true)
 	}
 }
 	sess.mu.Lock()
 	ac := sess.client
 	sess.mu.Unlock()
+	if ac != nil {
+		ac.pruneCaptureFrames(p)
+	}
 	copyCleared := ac != nil && ac.overlays.clearCopyModeForPane(p)
 	if ac != nil && (visible || copyCleared) {
-		d.paint(sess, ac, true)
+		d.invalidateRender(sess, ac, true, "floating.go")
 	}
 }
 	tb.mu.Unlock()
 	if ac != nil {
 		ac.overlays.clearCopyModeForPane(p)
+		ac.pruneCaptureFrames(p)
 	}
 	closeFloatingPane(p)
 }

A internal/usecase/daemon/floating_open_test.go +299 -0

@@ -0,0 +1,299 @@
	1	+	package daemon
	2	+
	3	+	import (
	4	+	"context"
	5	+	"sync"
	6	+	"testing"
	7	+	"time"
	8	+
	9	+	"github.com/stretchr/testify/mock"
	10	+	"github.com/stretchr/testify/require"
	11	+
	12	+	"github.com/bnema/vev/internal/domain"
	13	+	"github.com/bnema/vev/internal/ports"
	14	+	portsmocks "github.com/bnema/vev/internal/ports/mocks"
	15	+	)
	16	+
	17	+	type contextAwareFloatingPTY struct {
	18	+	ctx context.Context
	19	+	readerStarted chan struct{}
	20	+	closed chan struct{}
	21	+	readerOnce sync.Once
	22	+	closeOnce sync.Once
	23	+	closeCount int
	24	+	}
	25	+
	26	+	func newContextAwareFloatingPTY() *contextAwareFloatingPTY {
	27	+	return &contextAwareFloatingPTY{
	28	+	readerStarted: make(chan struct{}),
	29	+	closed: make(chan struct{}),
	30	+	}
	31	+	}
	32	+
	33	+	func (p *contextAwareFloatingPTY) Read([]byte) (int, error) {
	34	+	p.readerOnce.Do(func() { close(p.readerStarted) })
	35	+	<-p.ctx.Done()
	36	+	return 0, p.ctx.Err()
	37	+	}
	38	+
	39	+	func (*contextAwareFloatingPTY) Write(b []byte) (int, error) { return len(b), nil }
	40	+	func (*contextAwareFloatingPTY) Resize(domain.Size) error { return nil }
	41	+	func (*contextAwareFloatingPTY) Pid() int { return 1 }
	42	+	func (*contextAwareFloatingPTY) ForegroundPgid() (int, error) {
	43	+	return 1, nil
	44	+	}
	45	+
	46	+	func (p *contextAwareFloatingPTY) Close() error {
	47	+	p.closeOnce.Do(func() {
	48	+	p.closeCount++
	49	+	close(p.closed)
	50	+	})
	51	+	return nil
	52	+	}
	53	+
	54	+	type contextAwareFloatingFactory struct {
	55	+	pty *contextAwareFloatingPTY
	56	+	opened chan context.Context
	57	+	err error
	58	+	onOpen func()
	59	+	}
	60	+
	61	+	func (f *contextAwareFloatingFactory) Open(ctx context.Context, _ string, _ []string, _ []string, _ string, _ domain.Size) (ports.PTY, error) {
	62	+	if f.pty != nil {
	63	+	f.pty.ctx = ctx
	64	+	}
	65	+	f.opened <- ctx
	66	+	if f.onOpen != nil {
	67	+	f.onOpen()
	68	+	}
	69	+	return f.pty, f.err
	70	+	}
	71	+
	72	+	func TestFloatingSuccessfulOpenTransfersContextOwnershipToPane(t *testing.T) {
	73	+	teardowns := []struct {
	74	+	name string
	75	+	teardown func(Daemon, session, *tab)
	76	+	}{
	77	+	{
	78	+	name: "floating pane",
	79	+	teardown: func(d Daemon, _ session, tb *tab) {
	80	+	d.teardownFloating(tb, nil)
	81	+	},
	82	+	},
	83	+	{
	84	+	name: "tab",
	85	+	teardown: func(_ Daemon, _ session, tb *tab) {
	86	+	tb.closeAllPanes()
	87	+	},
	88	+	},
	89	+	{
	90	+	name: "session",
	91	+	teardown: func(d Daemon, sess session, _ *tab) {
	92	+	require.NoError(t, d.killSession(sess, ports.ReasonSessionKilled, false))
	93	+	},
	94	+	},
	95	+	}
	96	+	for _, tt := range teardowns {
	97	+	t.Run(tt.name, func(t *testing.T) {
	98	+	pty := newContextAwareFloatingPTY()
	99	+	factory := &contextAwareFloatingFactory{pty: pty, opened: make(chan context.Context, 1)}
	100	+	d := newTestDaemon(t, factory, stubClock{})
	101	+	tb := newFloatingTestTab(t)
	102	+	tabCtx, cancelTab := context.WithCancel(t.Context())
	103	+	tb.ctx, tb.cancel = tabCtx, cancelTab
	104	+	sessCtx, cancelSession := context.WithCancel(t.Context())
	105	+	sess := &session{id: "context-" + domain.SessionID(tt.name), name: "work", tabs: []*tab{tb}, ctx: sessCtx, cancel: cancelSession}
	106	+	d.mu.Lock()
	107	+	d.sessions[sess.id] = sess
	108	+	d.mu.Unlock()
	109	+	tb.mu.Lock()
	110	+	generation := tb.beginFloatingWarmLocked(true)
	111	+	tb.mu.Unlock()
	112	+
	113	+	d.openAndInstallFloating(sess, tb, floatingLaunchSpec{
	114	+	sessionName: "work",
	115	+	size: domain.Size{Cols: 20, Rows: 8},
	116	+	geometry: floatingGeometry{Inner: domain.Rect{Width: 20, Height: 8}},
	117	+	paneStableID: "p_floating",
	118	+	parentCtx: tabCtx,
	119	+	}, generation)
	120	+
	121	+	openCtx := <-factory.opened
	122	+	require.NoError(t, openCtx.Err(), "the PTY context must outlive the launch worker")
	123	+	tb.mu.Lock()
	124	+	floating := tb.floating.pane
	125	+	tb.mu.Unlock()
	126	+	require.NotNil(t, floating)
	127	+	sess.mu.Lock()
	128	+	sess.client = &attachedClient{}
	129	+	sess.mu.Unlock()
	130	+	require.True(t, d.paneRenderable(sess, tb, floating), "installed visible popup must remain renderable")
	131	+	sess.mu.Lock()
	132	+	sess.client = nil
	133	+	sess.mu.Unlock()
	134	+	select {
	135	+	case <-pty.readerStarted:
	136	+	case <-time.After(time.Second):
	137	+	t.Fatal("floating reader did not start")
	138	+	}
	139	+
	140	+	tt.teardown(d, sess, tb)
	141	+	select {
	142	+	case <-openCtx.Done():
	143	+	case <-time.After(time.Second):
	144	+	t.Fatal("floating PTY context was not cancelled by teardown")
	145	+	}
	146	+	select {
	147	+	case <-pty.closed:
	148	+	case <-time.After(time.Second):
	149	+	t.Fatal("floating PTY was not closed by teardown")
	150	+	}
	151	+	d.sessWg.Wait()
	152	+	require.Equal(t, 1, pty.closeCount, "floating PTY must close exactly once")
	153	+	})
	154	+	}
	155	+	}
	156	+
	157	+	func TestFloatingFailedAndStaleOpenCancelContext(t *testing.T) {
	158	+	tests := []struct {
	159	+	name string
	160	+	setup func(tab, contextAwareFloatingFactory)
	161	+	}{
	162	+	{
	163	+	name: "failed Open",
	164	+	setup: func(_ tab, factory contextAwareFloatingFactory) {
	165	+	factory.err = context.Canceled
	166	+	},
	167	+	},
	168	+	{
	169	+	name: "stale install",
	170	+	setup: func(tb tab, factory contextAwareFloatingFactory) {
	171	+	factory.onOpen = func() {
	172	+	tb.mu.Lock()
	173	+	tb.takeFloatingLocked()
	174	+	tb.mu.Unlock()
	175	+	}
	176	+	},
	177	+	},
	178	+	}
	179	+	for _, tt := range tests {
	180	+	t.Run(tt.name, func(t *testing.T) {
	181	+	pty := newContextAwareFloatingPTY()
	182	+	factory := &contextAwareFloatingFactory{pty: pty, opened: make(chan context.Context, 1)}
	183	+	d := newTestDaemon(t, factory, stubClock{})
	184	+	tb := newFloatingTestTab(t)
	185	+	sess := &session{name: "work", tabs: []*tab{tb}, ctx: t.Context()}
	186	+	tb.mu.Lock()
	187	+	generation := tb.beginFloatingWarmLocked(true)
	188	+	tb.mu.Unlock()
	189	+	tt.setup(tb, factory)
	190	+
	191	+	d.openAndInstallFloating(sess, tb, floatingLaunchSpec{
	192	+	sessionName: "work",
	193	+	size: domain.Size{Cols: 20, Rows: 8},
	194	+	geometry: floatingGeometry{Inner: domain.Rect{Width: 20, Height: 8}},
	195	+	paneStableID: "p_floating",
	196	+	parentCtx: tb.ctx,
	197	+	}, generation)
	198	+
	199	+	openCtx := <-factory.opened
	200	+	select {
	201	+	case <-openCtx.Done():
	202	+	case <-time.After(time.Second):
	203	+	t.Fatal("failed or stale floating Open retained its context")
	204	+	}
	205	+	tb.mu.Lock()
	206	+	require.Nil(t, tb.floating.pane)
	207	+	tb.mu.Unlock()
	208	+	if tt.name == "stale install" {
	209	+	select {
	210	+	case <-pty.closed:
	211	+	case <-time.After(time.Second):
	212	+	t.Fatal("stale floating PTY was not closed")
	213	+	}
	214	+	require.Equal(t, 1, pty.closeCount)
	215	+	}
	216	+	})
	217	+	}
	218	+	}
	219	+
	220	+	func TestFloatingOpenCancellationBoundsSessionTeardownAndClosesLatePTY(t *testing.T) {
	221	+	factory := portsmocks.NewMockPTYFactory(t)
	222	+	opened := make(chan struct{})
	223	+	latePTY := portsmocks.NewMockPTY(t)
	224	+	closed := make(chan struct{})
	225	+	latePTY.EXPECT().Close().RunAndReturn(func() error { close(closed); return nil }).Once()
	226	+	factory.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything).RunAndReturn(
	227	+	func(ctx context.Context, _ string, _ []string, _ []string, _ string, _ domain.Size) (ports.PTY, error) {
	228	+	close(opened)
	229	+	<-ctx.Done()
	230	+	return latePTY, nil
	231	+	}).Once()
	232	+	d := newTestDaemon(t, factory, stubClock{})
	233	+	tb := newFloatingTestTab(t)
	234	+	ctx, cancel := context.WithCancel(t.Context())
	235	+	sess := &session{id: "blocked-open", name: "work", tabs: []*tab{tb}, ctx: ctx, cancel: cancel}
	236	+	d.mu.Lock()
	237	+	d.sessions[sess.id] = sess
	238	+	d.mu.Unlock()
	239	+	d.startFloating(sess, tb, true)
	240	+	select {
	241	+	case <-opened:
	242	+	case <-time.After(time.Second):
	243	+	t.Fatal("floating Open did not begin")
	244	+	}
	245	+
	246	+	killed := make(chan error, 1)
	247	+	go func() { killed <- d.killSession(sess, ports.ReasonSessionKilled, false) }()
	248	+	select {
	249	+	case err := <-killed:
	250	+	require.NoError(t, err)
	251	+	case <-time.After(time.Second):
	252	+	t.Fatal("session teardown waited for blocked Open")
	253	+	}
	254	+	select {
	255	+	case <-closed:
	256	+	case <-time.After(time.Second):
	257	+	t.Fatal("late Open PTY was not closed")
	258	+	}
	259	+	tb.mu.Lock()
	260	+	require.Nil(t, tb.floating.pane, "late PTY must not be installed")
	261	+	tb.mu.Unlock()
	262	+	}
	263	+
	264	+	func TestFloatingOpensFromUnrelatedSessionsAreConcurrent(t *testing.T) {
	265	+	factory := portsmocks.NewMockPTYFactory(t)
	266	+	entered := make(chan struct{}, 2)
	267	+	release := make(chan struct{})
	268	+	var once sync.Once
	269	+	factory.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything).RunAndReturn(
	270	+	func(_ context.Context, _ string, _ []string, _ []string, _ string, _ domain.Size) (ports.PTY, error) {
	271	+	entered <- struct{}{}
	272	+	<-release
	273	+	return newQuietPTY(), nil
	274	+	}).Twice()
	275	+	d := newTestDaemon(t, factory, stubClock{})
	276	+	tabs := make([]*tab, 0, 2)
	277	+	cancels := make([]context.CancelFunc, 0, 2)
	278	+	for _, id := range []domain.SessionID{"one", "two"} {
	279	+	tb := newFloatingTestTab(t)
	280	+	ctx, cancel := context.WithCancel(t.Context())
	281	+	sess := &session{id: id, name: string(id), tabs: []*tab{tb}, ctx: ctx, cancel: cancel}
	282	+	tabs = append(tabs, tb)
	283	+	cancels = append(cancels, cancel)
	284	+	d.startFloating(sess, tb, false)
	285	+	}
	286	+	for range 2 {
	287	+	select {
	288	+	case <-entered:
	289	+	case <-time.After(time.Second):
	290	+	t.Fatal("unrelated floating launch was serialized")
	291	+	}
	292	+	}
	293	+	once.Do(func() { close(release) })
	294	+	for i, cancel := range cancels {
	295	+	cancel()
	296	+	d.teardownFloating(tabs[i], nil)
	297	+	}
	298	+	d.sessWg.Wait()
	299	+	}

M internal/usecase/daemon/floating_render.go +12 -113


@@ -85,60 +85,26 @@ func calculateContentFloatingGeometry(content domain.Size, cfg domain.FloatingCo
85		}	85		}
86		}	86		}
87			87
88	-	// composeFloatingFrame applies the popup to a copy of the normal composed	88	+	func composeCapturedFloatingFrame(base renderer.Frame, baseDamage []renderer.Damage, floating capturedFloatingRenderState, content domain.Rect, layoutSnap capturedTabLayout, theme themeui.Theme, cache composeCacheInput, full bool) (renderer.Frame, []renderer.Damage) {
89	-	// destination. In particular it never writes the VT screen frame: that frame	89	+	frame := base.Clone()
90	-	// is shared with the PTY reader and is the source for future renders.	90	+	legacyLayout := tabLayoutSnapshot{placements: layoutSnap.placements, area: layoutSnap.area, focus: layoutSnap.focus, ok: layoutSnap.valid}
91	-	func composeFloatingFrame(base renderer.Frame, baseDamage []renderer.Damage, p pane, generation uint64, content domain.Rect, desired floatingGeometry, layoutSnap tabLayoutSnapshot, theme themeui.Theme, cache composedFrameCache, full bool) (renderer.Frame, []renderer.Damage, floatingGeometry) {	91	+	(overlayBackdrop{DimPaneContents: true}).apply(frame, content, legacyLayout, theme)
92	-	var frame renderer.Frame	92	+	geometry := floating.geometry.translate(content.X, content.Y)
93	-	if cache != nil && cache.floatingFrame.Width == base.Width && cache.floatingFrame.Height == base.Height {	93	+	blitPaneFrame(frame, geometry.Inner, floating.pane.frame, false, theme)
94	-	frame = cache.floatingFrame
95	-	for y := 0; y < base.Height; y++ {
96	-	copy(frame.Row(y), base.Row(y))
97	-	}
98	-	} else {
99	-	frame = base.Clone()
100	-	if cache != nil {
101	-	cache.floatingFrame = frame
102	-	}
103	-	}
104	-	(overlayBackdrop{DimPaneContents: true}).apply(frame, content, layoutSnap, theme)
105	-	// The PTY reader mutates both the frame and its damage under p.mu. Keep
106	-	// every read of them, including the blit, in the same critical section.
107	-	// In particular, Frame is not safe to retain as an alias after unlocking.
108	-	p.mu.Lock()
109	-	// Committed popup geometry is content-relative; translate it exactly once
110	-	// while holding the same lock that protects the committed state and frame.
111	-	geometry := p.committedFloatingGeometryLocked(desired).translate(content.X, content.Y)
112	-	title := p.displayTitleLocked()
113	-	titleGeneration := p.title.generation
114	-	screenDamage := p.screen.Damage()
115	-	blitPaneFrame(frame, geometry.Inner, p.screen.Frame, false, theme)
116		damage := append([]renderer.Damage(nil), baseDamage...)	94		damage := append([]renderer.Damage(nil), baseDamage...)
117	-	for _, d := range screenDamage {	95	+	for _, d := range floating.pane.damage {
118		damage = append(damage, translatePaneDamage(d, geometry.Inner, content)...)	96		damage = append(damage, translatePaneDamage(d, geometry.Inner, content)...)
119		}	97		}
120	-	p.screen.ClearDamage()	98	+	popupChanged := !cache.valid \|\| cache.floatingGeneration != floating.generation \|\| cache.floatingGeometry != geometry
121	-	p.mu.Unlock()	99	+	titleChanged := popupChanged \|\| cache.floatingTitleGeneration != floating.titleGeneration
122	-		100	+	drawFloatingBorder(frame, geometry.Bounds, floating.title, newThemeStyles(theme).border)
123	-	// The committed geometry, rather than the requested config geometry, is
124	-	// what was actually rendered. Track it so movement (including a same-size
125	-	// resize) cannot reuse a cached popup at its former position.
126	-	popupChanged := cache == nil \|\| cache.floating != p \|\| cache.floatingGeneration != generation \|\| cache.floatingGeometry != geometry
127	-	titleChanged := popupChanged \|\| cache == nil \|\| cache.floatingTitleGeneration != titleGeneration
128	-	drawFloatingBorder(frame, geometry.Bounds, title, newThemeStyles(theme).border)
129	-	if cache != nil {
130	-	cache.floating = p
131	-	cache.floatingGeneration = generation
132	-	cache.floatingGeometry = geometry
133	-	cache.floatingTitleGeneration = titleGeneration
134	-	}
135		if full \|\| popupChanged {	101		if full \|\| popupChanged {
136	-	return frame, []renderer.Damage{renderer.FullRedraw()}, geometry	102	+	return frame, []renderer.Damage{renderer.FullRedraw()}
137		}	103		}
138		if titleChanged && geometry.Bounds.Height >= 3 {	104		if titleChanged && geometry.Bounds.Height >= 3 {
139		damage = append(damage, renderer.Damage{Kind: renderer.DamageText, X: geometry.Bounds.X, Y: geometry.Bounds.Y, Width: geometry.Bounds.Width, Height: 1})	105		damage = append(damage, renderer.Damage{Kind: renderer.DamageText, X: geometry.Bounds.X, Y: geometry.Bounds.Y, Width: geometry.Bounds.Width, Height: 1})
140		}	106		}
141	-	return frame, damage, geometry	107	+	return frame, damage
142		}	108		}
143			109
144		func drawFloatingBorder(frame renderer.Frame, bounds domain.Rect, title string, style renderer.Style) {	110		func drawFloatingBorder(frame renderer.Frame, bounds domain.Rect, title string, style renderer.Style) {

@@ -167,70 +133,3 @@ func drawFloatingBorder(frame renderer.Frame, bounds domain.Rect, title string,
167		ui.DrawText(frame, bounds.X+2, bounds.Y, bounds.X+bounds.Width-2, title, style)	133		ui.DrawText(frame, bounds.X+2, bounds.Y, bounds.X+bounds.Width-2, title, style)
168		}	134		}
169		}	135		}
170	-
171	-	// resizeFloatingPane serializes PTY resizes for a floating pane. PTY.Resize
172	-	// is intentionally outside pane.mu; failed resizes do not publish any state.
173	-	func (d Daemon) resizeFloatingPane(p pane, geometry floatingGeometry) bool {
174	-	inner := geometry.Inner
175	-	if p == nil \|\| !geometry.valid() {
176	-	return false
177	-	}
178	-	p.resizeMu.Lock()
179	-	defer p.resizeMu.Unlock()
180	-
181	-	p.mu.Lock()
182	-	old := p.rect
183	-	pty := p.pty
184	-	p.mu.Unlock()
185	-	if old.Width == inner.Width && old.Height == inner.Height {
186	-	// Position and outer bounds can change while the terminal dimensions do
187	-	// not. Commit that geometry without needlessly resetting the PTY/screen.
188	-	p.mu.Lock()
189	-	p.rect = inner
190	-	p.popupGeometry = geometry
191	-	p.mu.Unlock()
192	-	return true
193	-	}
194	-	if pty != nil {
195	-	if err := pty.Resize(rectSize(inner)); err != nil {
196	-	d.log.Warn("floating pty resize failed", "err", err)
197	-	return false
198	-	}
199	-	}
200	-	p.mu.Lock()
201	-	p.screen.Resize(inner.Width, inner.Height)
202	-	p.rect = inner
203	-	p.popupGeometry = geometry
204	-	p.mu.Unlock()
205	-	return true
206	-	}
207	-
208	-	// resizeInstalledFloating updates an installed floating pane, including a
209	-	// retained hidden pane when its tab becomes active.
210	-	func (d Daemon) resizeInstalledFloating(tb tab) bool {
211	-	if d == nil \|\| tb == nil {
212	-	return false
213	-	}
214	-	tb.mu.Lock()
215	-	p := tb.floating.pane
216	-	content := domain.Size{Cols: tb.size.Cols, Rows: tb.size.Rows}
217	-	tb.mu.Unlock()
218	-	if p == nil {
219	-	return false
220	-	}
221	-	return d.resizeFloatingPane(p, calculateContentFloatingGeometry(content, d.currentFloatingConfig()))
222	-	}
223	-
224	-	// resizeActiveFloating updates a visible floating pane during a client resize.
225	-	func (d Daemon) resizeActiveFloating(tb tab) bool {
226	-	if d == nil \|\| tb == nil {
227	-	return false
228	-	}
229	-	tb.mu.Lock()
230	-	visible := tb.floating.state == floatingVisible
231	-	tb.mu.Unlock()
232	-	if !visible {
233	-	return false
234	-	}
235	-	return d.resizeInstalledFloating(tb)
236	-	}

A internal/usecase/daemon/floating_render_legacy_test.go +72 -0

@@ -0,0 +1,72 @@
	1	+	package daemon
	2	+
	3	+	import (
	4	+	"github.com/bnema/vev/internal/domain"
	5	+	themeui "github.com/bnema/vev/internal/usecase/theme"
	6	+	"github.com/bnema/vev/pkg/renderer"
	7	+	)
	8	+
	9	+	// composeFloatingFrame applies the popup to a copy of the normal composed
	10	+	// destination. In particular it never writes the VT screen frame: that frame
	11	+	// is shared with the PTY reader and is the source for future renders.
	12	+	func composeFloatingFrame(base renderer.Frame, baseDamage []renderer.Damage, p pane, generation uint64, content domain.Rect, desired floatingGeometry, layoutSnap tabLayoutSnapshot, theme themeui.Theme, cache composedFrameCache, full bool) (renderer.Frame, []renderer.Damage, floatingGeometry) {
	13	+	var frame renderer.Frame
	14	+	if cache != nil && cache.floatingFrame.Width == base.Width && cache.floatingFrame.Height == base.Height {
	15	+	frame = cache.floatingFrame
	16	+	for y := 0; y < base.Height; y++ {
	17	+	copy(frame.Row(y), base.Row(y))
	18	+	}
	19	+	} else {
	20	+	frame = base.Clone()
	21	+	if cache != nil {
	22	+	cache.floatingFrame = frame
	23	+	}
	24	+	}
	25	+	(overlayBackdrop{DimPaneContents: true}).apply(frame, content, layoutSnap, theme)
	26	+	// The PTY reader mutates both the frame and its damage under p.mu. Keep
	27	+	// every read of them, including the blit, in the same critical section.
	28	+	// In particular, Frame is not safe to retain as an alias after unlocking.
	29	+	p.mu.Lock()
	30	+	// Committed popup geometry is content-relative; translate it exactly once
	31	+	// while holding the same lock that protects the committed state and frame.
	32	+	geometry := p.committedFloatingGeometryLocked(desired).translate(content.X, content.Y)
	33	+	previousCapture := capturedPaneRenderState{}
	34	+	if cache != nil {
	35	+	previousCapture = cache.floatingCaptured
	36	+	}
	37	+	captured := capturePaneRenderStateLockedInto(p, geometry.Inner, previousCapture)
	38	+	// This retained legacy test helper models its historical synchronous
	39	+	// consume path; production commits receipts in emitFrame instead.
	40	+	p.screen.AcknowledgeDamage(captured.damageGeneration)
	41	+	if cache != nil {
	42	+	cache.floatingCaptured = captured
	43	+	}
	44	+	p.mu.Unlock()
	45	+	title := captured.title
	46	+	titleGeneration := captured.titleGeneration
	47	+	blitPaneFrame(frame, geometry.Inner, captured.frame, false, theme)
	48	+	damage := append([]renderer.Damage(nil), baseDamage...)
	49	+	for _, d := range captured.damage {
	50	+	damage = append(damage, translatePaneDamage(d, geometry.Inner, content)...)
	51	+	}
	52	+
	53	+	// The committed geometry, rather than the requested config geometry, is
	54	+	// what was actually rendered. Track it so movement (including a same-size
	55	+	// resize) cannot reuse a cached popup at its former position.
	56	+	popupChanged := cache == nil \|\| cache.floating != p \|\| cache.floatingGeneration != generation \|\| cache.floatingGeometry != geometry
	57	+	titleChanged := popupChanged \|\| cache == nil \|\| cache.floatingTitleGeneration != titleGeneration
	58	+	drawFloatingBorder(frame, geometry.Bounds, title, newThemeStyles(theme).border)
	59	+	if cache != nil {
	60	+	cache.floating = p
	61	+	cache.floatingGeneration = generation
	62	+	cache.floatingGeometry = geometry
	63	+	cache.floatingTitleGeneration = titleGeneration
	64	+	}
	65	+	if full \|\| popupChanged {
	66	+	return frame, []renderer.Damage{renderer.FullRedraw()}, geometry
	67	+	}
	68	+	if titleChanged && geometry.Bounds.Height >= 3 {
	69	+	damage = append(damage, renderer.Damage{Kind: renderer.DamageText, X: geometry.Bounds.X, Y: geometry.Bounds.Y, Width: geometry.Bounds.Width, Height: 1})
	70	+	}
	71	+	return frame, damage, geometry
	72	+	}

M internal/usecase/daemon/floating_render_test.go +28 -44

 	"github.com/bnema/vev/pkg/renderer"
 )
+// applyFloatingResizePlanForTest exercises the transactional apply primitive.
+func applyFloatingResizePlanForTest(d *Daemon, p *pane, geometry floatingGeometry) bool {
+	if p == nil || !geometry.valid() {
+		return false
+	}
+	plan := resizePlan{members: []resizeMember{{pane: p, rect: geometry.Inner, floating: geometry, isFloating: true}}}
+	d.applyResize(&plan)
+	member := plan.members[0]
+	if !member.ok {
+		return false
+	}
+	p.mu.Lock()
+	p.rect = geometry.Inner
+	p.popupGeometry = geometry
+	p.screen.Resize(geometry.Inner.Width, geometry.Inner.Height)
+	p.mu.Unlock()
+	return true
+}
 func TestFloatingGeometryTranslate(t *testing.T) {
 	geometry := floatingGeometry{
 		Bounds: domain.Rect{X: 3, Y: 5, Width: 10, Height: 8},
 			p.rect, p.popupGeometry = initial.Inner, initial
 			d := newTestDaemon(t, nil, stubClock{})
-			require.True(t, d.resizeFloatingPane(p, tt.next))
+			require.True(t, applyFloatingResizePlanForTest(d, p, tt.next))
 			require.Empty(t, pty.sizes(), "same-size geometry must not resize the PTY")
 			require.Equal(t, tt.next.Inner, p.rect)
 			require.Equal(t, tt.next, p.popupGeometry)
 	}
 }
-func TestRenderHiddenFloatingSkipsPreviewAndActiveClientAndClearsDamage(t *testing.T) {
-	normalPTY, releaseNormal := newBlockingPTY(t)
-	defer releaseNormal()
-	d, sess, _, activeSends := newManualSessionWithPTYs(t, normalPTY)
-	previewTransport, previewSends := newCapturingTransport(t)
-	preview := &attachedClient{tr: previewTransport, output: newOutputStateStream(), size: domain.Size{Cols: 80, Rows: 24}}
-	preview.initOverlays()
-	preview.setSession(sess)
-	tb := sess.activeTab()
-	floating := newPane("floating", nil, domain.Size{Cols: 20, Rows: 8})
-	floating.screen.Write([]byte("hidden output"))
-	tb.mu.Lock()
-	generation := tb.beginFloatingWarmLocked(false)
-	require.True(t, tb.installFloatingLocked(floating, generation))
-	tb.previewClient = preview
-	tb.mu.Unlock()
-	d.render(sess, tb, floating)
-	floating.mu.Lock()
-	require.Empty(t, floating.screen.Damage())
-	floating.mu.Unlock()
-	select {
-	case frame := <-previewSends:
-		t.Fatalf("hidden floating output painted picker preview: %#v", frame)
-	default:
-	}
-	select {
-	case frame := <-activeSends:
-		t.Fatalf("hidden floating output painted active client: %#v", frame)
-	default:
-	}
-}
 func BenchmarkComposeFloatingFrameCached(b *testing.B) {
 	p := newPane("floating", nil, domain.Size{Cols: 62, Rows: 18})
 	p.title.displayFallback = "float"
 	d.resize(sess, nil, domain.Size{Cols: 100, Rows: 42})
 	require.Empty(t, pty.sizes())
-	// Showing attempts the current size before paint. A failed resize leaves the
+	// Showing attempts the current size before paint. A failed resize retains
-	// old screen and rect untouched, so a later hide/show can retry.
+	// the old screen, but commits the new popup rect so capture clips/pads it.
 	require.NoError(t, d.toggleFloating(sess, nil))
 	require.Equal(t, []domain.Size{rectSize(current)}, pty.sizes())
-	require.Equal(t, initial, floating.rect)
+	require.Equal(t, current, floating.rect)
 	require.Equal(t, initial.Width, floating.screen.Frame.Width)
 	require.Equal(t, initial.Height, floating.screen.Frame.Height)
-	require.Equal(t, initialGeometry, floating.popupGeometry)
+	require.Equal(t, currentGeometry, floating.popupGeometry)
 	require.NoError(t, d.toggleFloating(sess, nil)) // hide
 	require.NoError(t, d.toggleFloating(sess, nil)) // retry show
 	d := newTestDaemon(t, nil, stubClock{})
 	d.ApplyConfig(domain.Config{Floating: cfg})
-	require.False(t, d.resizeFloatingPane(p, newGeometry))
+	require.False(t, applyFloatingResizePlanForTest(d, p, newGeometry))
 	tb.mu.Lock()
 	_, inputGeometry, visible := tb.visibleFloatingSnapshotLocked(cfg)
 	copyRect := copyTargetRectLocked(tabLayoutSnapshot{}, newContent, p, p, true, oldGeometry.translate(newContent.X, newContent.Y))
 		p.rect = domain.Rect{X: 8, Y: 3, Width: 5, Height: 4}
 		d := newTestDaemon(t, nil, stubClock{})
 		requested := floatingGeometry{Bounds: domain.Rect{X: 1, Y: 1, Width: 11, Height: 9}, Inner: domain.Rect{X: 2, Y: 2, Width: 9, Height: 7}}
-		require.False(t, d.resizeFloatingPane(p, requested))
+		require.False(t, applyFloatingResizePlanForTest(d, p, requested))
 		require.Equal(t, domain.Rect{X: 8, Y: 3, Width: 5, Height: 4}, p.rect)
 		require.Equal(t, floatingGeometry{}, p.popupGeometry)
 		require.Equal(t, 5, p.screen.Frame.Width)
 		done1 := make(chan bool, 1)
 		done2 := make(chan bool, 1)
 		secondStarted := make(chan struct{})
-		go func() { done1 <- d.resizeFloatingPane(p, first) }()
+		go func() { done1 <- applyFloatingResizePlanForTest(d, p, first) }()
 		<-pty.entered
 		go func() {
 			close(secondStarted)
-			done2 <- d.resizeFloatingPane(p, second)
+			done2 <- applyFloatingResizePlanForTest(d, p, second)
 		}()
 		<-secondStarted
 		close(pty.release)

M internal/usecase/daemon/floating_test.go +5 -5


@@ -36,8 +36,8 @@ func newGatedOpenFactory(t testing.T, result ports.PTY, openErr error) (portsm
36		var releaseOnce sync.Once	36		var releaseOnce sync.Once
37		releaseOpen := func() { releaseOnce.Do(func() { close(release) }) }	37		releaseOpen := func() { releaseOnce.Do(func() { close(release) }) }
38		t.Cleanup(releaseOpen)	38		t.Cleanup(releaseOpen)
39	-	factory.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything).RunAndReturn(	39	+	factory.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything).RunAndReturn(
40	-	func(command string, args, _ []string, dir string, size domain.Size) (ports.PTY, error) {	40	+	func(_ context.Context, command string, args, _ []string, dir string, size domain.Size) (ports.PTY, error) {
41		opened <- gatedFloatingOpen{command: command, args: append([]string(nil), args...), dir: dir, size: size}	41		opened <- gatedFloatingOpen{command: command, args: append([]string(nil), args...), dir: dir, size: size}
42		<-release	42		<-release
43		return result, openErr	43		return result, openErr

@@ -364,7 +364,7 @@ func TestFloatingEOFRepaintsVisibleSlotOnly(t *testing.T) {
364			364
365		// Establish a renderer shadow first. The EOF repaint must reset it and	365		// Establish a renderer shadow first. The EOF repaint must reset it and
366		// redraw the underlying cell rather than depending on popup damage.	366		// redraw the underlying cell rather than depending on popup damage.
367	-	d.paint(sess, ac, true)	367	+	d.paint(sess, ac, true, nil)
368		baseline := awaitFrame(t, sends, ports.MsgOutput)	368		baseline := awaitFrame(t, sends, ports.MsgOutput)
369		baselineOutput, err := ports.UnmarshalOutput(baseline.Payload)	369		baselineOutput, err := ports.UnmarshalOutput(baseline.Payload)
370		require.NoError(t, err)	370		require.NoError(t, err)

@@ -467,8 +467,8 @@ func TestFloatingLaunchUsesLiveOrValidatedSessionCwd(t *testing.T) {
467		t.Run(tt.name, func(t *testing.T) {	467		t.Run(tt.name, func(t *testing.T) {
468		factory := portsmocks.NewMockPTYFactory(t)	468		factory := portsmocks.NewMockPTYFactory(t)
469		opened := make(chan string, 1)	469		opened := make(chan string, 1)
470	-	factory.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything).RunAndReturn(	470	+	factory.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything).RunAndReturn(
471	-	func(_ string, _ []string, _ []string, cwd string, _ domain.Size) (ports.PTY, error) {	471	+	func(_ context.Context, _ string, _ []string, _ []string, cwd string, _ domain.Size) (ports.PTY, error) {
472		opened <- cwd	472		opened <- cwd
473		return nil, io.ErrUnexpectedEOF	473		return nil, io.ErrUnexpectedEOF
474		}).Once()	474		}).Once()

A internal/usecase/daemon/history_test_helper_test.go +10 -0

@@ -0,0 +1,10 @@
	1	+	package daemon
	2	+
	3	+	import "github.com/bnema/vev/pkg/vt"
	4	+
	5	+	func newTestHistory(rows int) *vt.History { return vt.NewHistory(vt.HistoryConfig{MaxRows: rows}) }
	6	+
	7	+	func installTestHistory(p *pane, config vt.HistoryConfig) {
	8	+	p.screen = vt.NewScreenWithHistory(p.screen.Frame.Width, p.screen.Frame.Height, config)
	9	+	p.history = p.screen.History()
	10	+	}

M internal/usecase/daemon/input.go +9 -32

-// Package daemon holds vev's server-side session multiplexer use case: the
+// Package daemon holds vev's server-side session multiplexer use case.
-// accept loop, the ephemeral/named session registry, the per-tab PTY reader
-// and VT screen, and the per-client debounced render scheduler.
-//
-// Concurrency model (sessions own one or more PTY-backed tabs):
-//
-//   - Serve runs the accept loop. Each accepted connection is handled by its
-//     own goroutine (handleConn): it reads the first frame and routes it to a
-//     session create/attach, a list, or a kill.
-//   - Per session there are exactly two long-lived goroutines: the PTY reader
-//     (drains child output into the VT screen and pokes a cap-1 dirty channel)
-//     and the render scheduler (debounces dirties and paints the attached
-//     client). Both are tied to the session context and unwind when the
-//     session is killed (pty.Close unblocks the reader; ctx cancel stops the
-//     scheduler).
-//   - The daemon exits (Serve returns) when the last session is removed, or
-//     when the parent context is cancelled (graceful shutdown notifies any
-//     attached clients with ReasonServerShutdown).
-//
-// Locking: a pane's screen/scrollback and per-client renderer shadow are
-// guarded by pane.mu/tab.mu as appropriate; the attached-client pointer by
-// session.mu; the registry by Daemon.mu. When more than one is held the order
-// is always attachedClient.sendMu > Daemon.mu > session.mu > tab.mu > pane.mu.
-// The PTY reader only ever takes pane.mu, so it never blocks on a slow client.
 package daemon
 import (
 					if target.ID != oldFocus {
 						d.refreshPaneTitleOnFocus(sess, target.ID)
 					}
-					d.paint(sess, ac, true)
+					d.invalidateRender(sess, ac, true, "input.go")
 					return
 				}
 			}
 			d.exitCopyMode(ac)
 			d.refreshPaneTitleOnFocus(sess, pl.ID)
 		}
-		d.paint(sess, ac, true)
+		d.invalidateRender(sess, ac, true, "input.go")
 		return
 	}
 	var p *pane
 		if isMouseFocusPress(ev) && pl.ID != oldFocus {
 			d.exitCopyMode(ac)
 			d.refreshPaneTitleOnFocus(sess, pl.ID)
-			d.paint(sess, ac, true)
+			d.invalidateRender(sess, ac, true, "input.go")
 		}
 		if multi {
 			ev = translateMouseEvent(ev, pl.Content.X, pl.Content.Y)
 	mouseMode, mouseSGR := p.screen.MouseMode()
 	altScreen := p.screen.AltScreenActive()
 	scrollbackRows := 0
-	if p.scrollback != nil {
+	if p.history != nil {
-		scrollbackRows = p.scrollback.Len()
+		scrollbackRows = p.history.Len()
 	}
 	p.mu.Unlock()
 			// The immutable snapshot is the only live-pane access needed for drag entry.
 			p.mu.Lock()
-			document := scopy.NewSnapshot(p.scrollback, p.screen.Frame)
+			document := scopy.NewSnapshot(p.history, p.screen.Frame)
 			p.mu.Unlock()
 			tb := sess.activeTab()
 			if !d.publishCopyMode(sess, ac, tb, p, document, func(mode *scopy.Mode) {
 				rt.copyDragging = true
 			}
 			rt.copyMu.Unlock()
-			d.paint(sess, ac, true)
+			d.invalidateRender(sess, ac, true, "input.go")
 		case mouse.Release:
 			rt.copyMu.Lock()
 			rt.normalMousePressValid = false
 		idx := int(action - keys.ActionSwitchTab1)
 		if sess.switchTab(idx) {
 			h.d.activateTab(sess, sess.activeTab())
-			h.d.paint(sess, h.ac, true)
+			h.d.invalidateRender(sess, h.ac, true, "input.go")
 		}
 	}
 }

M internal/usecase/daemon/input_test.go +188 -13

@@ -1,6 +1,7 @@
1		package daemon	1		package daemon
2			2
3		import (	3		import (
			4	+	"context"
4		"encoding/base64"	5		"encoding/base64"
5		"fmt"	6		"fmt"
6		"strings"	7		"strings"

@@ -14,10 +15,10 @@ import (
14		"github.com/bnema/vev/internal/domain"	15		"github.com/bnema/vev/internal/domain"
15		"github.com/bnema/vev/internal/ports"	16		"github.com/bnema/vev/internal/ports"
16		portsmocks "github.com/bnema/vev/internal/ports/mocks"	17		portsmocks "github.com/bnema/vev/internal/ports/mocks"
17	-	scopy "github.com/bnema/vev/internal/usecase/copy"
18		"github.com/bnema/vev/internal/usecase/keys"	18		"github.com/bnema/vev/internal/usecase/keys"
19		"github.com/bnema/vev/internal/usecase/layout"	19		"github.com/bnema/vev/internal/usecase/layout"
20		"github.com/bnema/vev/internal/usecase/picker"	20		"github.com/bnema/vev/internal/usecase/picker"
			21	+	"github.com/bnema/vev/pkg/vt"
21		)	22		)
22			23
23		// --- test doubles -----------------------------------------------------------	24		// --- test doubles -----------------------------------------------------------

@@ -57,11 +58,11 @@ func TestAltDigitSwitchesBetweenThreeTabs(t *testing.T) {
57		return len(sessions.Sessions) == 1 && sessions.Sessions[0].Tabs == 3	58		return len(sessions.Sessions) == 1 && sessions.Sessions[0].Tabs == 3
58		}, 2time.Second, 5time.Millisecond)	59		}, 2time.Second, 5time.Millisecond)
59		require.Eventually(t, func() bool { return len(writes1) == 1 }, 2time.Second, 5time.Millisecond)	60		require.Eventually(t, func() bool { return len(writes1) == 1 }, 2time.Second, 5time.Millisecond)
60	-	require.Equal(t, []byte("A"), <-writes1)	61	+	requirePTYWrite(t, writes1, []byte("A"))
61		require.Eventually(t, func() bool { return len(writes2) == 1 }, 2time.Second, 5time.Millisecond)	62		require.Eventually(t, func() bool { return len(writes2) == 1 }, 2time.Second, 5time.Millisecond)
62	-	require.Equal(t, []byte("B"), <-writes2)	63	+	requirePTYWrite(t, writes2, []byte("B"))
63		require.Eventually(t, func() bool { return len(writes3) == 1 }, 2time.Second, 5time.Millisecond)	64		require.Eventually(t, func() bool { return len(writes3) == 1 }, 2time.Second, 5time.Millisecond)
64	-	require.Equal(t, []byte("C"), <-writes3)	65	+	requirePTYWrite(t, writes3, []byte("C"))
65			66
66		d.mu.Lock()	67		d.mu.Lock()
67		var sess *session	68		var sess *session

@@ -87,6 +88,77 @@ func TestAltDigitSwitchesBetweenThreeTabs(t *testing.T) {
87		d.sessWg.Wait()	88		d.sessWg.Wait()
88		}	89		}
89			90
			91	+	func TestSwitchTabFirstFrameDoesNotReuseSamePaneIDCapture(t *testing.T) {
			92	+	d, sess, ac, sends := newManualSessionWithPTYs(t, nil, nil)
			93	+	source := sess.tabs[0].focusedPane()
			94	+	target := sess.tabs[1].focusedPane()
			95	+	require.Equal(t, source.id, target.id, "tabs deliberately reuse pane-1")
			96	+
			97	+	source.mu.Lock()
			98	+	source.screen.Write([]byte("source"))
			99	+	source.mu.Unlock()
			100	+	d.paint(sess, ac, true, nil)
			101	+	first := awaitFrame(t, sends, ports.MsgOutput)
			102	+	firstOutput, err := ports.UnmarshalOutput(first.Payload)
			103	+	require.NoError(t, err)
			104	+	terminal := vt.NewScreen(ac.size.Cols, ac.size.Rows)
			105	+	terminal.Write(firstOutput.Data)
			106	+
			107	+	// A clean pane relies on the attachment capture cache for its retained
			108	+	// snapshot. Both tabs deliberately use the tab-local default pane ID.
			109	+	target.mu.Lock()
			110	+	target.screen.ClearDamage()
			111	+	target.mu.Unlock()
			112	+	// Use the real key action: it requests the mandatory complete repaint for
			113	+	// the first target-tab frame.
			114	+	daemonKeyHandler{d: d, ac: ac}.Action(keys.ActionSwitchTab2)
			115	+	require.Equal(t, 1, activeTabIndex(sess))
			116	+
			117	+	second := awaitFrame(t, sends, ports.MsgOutput)
			118	+	secondOutput, err := ports.UnmarshalOutput(second.Payload)
			119	+	require.NoError(t, err)
			120	+	require.Zero(t, secondOutput.BaseStateNum, "tab switch must emit the complete target frame first")
			121	+	terminal.Write(secondOutput.Data)
			122	+	require.NotContains(t, strings.Join(frameRows(terminal.Frame), "\n"), "source", "the first target-tab frame must not retain source pane cells")
			123	+	}
			124	+
			125	+	func TestClosePanePrunesAttachedCaptureFrameAndKeepsSurvivor(t *testing.T) {
			126	+	d, sess, ac, _ := newManualSessionWithPTYs(t, nil)
			127	+	tb := sess.tabs[0]
			128	+	survivor := tb.panes["pane-1"]
			129	+	closed := newPane("pane-2", nil, domain.Size{Cols: 40, Rows: 23})
			130	+	tb.mu.Lock()
			131	+	tb.tree = &layout.Tree{Root: &layout.Node{Kind: layout.Split, Dir: layout.Horizontal, Children: []*layout.Node{layout.NewLeaf("pane-1"), layout.NewLeaf("pane-2")}}, Focus: "pane-2"}
			132	+	tb.panes[closed.id] = closed
			133	+	tb.mu.Unlock()
			134	+
			135	+	cachedSurvivor := capturedPaneRenderState{title: "survivor"}
			136	+	ac.captureFrames = map[*pane]capturedPaneRenderState{
			137	+	survivor: cachedSurvivor,
			138	+	closed: {title: "closed"},
			139	+	}
			140	+
			141	+	require.NoError(t, d.closePane(sess, tb, closed.id, ac, false))
			142	+	require.NotContains(t, ac.captureFrames, closed, "closed pane must not remain strongly retained by its attachment")
			143	+	require.Equal(t, cachedSurvivor, ac.captureFrames[survivor], "surviving pane keeps its incremental capture")
			144	+	}
			145	+
			146	+	func TestCloseTabPrunesAttachedCaptureFrames(t *testing.T) {
			147	+	d, sess, ac, _ := newManualSessionWithPTYs(t, nil, nil)
			148	+	closedTab, survivorTab := sess.tabs[0], sess.tabs[1]
			149	+	closed := closedTab.panes["pane-1"]
			150	+	survivor := survivorTab.panes["pane-1"]
			151	+	cachedSurvivor := capturedPaneRenderState{title: "survivor"}
			152	+	ac.captureFrames = map[*pane]capturedPaneRenderState{
			153	+	closed: {title: "closed"},
			154	+	survivor: cachedSurvivor,
			155	+	}
			156	+
			157	+	d.closeTab(sess, closedTab, false)
			158	+	require.NotContains(t, ac.captureFrames, closed, "closed tab panes must not remain strongly retained by its attachment")
			159	+	require.Equal(t, cachedSurvivor, ac.captureFrames[survivor], "surviving tab pane keeps its incremental capture")
			160	+	}
			161	+
90		func TestAltCForwardsToPTY(t *testing.T) {	162		func TestAltCForwardsToPTY(t *testing.T) {
91		writes := make(chan []byte, 2)	163		writes := make(chan []byte, 2)
92		p, releasePTY := newBlockingPTYWithWrites(t, writes)	164		p, releasePTY := newBlockingPTYWithWrites(t, writes)

@@ -94,10 +166,74 @@ func TestAltCForwardsToPTY(t *testing.T) {
94			166
95		d.handleInput(sess, ac, []byte("\x1bc"))	167		d.handleInput(sess, ac, []byte("\x1bc"))
96			168
97	-	require.Equal(t, []byte("\x1bc"), <-writes)	169	+	requirePTYWrite(t, writes, []byte("\x1bc"))
98		releasePTY()	170		releasePTY()
99		}	171		}
100			172
			173	+	func TestAltFToggleRetainedFloatingPaneRepaintsImmediately(t *testing.T) {
			174	+	normal, releaseNormal := newBlockingPTY(t)
			175	+	floatingPTY, releaseFloating := newBlockingPTY(t)
			176	+	defer releaseNormal()
			177	+	defer releaseFloating()
			178	+	d, sess, ac, sends := newManualSessionWithPTYs(t, normal)
			179	+	client := vt.NewScreen(80, 25)
			180	+	floatingCtx, cancelFloating := context.WithCancel(sess.ctx)
			181	+	defer cancelFloating()
			182	+	floating := newPane("floating", floatingPTY, domain.Size{Cols: 20, Rows: 5})
			183	+	floating.ctx = floatingCtx
			184	+	floating.screen.Write([]byte("popup-content"))
			185	+	installTestFloating(sess.activeTab(), floating, false)
			186	+
			187	+	// Establish the client shadow while the retained popup is hidden.
			188	+	d.paint(sess, ac, true, nil)
			189	+	mustApplyOutput(t, client, awaitFrame(t, sends, ports.MsgOutput))
			190	+
			191	+	// Route the real Alt+F binding. Showing a retained pane must paint once.
			192	+	d.handleInput(sess, ac, []byte("\x1bf"))
			193	+	mustApplyOutput(t, client, awaitFrame(t, sends, ports.MsgOutput))
			194	+	require.Contains(t, strings.Join(frameRows(client.Frame), "\n"), "popup-content")
			195	+	require.Contains(t, strings.Join(frameRows(client.Frame), "\n"), "┌")
			196	+	select {
			197	+	case extra := <-sends:
			198	+	t.Fatalf("show emitted duplicate output: %#v", extra)
			199	+	default:
			200	+	}
			201	+
			202	+	// The second real key must repaint immediately, rather than waiting for
			203	+	// unrelated output, while retaining the existing PTY and context.
			204	+	d.handleInput(sess, ac, []byte("\x1bf"))
			205	+	mustApplyOutput(t, client, awaitFrame(t, sends, ports.MsgOutput))
			206	+	require.NotContains(t, strings.Join(frameRows(client.Frame), "\n"), "popup-content")
			207	+	tb := sess.activeTab()
			208	+	tb.mu.Lock()
			209	+	require.Equal(t, floatingHidden, tb.floating.state)
			210	+	require.Same(t, floating, tb.floating.pane)
			211	+	generation := tb.floating.generation
			212	+	tb.mu.Unlock()
			213	+	require.Same(t, floatingPTY, floating.pty)
			214	+	select {
			215	+	case <-floating.ctx.Done():
			216	+	t.Fatal("hiding retained popup cancelled its context")
			217	+	default:
			218	+	}
			219	+
			220	+	// A third show uses the installed pane; it must not launch or emit a second
			221	+	// resize/start frame.
			222	+	d.handleInput(sess, ac, []byte("\x1bf"))
			223	+	mustApplyOutput(t, client, awaitFrame(t, sends, ports.MsgOutput))
			224	+	require.Contains(t, strings.Join(frameRows(client.Frame), "\n"), "popup-content")
			225	+	tb.mu.Lock()
			226	+	require.Equal(t, floatingVisible, tb.floating.state)
			227	+	require.Same(t, floating, tb.floating.pane)
			228	+	require.Equal(t, generation, tb.floating.generation)
			229	+	tb.mu.Unlock()
			230	+	select {
			231	+	case extra := <-sends:
			232	+	t.Fatalf("retained show emitted duplicate output: %#v", extra)
			233	+	default:
			234	+	}
			235	+	}
			236	+
101		func requireFloatingInitialized(t testing.T, tb tab) {	237		func requireFloatingInitialized(t testing.T, tb tab) {
102		t.Helper()	238		t.Helper()
103		tb.mu.Lock()	239		tb.mu.Lock()

@@ -260,7 +396,7 @@ func TestBracketedMultilinePasteForwardsDelimitersAndNewlines(t *testing.T) {
260		paste := []byte("\x1b[200~first line\nsecond line\n\x1b[201~")	396		paste := []byte("\x1b[200~first line\nsecond line\n\x1b[201~")
261		d.handleInput(sess, ac, paste)	397		d.handleInput(sess, ac, paste)
262			398
263	-	require.Equal(t, paste, <-writes)	399	+	requirePTYWrite(t, writes, paste)
264		releasePTY()	400		releasePTY()
265		}	401		}
266			402

@@ -342,7 +478,7 @@ func TestOverlayInputPrecedence(t *testing.T) {
342		input: []byte("Z"),	478		input: []byte("Z"),
343		check: func(t testing.T, ac attachedClient, writes chan []byte) {	479		check: func(t testing.T, ac attachedClient, writes chan []byte) {
344		t.Helper()	480		t.Helper()
345	-	require.Equal(t, []byte("Z"), <-writes)	481	+	requirePTYWrite(t, writes, []byte("Z"))
346		},	482		},
347		},	483		},
348		}	484		}

@@ -466,6 +602,45 @@ func TestStaleBackSessionClearPreservesConcurrentTarget(t *testing.T) {
466		require.Same(t, updated, ac.previousSession.Get())	602		require.Same(t, updated, ac.previousSession.Get())
467		}	603		}
468			604
			605	+	// backSession's invalid-target fallback is a render producer: it must publish
			606	+	// exactly one coordinator invalidation instead of composing directly, while
			607	+	// the valid-target hand-off (TestPaletteBackSessionTogglesPreviousSession) and
			608	+	// the guarded stale clear (TestStaleBackSessionClearPreservesConcurrentTarget)
			609	+	// keep their PR #73 behavior unchanged.
			610	+	func TestBackSessionInvalidTargetsFallBackThroughOneInvalidation(t *testing.T) {
			611	+	for _, tc := range []struct {
			612	+	name string
			613	+	prepare func(current session, ac attachedClient)
			614	+	}{
			615	+	{name: "no previous target", prepare: func(session, attachedClient) {}},
			616	+	{name: "stale previous target", prepare: func(_ session, ac attachedClient) {
			617	+	ac.previousSession.Set(&session{id: "gone"})
			618	+	}},
			619	+	{name: "previous target equals current", prepare: func(current session, ac attachedClient) {
			620	+	ac.previousSession.Set(current)
			621	+	}},
			622	+	} {
			623	+	t.Run(tc.name, func(t *testing.T) {
			624	+	d, current, ac, sends, releases := newRecentNavigationTestSessions(t)
			625	+	defer releaseAll(releases)
			626	+	invs := make(chan renderInvalidation, 4)
			627	+	current.installRenderCoordinator(newRenderCoordinator(renderCoordinatorOptions{
			628	+	clock: d.clock,
			629	+	wake: func(renderWake) {},
			630	+	onInvalidate: func(inv renderInvalidation) { invs <- inv },
			631	+	}))
			632	+	tc.prepare(current, ac)
			633	+
			634	+	d.backSession(current, ac)
			635	+
			636	+	require.Same(t, current, ac.currentSession(), "an invalid target must not move the attachment")
			637	+	awaitInvalidation(t, invs)
			638	+	requireNoInvalidation(t, invs)
			639	+	requireNoOutputFrame(t, sends)
			640	+	})
			641	+	}
			642	+	}
			643	+
469		func TestSwitchSourcePreviousSessionContracts(t *testing.T) {	644		func TestSwitchSourcePreviousSessionContracts(t *testing.T) {
470		for _, tc := range []struct {	645		for _, tc := range []struct {
471		name string	646		name string

@@ -592,7 +767,7 @@ func TestAltXClosesActiveTabAndSelectsRemaining(t *testing.T) {
592		require.Equal(t, 1, activeTabIndex(sess), "closing middle tab selects the next remaining tab")	767		require.Equal(t, 1, activeTabIndex(sess), "closing middle tab selects the next remaining tab")
593		d.handleInput(sess, ac, []byte("Z"))	768		d.handleInput(sess, ac, []byte("Z"))
594		require.Eventually(t, func() bool { return len(writes) == 1 }, 2time.Second, 5time.Millisecond)	769		require.Eventually(t, func() bool { return len(writes) == 1 }, 2time.Second, 5time.Millisecond)
595	-	require.Equal(t, []byte("Z"), <-writes)	770	+	requirePTYWrite(t, writes, []byte("Z"))
596			771
597		releasePTY1()	772		releasePTY1()
598		releasePTY2()	773		releasePTY2()

@@ -696,8 +871,8 @@ func TestMouseAltScreenWheelMapsToArrows(t *testing.T) {
696		d.handleInput(sess, ac, []byte("\x1b[<64;1;1M"))	871		d.handleInput(sess, ac, []byte("\x1b[<64;1;1M"))
697		d.handleInput(sess, ac, []byte("\x1b[<65;1;1M"))	872		d.handleInput(sess, ac, []byte("\x1b[<65;1;1M"))
698			873
699	-	require.Equal(t, []byte("\x1b[A\x1b[A\x1b[A"), <-writes)	874	+	requirePTYWrite(t, writes, []byte("\x1b[A\x1b[A\x1b[A"))
700	-	require.Equal(t, []byte("\x1b[B\x1b[B\x1b[B"), <-writes)	875	+	requirePTYWrite(t, writes, []byte("\x1b[B\x1b[B\x1b[B"))
701		require.Nil(t, ac.overlays.copyMode)	876		require.Nil(t, ac.overlays.copyMode)
702		}	877		}
703			878

@@ -752,7 +927,7 @@ func TestMouseChildForwardingStatusDropAndPressDrop(t *testing.T) {
752			927
753		sess.tabs[0].focusedPane().screen.Write([]byte("\x1b[?1006h"))	928		sess.tabs[0].focusedPane().screen.Write([]byte("\x1b[?1006h"))
754		d.handleInput(sess, ac, raw)	929		d.handleInput(sess, ac, raw)
755	-	require.Equal(t, []byte("\x1b[<0;2;2M"), <-writes)	930	+	requirePTYWrite(t, writes, []byte("\x1b[<0;2;2M"))
756			931
757		d.handleInput(sess, ac, []byte("\x1b[<0;1;1M"))	932		d.handleInput(sess, ac, []byte("\x1b[<0;1;1M"))
758		select {	933		select {

@@ -892,7 +1067,7 @@ func TestCopyModeStatusRowPressClearsDragState(t *testing.T) {
892		func TestMouseNormalScreenDragExtendsToCurrentScrollbackOffset(t *testing.T) {	1067		func TestMouseNormalScreenDragExtendsToCurrentScrollbackOffset(t *testing.T) {
893		p, _ := newBlockingPTY(t)	1068		p, _ := newBlockingPTY(t)
894		d, sess, ac, _ := newManualSessionWithPTYs(t, p)	1069		d, sess, ac, _ := newManualSessionWithPTYs(t, p)
895	-	sess.tabs[0].focusedPane().scrollback = scopy.NewScrollback(50)	1070	+	installTestHistory(sess.tabs[0].focusedPane(), vt.HistoryConfig{MaxRows: 50})
896		require.Equal(t, 23, sess.tabs[0].focusedPane().screen.Frame.Height, "fixture assumption: status row is wire row 24")	1071		require.Equal(t, 23, sess.tabs[0].focusedPane().screen.Frame.Height, "fixture assumption: status row is wire row 24")
897			1072
898		// Press on a content row while scrollback is empty: the anchor is	1073		// Press on a content row while scrollback is empty: the anchor is

@@ -902,7 +1077,7 @@ func TestMouseNormalScreenDragExtendsToCurrentScrollbackOffset(t *testing.T) {
902		// Simulate 5 lines evicted into scrollback between the Press and the	1077		// Simulate 5 lines evicted into scrollback between the Press and the
903		// first Motion (e.g. the child kept producing output).	1078		// first Motion (e.g. the child kept producing output).
904		for range 5 {	1079		for range 5 {
905	-	sess.tabs[0].focusedPane().scrollback.Append(testRow("evicted"))	1080	+	sess.tabs[0].focusedPane().history.Append(testRow("evicted"))
906		}	1081		}
907			1082
908		// Motion lands on wire row 3 (0-based row 2 of the current screen).	1083		// Motion lands on wire row 3 (0-based row 2 of the current screen).

A internal/usecase/daemon/lease_callback_guard_test.go +67 -0

@@ -0,0 +1,67 @@
	1	+	package daemon
	2	+
	3	+	import (
	4	+	"runtime"
	5	+	"testing"
	6	+	"time"
	7	+
	8	+	"github.com/stretchr/testify/require"
	9	+
	10	+	"github.com/bnema/vev/internal/domain"
	11	+	)
	12	+
	13	+	func TestResizeAndRetryCallbacksRejectSameObjectLeaseReplacement(t *testing.T) {
	14	+	for _, tc := range []struct {
	15	+	name string
	16	+	schedule func(renderCoordinator, attachedClient, *attachmentLease, func())
	17	+	done func(testing.T, renderCoordinator) <-chan struct{}
	18	+	}{
	19	+	{
	20	+	name: "resize",
	21	+	schedule: func(rc renderCoordinator, ac attachedClient, lease *attachmentLease, run func()) {
	22	+	rc.scheduleResizeForLease(domain.Size{Cols: 100, Rows: 30}, ac, lease, func(uint64) { run() })
	23	+	},
	24	+	done: captureResizeCallbackDone,
	25	+	},
	26	+	{
	27	+	name: "retry",
	28	+	schedule: func(rc renderCoordinator, ac attachedClient, lease *attachmentLease, run func()) {
	29	+	epoch := rc.recordResizeRequestForLease(domain.Size{Cols: 100, Rows: 30}, ac, lease)
	30	+	require.NotZero(t, epoch)
	31	+	require.True(t, rc.resizeCurrentForLease(epoch, ac, lease, true))
	32	+	rc.scheduleResizeRetryForLease(epoch, ac, lease, run)
	33	+	},
	34	+	done: func(t testing.T, rc renderCoordinator) <-chan struct{} {
	35	+	t.Helper()
	36	+	rc.mu.Lock()
	37	+	defer rc.mu.Unlock()
	38	+	if rc.retryLane.token == nil {
	39	+	t.Fatal("coordinator did not publish a retry callback completion")
	40	+	}
	41	+	return rc.retryLane.token.done
	42	+	},
	43	+	},
	44	+	} {
	45	+	t.Run(tc.name, func(t *testing.T) {
	46	+	clock := newCoordinatorMockClock(t, 1)
	47	+	rc := newRenderCoordinator(renderCoordinatorOptions{clock: clock.clock})
	48	+	ac := &attachedClient{}
	49	+	rc.attach(ac)
	50	+	lease := rc.attachmentLease(ac)
	51	+	ran := false
	52	+	tc.schedule(rc, ac, lease, func() { ran = true })
	53	+	timer := awaitCoordinatorScheduledTimer(t, clock)
	54	+	done := tc.done(t, rc)
	55	+	require.NotNil(t, done)
	56	+
	57	+	rc.notePark(ac)
	58	+	rc.attach(ac) // resume the exact same object under a new incarnation
	59	+	timer.ch <- time.Time{}
	60	+	awaitTestCompletion(t, done, "stale callback did not complete")
	61	+	for range 32 {
	62	+	runtime.Gosched()
	63	+	}
	64	+	require.False(t, ran, "stale %s callback ran after lease replacement", tc.name)
	65	+	})
	66	+	}
	67	+	}

M internal/usecase/daemon/mouse_panes.go +0 -4


@@ -64,10 +64,6 @@ func solveTabLayoutLocked(tb *tab) tabLayoutSnapshot {
64		return tabLayoutSnapshot{root: tb.tree.Root, fingerprint: layoutFingerprint(tb.tree.Root), area: area, focus: tb.tree.Focus, placements: placements, ok: ok}	64	return tabLayoutSnapshot{root: tb.tree.Root, fingerprint: layoutFingerprint(tb.tree.Root), area: area, focus: tb.tree.Focus, placements: placements, ok: ok}
65		}	65	}
66			66
67	-	func (snap tabLayoutSnapshot) matchesLocked(tb *tab) bool {
68	-	return snap.ok && tb != nil && tb.tree != nil && tb.tree.Root == snap.root && tb.tree.Focus == snap.focus && tb.size.Cols == snap.area.Width && tb.size.Rows == snap.area.Height && snap.fingerprint == layoutFingerprint(tb.tree.Root)
69	-	}
70	-
71		func layoutFingerprint(root *layout.Node) string {	67	func layoutFingerprint(root *layout.Node) string {
72		var b strings.Builder	68	var b strings.Builder
73		writeLayoutFingerprint(&b, root)	69	writeLayoutFingerprint(&b, root)

M internal/usecase/daemon/nil_pty_test.go +1 -1


@@ -24,7 +24,7 @@ func TestNilPTYLifecyclePathsDoNotPanic(t *testing.T) {
24		sess.tabs = []*tab{first, second}	24		sess.tabs = []*tab{first, second}
25			25
26		require.NotPanics(t, func() {	26		require.NotPanics(t, func() {
27	-	d.resizePane(first.focusedPane(), domain.Rect{Width: 20, Height: 5})	27	+	d.applyPaneResize(first.focusedPane(), domain.Rect{Width: 20, Height: 5})
28		})	28		})
29		require.Equal(t, 20, first.focusedPane().screen.Frame.Width)	29		require.Equal(t, 20, first.focusedPane().screen.Frame.Width)
30		require.Equal(t, 5, first.focusedPane().screen.Frame.Height)	30		require.Equal(t, 5, first.focusedPane().screen.Frame.Height)

M internal/usecase/daemon/output_state_stream.go +0 -29


@@ -17,8 +17,6 @@ type outputStateStream struct {
17		renderer *renderer.Renderer	17	renderer *renderer.Renderer
18		next uint64	18	next uint64
19		acked uint64	19	acked uint64
20	-	deferred bool
21	-	deferredReset bool
22		maxOutstanding uint64	20	maxOutstanding uint64
23		}	21	}
24			22

@@ -89,37 +87,10 @@ func (s *outputStateStream) ack(state uint64) {
89		func (s *outputStateStream) rebase() {	87	func (s *outputStateStream) rebase() {
90		s.acked = s.next	88	s.acked = s.next
91		s.renderer.Reset()	89	s.renderer.Reset()
92	-	s.clearDeferred()
93		}	90	}
94			91
95		func (s *outputStateStream) atCapacity() bool {	92	func (s *outputStateStream) atCapacity() bool {
96		return s.outstanding() >= s.maxOutstanding	93	return s.outstanding() >= s.maxOutstanding
97		}	94	}
98			95
99	-	func (s *outputStateStream) deferIfAtCapacity(reset bool) bool {
100	-	if !s.atCapacity() {
101	-	if reset {
102	-	s.clearDeferred()
103	-	}
104	-	return false
105	-	}
106	-	s.deferred = true
107	-	s.deferredReset = s.deferredReset \|\| reset
108	-	return true
109	-	}
110	-
111	-	func (s *outputStateStream) takeDeferred() (reset bool, ok bool) {
112	-	if !s.deferred \|\| s.atCapacity() {
113	-	return false, false
114	-	}
115	-	reset = s.deferredReset
116	-	s.clearDeferred()
117	-	return reset, true
118	-	}
119	-
120	-	func (s *outputStateStream) clearDeferred() {
121	-	s.deferred = false
122	-	s.deferredReset = false
123	-	}
124	-
125		func (s *outputStateStream) outstanding() uint64 { return s.next - s.acked }	96	func (s *outputStateStream) outstanding() uint64 { return s.next - s.acked }

M internal/usecase/daemon/output_state_stream_test.go +0 -75


@@ -68,53 +68,6 @@ func TestOutputStateStreamRepeatedUnackedScrollRemainsClientCorrect(t *testing.T
68		}	68	}
69		}	69	}
70			70
71	-	func TestOutputStateStreamAckProgressionReleasesDeferredPaint(t *testing.T) {
72	-	stream := newOutputStateStream()
73	-	for range maxUnackedOutputStates {
74	-	stream.frame([]byte("output"), false, 0)
75	-	}
76	-	require.True(t, stream.deferIfAtCapacity(false))
77	-	_, ok := stream.takeDeferred()
78	-	require.False(t, ok)
79	-	stream.ack(maxUnackedOutputStates - 1)
80	-	reset, ok := stream.takeDeferred()
81	-	require.True(t, ok)
82	-	require.False(t, reset)
83	-	require.Equal(t, uint64(1), stream.outstanding())
84	-
85	-	// Stale and future ACKs do not move the cumulative acknowledgement.
86	-	stream.ack(maxUnackedOutputStates - 2)
87	-	stream.ack(maxUnackedOutputStates + 1)
88	-	require.Equal(t, uint64(maxUnackedOutputStates-1), stream.acked)
89	-	}
90	-
91	-	func TestOutputStateStreamResetIsDependencyFreeAndClearsDeferral(t *testing.T) {
92	-	stream := newOutputStateStream()
93	-	for range maxUnackedOutputStates {
94	-	stream.frame([]byte("output"), false, 0)
95	-	}
96	-	require.True(t, stream.deferIfAtCapacity(false))
97	-	for range 32 {
98	-	require.True(t, stream.deferIfAtCapacity(true), "reset paints must coalesce while the stream is at capacity")
99	-	}
100	-	require.Equal(t, uint64(maxUnackedOutputStates), stream.outstanding())
101	-	stream.ack(maxUnackedOutputStates)
102	-	reset, ok := stream.takeDeferred()
103	-	require.True(t, ok)
104	-	require.True(t, reset, "a coalesced reset must flush as a dependency-free repaint")
105	-
106	-	frame := renderer.NewFrame(2, 1)
107	-	fillOutputStateRows(frame, []string{"ok"})
108	-	outFrame, ok, err := drawOutputState(t, stream, frame, nil, true, 0)
109	-	require.NoError(t, err)
110	-	require.True(t, ok)
111	-	out, err := ports.UnmarshalOutput(outFrame.Payload)
112	-	require.NoError(t, err)
113	-	require.Zero(t, out.BaseStateNum)
114	-	_, ok = stream.takeDeferred()
115	-	require.False(t, ok)
116	-	}
117	-
118		func drawOutputState(t testing.T, stream outputStateStream, frame renderer.Frame, damage []renderer.Damage, reset bool, echoAck uint64) (ports.Frame, bool, error) {	71	func drawOutputState(t testing.T, stream outputStateStream, frame renderer.Frame, damage []renderer.Damage, reset bool, echoAck uint64) (ports.Frame, bool, error) {
119		t.Helper()	72	t.Helper()
120		data, err := stream.render(frame, damage, reset)	73	data, err := stream.render(frame, damage, reset)

@@ -124,34 +77,6 @@ func drawOutputState(t testing.T, stream outputStateStream, frame renderer.Fra
124		return stream.frame(data, reset, echoAck), true, nil	77	return stream.frame(data, reset, echoAck), true, nil
125		}	78	}
126			79
127	-	func TestOutputStateStreamWindowOneCoalescesUntilCumulativeAck(t *testing.T) {
128	-	stream := newOutputStateStream(1)
129	-	stream.frame([]byte("first"), true, 0)
130	-	for range 100 {
131	-	require.True(t, stream.deferIfAtCapacity(false))
132	-	}
133	-	require.Equal(t, uint64(1), stream.outstanding())
134	-
135	-	// A stale ACK cannot release the coalesced paint; the cumulative ACK for
136	-	// the actually sent state does.
137	-	stream.ack(0)
138	-	_, ok := stream.takeDeferred()
139	-	require.False(t, ok)
140	-	stream.ack(1)
141	-	reset, ok := stream.takeDeferred()
142	-	require.True(t, ok)
143	-	require.False(t, reset)
144	-	stream.frame([]byte("latest"), reset, 0)
145	-	require.Equal(t, uint64(1), stream.outstanding())
146	-
147	-	sideEffect := stream.sideEffect([]byte("reliable"), 0)
148	-	output, err := ports.UnmarshalOutput(sideEffect.Payload)
149	-	require.NoError(t, err)
150	-	require.Zero(t, output.BaseStateNum)
151	-	require.Zero(t, output.NewStateNum)
152	-	require.Equal(t, uint64(2), stream.next)
153	-	}
154	-
155		func TestOutputStateStreamDefaultsAndNormalizesWindow(t *testing.T) {	80	func TestOutputStateStreamDefaultsAndNormalizesWindow(t *testing.T) {
156		require.Equal(t, uint64(8), newOutputStateStream().maxOutstanding)	81	require.Equal(t, uint64(8), newOutputStateStream().maxOutstanding)
157		require.Equal(t, uint64(8), newOutputStateStream(0).maxOutstanding)	82	require.Equal(t, uint64(8), newOutputStateStream(0).maxOutstanding)

M internal/usecase/daemon/output_state_sync_test.go +52 -25


@@ -82,7 +82,7 @@ func TestDatagramAttachPipelinesRendererBeforeAck(t *testing.T) {
82		require.Same(t, sess, routed)	82		require.Same(t, sess, routed)
83			83
84		sess.tabs[0].focusedPane().screen.Write([]byte("A"))	84		sess.tabs[0].focusedPane().screen.Write([]byte("A"))
85	-	d.paint(sess, ac, false)	85	+	d.paint(sess, ac, false, nil)
86		first := awaitFrame(t, tr.sends, ports.MsgOutput)	86		first := awaitFrame(t, tr.sends, ports.MsgOutput)
87		out, err := ports.UnmarshalOutput(first.Payload)	87		out, err := ports.UnmarshalOutput(first.Payload)
88		require.NoError(t, err)	88		require.NoError(t, err)

@@ -90,7 +90,7 @@ func TestDatagramAttachPipelinesRendererBeforeAck(t *testing.T) {
90			90
91		// Before the MsgAck, the renderer has already advanced along the ordered	91		// Before the MsgAck, the renderer has already advanced along the ordered
92		// output dependency chain, so an unchanged repaint is a no-op.	92		// output dependency chain, so an unchanged repaint is a no-op.
93	-	d.paint(sess, ac, false)	93	+	d.paint(sess, ac, false, nil)
94		requireNoOutputFrame(t, tr.sends)	94		requireNoOutputFrame(t, tr.sends)
95			95
96		tr.recv <- ports.Frame{Type: ports.MsgAck, Payload: ports.MarshalAck(ports.Ack{AckedStateNum: out.NewStateNum})}	96		tr.recv <- ports.Frame{Type: ports.MsgAck, Payload: ports.MarshalAck(ports.Ack{AckedStateNum: out.NewStateNum})}

@@ -99,7 +99,7 @@ func TestDatagramAttachPipelinesRendererBeforeAck(t *testing.T) {
99			99
100		// The production MsgAck path retires retained states without moving the	100		// The production MsgAck path retires retained states without moving the
101		// renderer baseline backward.	101		// renderer baseline backward.
102	-	d.paint(sess, ac, false)	102	+	d.paint(sess, ac, false, nil)
103		requireNoOutputFrame(t, tr.sends)	103		requireNoOutputFrame(t, tr.sends)
104		}	104		}
105			105

@@ -111,7 +111,7 @@ func TestPaintExplicitlyUsesAsyncTransportCapability(t *testing.T) {
111		ac.replaceTransport(tr)	111		ac.replaceTransport(tr)
112		sess.tabs[0].focusedPane().screen.Write([]byte("A"))	112		sess.tabs[0].focusedPane().screen.Write([]byte("A"))
113			113
114	-	d.paint(sess, ac, false)	114	+	d.paint(sess, ac, false, nil)
115			115
116		awaitFrame(t, tr.asyncSends, ports.MsgOutput)	116		awaitFrame(t, tr.asyncSends, ports.MsgOutput)
117		select {	117		select {

@@ -144,7 +144,7 @@ func TestDatagramMultipleUnackedScrollPaintsMatchLatestFrame(t *testing.T) {
144		}	144		}
145		pane.screen.ClearDamage()	145		pane.screen.ClearDamage()
146		client := vt.NewScreen(80, 25)	146		client := vt.NewScreen(80, 25)
147	-	d.paint(sess, ac, true)	147	+	d.paint(sess, ac, true, nil)
148		first := mustApplyOutput(t, client, awaitFrame(t, tr.sends, ports.MsgOutput))	148		first := mustApplyOutput(t, client, awaitFrame(t, tr.sends, ports.MsgOutput))
149		ac.ackOutputState(first.NewStateNum)	149		ac.ackOutputState(first.NewStateNum)
150			150

@@ -164,7 +164,7 @@ func TestDatagramMultipleUnackedScrollPaintsMatchLatestFrame(t *testing.T) {
164		pane.screen.Col = 0	164		pane.screen.Col = 0
165		pane.screen.Write([]byte("\nq"))	165		pane.screen.Write([]byte("\nq"))
166		pane.screen.Frame = desired.Clone()	166		pane.screen.Frame = desired.Clone()
167	-	d.paint(sess, ac, false)	167	+	d.paint(sess, ac, false, nil)
168		mustApplyOutput(t, client, awaitFrame(t, tr.sends, ports.MsgOutput))	168		mustApplyOutput(t, client, awaitFrame(t, tr.sends, ports.MsgOutput))
169		}	169		}
170			170

@@ -202,11 +202,11 @@ func TestLocalAttachStillAdvancesRendererOnSend(t *testing.T) {
202		d, sess, ac, sends := newManualSessionWithPTYs(t, p)	202		d, sess, ac, sends := newManualSessionWithPTYs(t, p)
203			203
204		sess.tabs[0].focusedPane().screen.Write([]byte("A"))	204		sess.tabs[0].focusedPane().screen.Write([]byte("A"))
205	-	d.paint(sess, ac, false)	205	+	d.paint(sess, ac, false, nil)
206		awaitFrame(t, sends, ports.MsgOutput)	206		awaitFrame(t, sends, ports.MsgOutput)
207			207
208		// The first paint updates the renderer shadow without waiting for MsgAck.	208		// The first paint updates the renderer shadow without waiting for MsgAck.
209	-	d.paint(sess, ac, false)	209	+	d.paint(sess, ac, false, nil)
210		requireNoOutputFrame(t, sends)	210		requireNoOutputFrame(t, sends)
211		}	211		}
212			212

@@ -217,17 +217,17 @@ func TestLocalOutputAckDoesNotMoveRendererShadowBackward(t *testing.T) {
217		pane := sess.tabs[0].focusedPane()	217		pane := sess.tabs[0].focusedPane()
218			218
219		pane.screen.Write([]byte("A"))	219		pane.screen.Write([]byte("A"))
220	-	d.paint(sess, ac, false)	220	+	d.paint(sess, ac, false, nil)
221		first := awaitFrame(t, sends, ports.MsgOutput)	221		first := awaitFrame(t, sends, ports.MsgOutput)
222		out, err := ports.UnmarshalOutput(first.Payload)	222		out, err := ports.UnmarshalOutput(first.Payload)
223		require.NoError(t, err)	223		require.NoError(t, err)
224			224
225		pane.screen.Write([]byte("\rB"))	225		pane.screen.Write([]byte("\rB"))
226	-	d.paint(sess, ac, false)	226	+	d.paint(sess, ac, false, nil)
227		awaitFrame(t, sends, ports.MsgOutput)	227		awaitFrame(t, sends, ports.MsgOutput)
228			228
229		ac.ackOutputState(out.NewStateNum)	229		ac.ackOutputState(out.NewStateNum)
230	-	d.paint(sess, ac, false)	230	+	d.paint(sess, ac, false, nil)
231		requireNoOutputFrame(t, sends)	231		requireNoOutputFrame(t, sends)
232		}	232		}
233			233

@@ -264,7 +264,7 @@ func TestFailedPaintSendRollsBackOutputState(t *testing.T) {
264		ac.replaceTransport(failingOutputTransport{})	264		ac.replaceTransport(failingOutputTransport{})
265		sess.tabs[0].focusedPane().screen.Write([]byte("A"))	265		sess.tabs[0].focusedPane().screen.Write([]byte("A"))
266			266
267	-	d.paint(sess, ac, false)	267	+	d.paint(sess, ac, false, nil)
268			268
269		require.Zero(t, ac.output.next)	269		require.Zero(t, ac.output.next)
270		require.Zero(t, ac.output.acked)	270		require.Zero(t, ac.output.acked)

@@ -279,30 +279,50 @@ func TestResizeGrowthFirstFrameIncludesConcurrentPTYRedraw(t *testing.T) {
279		{name: "stream window", window: maxUnackedOutputStates},	279		{name: "stream window", window: maxUnackedOutputStates},
280		} {	280		} {
281		t.Run(tc.name, func(t *testing.T) {	281		t.Run(tc.name, func(t *testing.T) {
282	-	p, releasePTY := newBlockingPTY(t)	282	+	// resizeReaderPTY delivers the child's SIGWINCH redraw through the
283	-	defer releasePTY()	283	+	// actual ptyReader. Its Resize callback does not return until Read has
			284	+	// accepted the bytes, at which point resizeApplying is still true.
			285	+	p := newResizeReaderPTY([]byte("\x1b[1;81H" + strings.Repeat("B", 40)))
284		d, sess, ac, sends := newManualSessionWithPTYs(t, p)	286		d, sess, ac, sends := newManualSessionWithPTYs(t, p)
285		ac.output.maxOutstanding = tc.window	287		ac.output.maxOutstanding = tc.window
286		pane := sess.activeTab().focusedPane()	288		pane := sess.activeTab().focusedPane()
287	-		289	+	p.applying = func() bool {
			290	+	pane.mu.Lock()
			291	+	defer pane.mu.Unlock()
			292	+	return pane.resizeApplying
			293	+	}
288		pane.screen.Write([]byte(strings.Repeat("A", 79)))	294		pane.screen.Write([]byte(strings.Repeat("A", 79)))
289	-	d.paint(sess, ac, true)	295	+	d.paint(sess, ac, true, nil)
290		client := vt.NewScreen(80, 24)	296		client := vt.NewScreen(80, 24)
291		initial := mustApplyOutput(t, client, awaitFrame(t, sends, ports.MsgOutput))	297		initial := mustApplyOutput(t, client, awaitFrame(t, sends, ports.MsgOutput))
292		ac.ackOutputState(initial.NewStateNum)	298		ac.ackOutputState(initial.NewStateNum)
293			299
294	-	// The child can redraw the newly exposed columns while the resize	300	+	d.sessWg.Add(1)
295	-	// output is being prepared. The first visible grown frame must contain	301	+	pane.onExit = func() {}
296	-	// that redraw, regardless of the negotiated output window.	302	+	go d.ptyReader(sess, sess.activeTab(), pane)
			303	+	defer func() { p.close(); d.sessWg.Wait() }()
			304	+
			305	+	// The resize deadline drives coordinator prepare/apply/commit. No test
			306	+	// writes the grown columns directly.
			307	+	clock := &signalClock{timers: make(chan *signalTimer, 2)}
			308	+	d.clock = clock
297		d.resize(sess, ac, domain.Size{Cols: 120, Rows: 24})	309		d.resize(sess, ac, domain.Size{Cols: 120, Rows: 24})
298	-	pane.screen.Write([]byte("\x1b[1;81H" + strings.Repeat("B", 40)))	310	+	timer := <-clock.timers
299	-	d.paint(sess, ac, false)	311	+	done := captureResizeCallbackDone(t, sess.renderCoordinator())
			312	+	timer.ch <- time.Time{}
			313	+	awaitTestCompletion(t, done, "resize callback did not complete")
			314	+	require.True(t, p.deliveredWhileApplying(), "redraw must arrive while PTY.Resize owns the apply gate")
			315	+	require.Equal(t, 'B', pane.screen.Frame.At(100, 0).Rune, "replay must parse the redraw before commit")
300			316
301		client.Resize(120, 24)	317		client.Resize(120, 24)
302		mustApplyOutput(t, client, awaitFrame(t, sends, ports.MsgOutput))	318		mustApplyOutput(t, client, awaitFrame(t, sends, ports.MsgOutput))
303		require.Equal(t, 'B', client.Frame.At(100, 1).Rune,	319		require.Equal(t, 'B', client.Frame.At(100, 1).Rune,
304		"first grown frame exposed stale pre-SIGWINCH pane content")	320		"first grown frame exposed stale pre-SIGWINCH pane content")
305	-	requireNoOutputFrame(t, sends)	321	+	select {
			322	+	case extra := <-sends:
			323	+	t.Fatalf("resize emitted more than one frame: %#v", extra)
			324	+	default:
			325	+	}
306		})	326		})
307		}	327		}
308		}	328		}

@@ -323,7 +343,9 @@ func TestResizeWithoutPTYOutputFlushesOneFullFrameAtDeadline(t *testing.T) {
323		}	343		}
324		requireNoOutputFrame(t, sends)	344		requireNoOutputFrame(t, sends)
325			345
			346	+	done := captureResizeCallbackDone(t, sess.renderCoordinator())
326		timer.ch <- time.Now()	347		timer.ch <- time.Now()
			348	+	awaitTestCompletion(t, done, "resize callback did not complete")
327		frame := awaitFrame(t, sends, ports.MsgOutput)	349		frame := awaitFrame(t, sends, ports.MsgOutput)
328		client := vt.NewScreen(120, 24)	350		client := vt.NewScreen(120, 24)
329		out := mustApplyOutput(t, client, frame)	351		out := mustApplyOutput(t, client, frame)

@@ -348,6 +370,7 @@ func TestResizeBurstFlushesOnlyLatestGeometry(t *testing.T) {
348		case <-time.After(time.Second):	370		case <-time.After(time.Second):
349		t.Fatal("first resize did not schedule a bounded paint")	371		t.Fatal("first resize did not schedule a bounded paint")
350		}	372		}
			373	+	firstDone := captureResizeCallbackDone(t, sess.renderCoordinator())
351		d.resize(sess, ac, domain.Size{Cols: 120, Rows: 24})	374		d.resize(sess, ac, domain.Size{Cols: 120, Rows: 24})
352		var latest *signalTimer	375		var latest *signalTimer
353		select {	376		select {

@@ -356,9 +379,12 @@ func TestResizeBurstFlushesOnlyLatestGeometry(t *testing.T) {
356		t.Fatal("latest resize did not replace the bounded paint")	379		t.Fatal("latest resize did not replace the bounded paint")
357		}	380		}
358			381
			382	+	latestDone := captureResizeCallbackDone(t, sess.renderCoordinator())
359		first.ch <- time.Now()	383		first.ch <- time.Now()
			384	+	awaitTestCompletion(t, firstDone, "obsolete resize callback did not complete")
360		requireNoOutputFrame(t, sends)	385		requireNoOutputFrame(t, sends)
361		latest.ch <- time.Now()	386		latest.ch <- time.Now()
			387	+	awaitTestCompletion(t, latestDone, "latest resize callback did not complete")
362		awaitFrame(t, sends, ports.MsgOutput)	388		awaitFrame(t, sends, ports.MsgOutput)
363		require.Equal(t, domain.Size{Cols: 120, Rows: 24}, ac.size)	389		require.Equal(t, domain.Size{Cols: 120, Rows: 24}, ac.size)
364		require.Equal(t, 120, sess.activeTab().focusedPane().screen.Frame.Width)	390		require.Equal(t, 120, sess.activeTab().focusedPane().screen.Frame.Width)

@@ -378,20 +404,21 @@ func TestDatagramWindowOneCoalescesPaintsUntilMsgAck(t *testing.T) {
378		MaxOutputInFlight: 1,	404		MaxOutputInFlight: 1,
379		}, tr)	405		}, tr)
380		require.NoError(t, err)	406		require.NoError(t, err)
			407	+	require.True(t, sess.renderCoordinator().markAttachmentReady(sess.renderCoordinator().attachmentLease(ac)))
381			408
382		pane := sess.tabs[0].focusedPane()	409		pane := sess.tabs[0].focusedPane()
383		pane.screen.Write([]byte("A"))	410		pane.screen.Write([]byte("A"))
384	-	d.paint(sess, ac, false)	411	+	d.paint(sess, ac, false, nil)
385		first := awaitFrame(t, tr.sends, ports.MsgOutput)	412		first := awaitFrame(t, tr.sends, ports.MsgOutput)
386		firstOutput, err := ports.UnmarshalOutput(first.Payload)	413		firstOutput, err := ports.UnmarshalOutput(first.Payload)
387		require.NoError(t, err)	414		require.NoError(t, err)
388		require.Equal(t, uint64(1), firstOutput.NewStateNum)	415		require.Equal(t, uint64(1), firstOutput.NewStateNum)
389		for range 100 {	416		for range 100 {
390		pane.screen.Write([]byte("x"))	417		pane.screen.Write([]byte("x"))
391	-	d.paint(sess, ac, false)	418	+	d.invalidateRender(sess, ac, false, "output_state_sync_test.go")
392		}	419		}
393		pane.screen.Write([]byte("LATEST"))	420		pane.screen.Write([]byte("LATEST"))
394	-	d.paint(sess, ac, false)	421	+	d.invalidateRender(sess, ac, false, "output_state_sync_test.go")
395		requireNoOutputFrame(t, tr.sends)	422		requireNoOutputFrame(t, tr.sends)
396		require.Equal(t, uint64(1), ac.output.outstanding(), "only one state-bearing datagram output may be unacknowledged")	423		require.Equal(t, uint64(1), ac.output.outstanding(), "only one state-bearing datagram output may be unacknowledged")
397			424

M internal/usecase/daemon/overlay_runtime.go +7 -5


@@ -13,11 +13,13 @@ import (
13		type overlayRuntime struct {	13		type overlayRuntime struct {
14		ac *attachedClient	14		ac *attachedClient
15			15
16	-	pickerMu sync.Mutex	16	+	pickerMu sync.Mutex
17	-	picker *picker.Model	17	+	picker *picker.Model
18	-	pickerPreview *tab	18	+	pickerPreview *tab
19	-	pickerPending []byte	19	+	pickerPreviewSession *session
20	-	pickerESC pendingByteTimer	20	+	pickerPreviewGeneration uint64
			21	+	pickerPending []byte
			22	+	pickerESC pendingByteTimer
21			23
22		paletteMu sync.Mutex	24		paletteMu sync.Mutex
23		palette *palette.Model	25		palette *palette.Model

M internal/usecase/daemon/overlay_runtime_test.go +1 -1


@@ -173,7 +173,7 @@ func TestOverlayRuntimeSnapshotCarriesCopyDocumentAndClearReleasesIt(t *testing.
173		func TestOverlayRuntimeCopyDocumentSurvivesRingOverwrite(t *testing.T) {	173		func TestOverlayRuntimeCopyDocumentSurvivesRingOverwrite(t *testing.T) {
174		ac := &attachedClient{}	174		ac := &attachedClient{}
175		ac.initOverlays()	175		ac.initOverlays()
176	-	history := scopy.NewScrollback(1)	176	+	history := newTestHistory(1)
177		history.Append(testRow("before"))	177		history.Append(testRow("before"))
178		document := scopy.NewSnapshot(history, renderer.NewFrame(6, 1))	178		document := scopy.NewSnapshot(history, renderer.NewFrame(6, 1))
179			179

M internal/usecase/daemon/paint_gate_test.go +0 -112

@@ -1,14 +1,10 @@
1		package daemon	1	package daemon
2			2
3		import (	3	import (
4	-	"io"
5		"testing"	4	"testing"
6		"time"	5	"time"
7			6
8		"github.com/bnema/vev/internal/ports"	7	"github.com/bnema/vev/internal/ports"
9	-	portsmocks "github.com/bnema/vev/internal/ports/mocks"
10	-	"github.com/bnema/vev/pkg/vt"
11	-	"github.com/stretchr/testify/require"
12		)	8	)
13			9
14		func requireNoOutputFrame(t *testing.T, sends chan ports.Frame) {	10	func requireNoOutputFrame(t *testing.T, sends chan ports.Frame) {

@@ -25,111 +21,3 @@ func requireNoOutputFrame(t *testing.T, sends chan ports.Frame) {
25		}	21	}
26		}	22	}
27		}	23	}
28	-
29	-	func TestPaintDefersWhenOutputAckLagsAtCap(t *testing.T) {
30	-	p, _ := newBlockingPTY(t)
31	-	d, sess, ac, sends := newManualSessionWithPTYs(t, p)
32	-	sess.tabs[0].focusedPane().screen.Write([]byte("A"))
33	-
34	-	// Client is maxUnackedOutputStates behind (outputAck stays 0).
35	-	ac.output.next = maxUnackedOutputStates
36	-	d.paint(sess, ac, false)
37	-
38	-	requireNoOutputFrame(t, sends)
39	-	ac.sendMu.Lock()
40	-	deferred := ac.output.deferred
41	-	ac.sendMu.Unlock()
42	-	require.True(t, deferred, "diff paint at ack cap must defer")
43	-	}
44	-
45	-	func TestDeferredPaintFlushesOnceWhenAcksCatchUp(t *testing.T) {
46	-	p, _ := newBlockingPTY(t)
47	-	d, sess, ac, sends := newManualSessionWithPTYs(t, p)
48	-	sess.tabs[0].focusedPane().screen.Write([]byte("A"))
49	-
50	-	ac.output.next = maxUnackedOutputStates
51	-	d.paint(sess, ac, false)
52	-	requireNoOutputFrame(t, sends)
53	-
54	-	// Drive a real MsgAck through the connection loop so the flush is exercised
55	-	// end-to-end: UnmarshalAck → advanceOutputAck → takeDeferredPaint → paint.
56	-	// One ack frame, then Recv blocks until the test ends.
57	-	recvCh := make(chan ports.Frame, 1)
58	-	done := make(chan struct{})
59	-	t.Cleanup(func() { close(done) })
60	-	tr, ok := ac.transport().(*portsmocks.MockTransport)
61	-	require.True(t, ok, "manual session transport must be the capturing mock")
62	-	tr.EXPECT().Recv().RunAndReturn(func() (ports.Frame, error) {
63	-	select {
64	-	case f := <-recvCh:
65	-	return f, nil
66	-	case <-done:
67	-	return ports.Frame{}, io.EOF
68	-	}
69	-	}).Maybe()
70	-	recvCh <- ports.Frame{Type: ports.MsgAck, Payload: ports.MarshalAck(ports.Ack{AckedStateNum: maxUnackedOutputStates})}
71	-	go d.runConnLoop(ac)
72	-
73	-	// The cumulative flush paint must be triggered by the MsgAck handler itself
74	-	// (the renderer shadow only advanced on send, so the skip coalesces here).
75	-	awaitFrame(t, sends, ports.MsgOutput)
76	-	requireNoOutputFrame(t, sends)
77	-	ac.sendMu.Lock()
78	-	deferred := ac.output.deferred
79	-	ac.sendMu.Unlock()
80	-	require.False(t, deferred, "flag must clear after the flush paint")
81	-	}
82	-
83	-	func TestResetPaintCoalescesAtAckGate(t *testing.T) {
84	-	p, _ := newBlockingPTY(t)
85	-	d, sess, ac, sends := newManualSessionWithPTYs(t, p)
86	-	sess.tabs[0].focusedPane().screen.Write([]byte("A"))
87	-
88	-	// At the cap, repeated reset paints coalesce instead of growing retained
89	-	// state and transport queues without bound.
90	-	ac.output.next = maxUnackedOutputStates
91	-	for range 32 {
92	-	d.paint(sess, ac, true)
93	-	}
94	-
95	-	requireNoOutputFrame(t, sends)
96	-	ac.sendMu.Lock()
97	-	deferred := ac.output.deferred
98	-	deferredReset := ac.output.deferredReset
99	-	retained := ac.output.outstanding()
100	-	ac.sendMu.Unlock()
101	-	require.True(t, deferred)
102	-	require.True(t, deferredReset)
103	-	require.LessOrEqual(t, retained, uint64(maxUnackedOutputStates))
104	-	}
105	-
106	-	func TestCoalescedResetFlushesLatestFrameAfterAck(t *testing.T) {
107	-	p, _ := newBlockingPTY(t)
108	-	d, sess, ac, sends := newManualSessionWithPTYs(t, p)
109	-	for range maxUnackedOutputStates {
110	-	ac.output.frame([]byte("in flight"), false, 0)
111	-	}
112	-
113	-	pane := sess.tabs[0].focusedPane()
114	-	var latest byte
115	-	for i := range 32 {
116	-	latest = byte('a' + i%26)
117	-	pane.screen.Write([]byte{'\r', latest})
118	-	d.paint(sess, ac, true)
119	-	}
120	-	requireNoOutputFrame(t, sends)
121	-	require.Equal(t, uint64(maxUnackedOutputStates), ac.output.outstanding())
122	-
123	-	ac.ackOutputState(maxUnackedOutputStates)
124	-	reset, ok := ac.takeDeferredPaint()
125	-	require.True(t, ok)
126	-	require.True(t, reset)
127	-	d.paint(sess, ac, reset)
128	-
129	-	client := vt.NewScreen(80, 25)
130	-	out := mustApplyOutput(t, client, awaitFrame(t, sends, ports.MsgOutput))
131	-	require.Zero(t, out.BaseStateNum)
132	-	require.Equal(t, rune(latest), client.Frame.At(0, 1).Rune)
133	-	require.Equal(t, uint64(1), ac.output.outstanding())
134	-	requireNoOutputFrame(t, sends)
135	-	}

M internal/usecase/daemon/palette.go +9 -9


@@ -42,7 +42,7 @@ func (d Daemon) enterPalette(sess session, ac *attachedClient) {
42		ac.overlays.paletteHints = palette.ContextualHints{}	42		ac.overlays.paletteHints = palette.ContextualHints{}
43		ac.overlays.palettePending = nil	43		ac.overlays.palettePending = nil
44		ac.overlays.paletteMu.Unlock()	44		ac.overlays.paletteMu.Unlock()
45	-	d.paint(sess, ac, true)	45	+	d.invalidateRender(sess, ac, true, "palette.go")
46		}	46		}
47			47
48		func recentSessionHints(recent []recentSession, args []string) palette.ContextualHints {	48		func recentSessionHints(recent []recentSession, args []string) palette.ContextualHints {

@@ -167,12 +167,12 @@ func (d Daemon) handlePaletteInput(ac attachedClient, data []byte) {
167		}	167		}
168		ac.overlays.paletteMu.Unlock()	168		ac.overlays.paletteMu.Unlock()
169		if cancel {	169		if cancel {
170	-	d.paint(sess, ac, true)	170	+	d.invalidateRender(sess, ac, true, "palette.go")
171		return	171		return
172		}	172		}
173		if !execute {	173		if !execute {
174		if changed {	174		if changed {
175	-	d.paint(sess, ac, true)	175	+	d.invalidateRender(sess, ac, true, "palette.go")
176		}	176		}
177		return	177		return
178		}	178		}

@@ -183,7 +183,7 @@ func (d Daemon) handlePaletteInput(ac attachedClient, data []byte) {
183		rank, err := command.ParsePositiveDecimal(args)	183		rank, err := command.ParsePositiveDecimal(args)
184		if err != nil {	184		if err != nil {
185		ac.paletteFailure(generation, rawQuery, "rank must be one positive decimal")	185		ac.paletteFailure(generation, rawQuery, "rank must be one positive decimal")
186	-	d.paint(sess, ac, true)	186	+	d.invalidateRender(sess, ac, true, "palette.go")
187		return	187		return
188		}	188		}
189		// The captured ID is handed off atomically. A target can disappear after	189		// The captured ID is handed off atomically. A target can disappear after

@@ -191,14 +191,14 @@ func (d Daemon) handlePaletteInput(ac attachedClient, data []byte) {
191		exec := paletteExec{d: d, sess: sess, ac: ac, recent: recent}	191		exec := paletteExec{d: d, sess: sess, ac: ac, recent: recent}
192		if err := exec.JumpRecentSession(rank); err != nil {	192		if err := exec.JumpRecentSession(rank); err != nil {
193		ac.paletteFailure(generation, rawQuery, "requested recent session is unavailable")	193		ac.paletteFailure(generation, rawQuery, "requested recent session is unavailable")
194	-	d.paint(sess, ac, true)	194	+	d.invalidateRender(sess, ac, true, "palette.go")
195		return	195		return
196		}	196		}
197		// Publication remains generation-safe; do not let stale work close a	197		// Publication remains generation-safe; do not let stale work close a
198		// newer palette interaction.	198		// newer palette interaction.
199		if ac.closeExecutedPalette(generation, rawQuery) {	199		if ac.closeExecutedPalette(generation, rawQuery) {
200		d.recordPaletteUse(cmd.Code)	200		d.recordPaletteUse(cmd.Code)
201	-	d.paint(ac.currentSession(), ac, true)	201	+	d.invalidateRender(ac.currentSession(), ac, true, "palette.go")
202		}	202		}
203		return	203		return
204		}	204		}

@@ -259,7 +259,7 @@ type paletteExec struct {
259		}	259		}
260			260
261		func (e paletteExec) CreateTab() error {	261		func (e paletteExec) CreateTab() error {
262	-	defer e.d.paint(e.sess, e.ac, true)	262	+	defer e.d.invalidateRender(e.sess, e.ac, true, "palette.go")
263		return e.d.createTab(e.sess, e.ac.size)	263		return e.d.createTab(e.sess, e.ac.size)
264		}	264		}
265			265

@@ -325,7 +325,7 @@ func (e paletteExec) NextTab() error {
325		if e.sess.switchRelative(1) {	325		if e.sess.switchRelative(1) {
326		e.d.activateTab(e.sess, e.sess.activeTab())	326		e.d.activateTab(e.sess, e.sess.activeTab())
327		}	327		}
328	-	e.d.paint(e.sess, e.ac, true)	328	+	e.d.invalidateRender(e.sess, e.ac, true, "palette.go")
329		return nil	329		return nil
330		}	330		}
331			331

@@ -333,7 +333,7 @@ func (e paletteExec) PrevTab() error {
333		if e.sess.switchRelative(-1) {	333		if e.sess.switchRelative(-1) {
334		e.d.activateTab(e.sess, e.sess.activeTab())	334		e.d.activateTab(e.sess, e.sess.activeTab())
335		}	335		}
336	-	e.d.paint(e.sess, e.ac, true)	336	+	e.d.invalidateRender(e.sess, e.ac, true, "palette.go")
337		return nil	337		return nil
338		}	338		}
339			339

M internal/usecase/daemon/palette_test.go +2 -2


@@ -371,7 +371,7 @@ func TestPaletteCreateTabErrorRepaintsAfterClose(t *testing.T) {
371		d, sess, ac, sends, releases := newManualTabSession(t, 1)	371		d, sess, ac, sends, releases := newManualTabSession(t, 1)
372		defer releases[0]()	372		defer releases[0]()
373		ptys := portsmocks.NewMockPTYFactory(t)	373		ptys := portsmocks.NewMockPTYFactory(t)
374	-	ptys.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything).Return(nil, errors.New("open failed"))	374	+	ptys.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything).Return(nil, errors.New("open failed"))
375		d.ptys = ptys	375		d.ptys = ptys
376			376
377		d.handleInput(sess, ac, []byte("\x1b "))	377		d.handleInput(sess, ac, []byte("\x1b "))

@@ -518,7 +518,7 @@ func TestPaletteRenderAndInputCanRunConcurrently(t *testing.T) {
518		go func() {	518		go func() {
519		defer wg.Done()	519		defer wg.Done()
520		for range 200 {	520		for range 200 {
521	-	d.paint(sess, ac, true)	521	+	d.paint(sess, ac, true, nil)
522		}	522		}
523		}()	523		}()
524		go func() {	524		go func() {

M internal/usecase/daemon/pane.go +26 -32

 	"github.com/bnema/vev/internal/domain"
 	"github.com/bnema/vev/internal/ports"
-	scopy "github.com/bnema/vev/internal/usecase/copy"
 	"github.com/bnema/vev/internal/usecase/layout"
 	"github.com/bnema/vev/pkg/vt"
 )
-// pane owns one PTY, its terminal screen state, and render scheduling channels.
+// pane owns one PTY and its terminal screen state.
 // Lock order when multiple locks are held is:
 // attachedClient.sendMu > Daemon.mu > session.mu > tab.mu > pane.mu.
 // The PTY reader takes only pane.mu, so child output never waits on client IO.
 type pane struct {
-	id                layout.PaneID
+	id       layout.PaneID
-	stableID          string
+	stableID string
-	pty               ports.PTY
+	pty      ports.PTY
-	mu                sync.Mutex // guards screen, scrollback, syncGen, rect, and title
+	mu       sync.Mutex // guards screen, history, syncGen, rect, resizeApplying, resizePending, PTY side effects, and title
-	resizeMu          sync.Mutex // serializes PTY resizes without holding mu
+	resizeMu sync.Mutex // serializes PTY resizes without holding mu
-	screen            *vt.Screen
+	screen   *vt.Screen
-	scrollback        *scopy.Scrollback
+	history  *vt.History
-	dirty             chan struct{}
+	syncGen  uint64
-	flush             chan struct{}
+	rect     domain.Rect
-	syncGen           uint64
+	// resizeApplying gates VT parsing across PTY.Resize. The reader continues
-	rect              domain.Rect
+	// draining into resizePending so output is replayed against the target (or
+	// retained old) screen only after apply resolves.
+	resizeApplying    bool
+	resizePending     []byte
+	ptyResponses      []byte
+	ptyClipboards     []string
+	ptyAttention      bool
 	popupGeometry     floatingGeometry // last geometry committed after a successful floating resize
 	title             paneTitleState
 	ctx               context.Context
 }
 func newPaneWithStableID(id layout.PaneID, stableID string, pty ports.PTY, sz domain.Size) *pane {
-	sb := scopy.NewScrollback(defaultScrollbackRows)
+	screen := vt.NewScreenWithHistory(sz.Cols, sz.Rows, vt.HistoryConfig{MaxRows: defaultScrollbackRows})
-	screen := vt.NewScreen(sz.Cols, sz.Rows)
-	screen.OnLineEvicted = sb.Append
 	return &pane{
-		id:         id,
+		id:       id,
-		stableID:   stableID,
+		stableID: stableID,
-		pty:        pty,
+		pty:      pty,
-		screen:     screen,
+		screen:   screen,
-		scrollback: sb,
+		history:  screen.History(),
-		dirty:      make(chan struct{}, 1),
+		rect:     domain.Rect{Width: sz.Cols, Height: sz.Rows},
-		flush:      make(chan struct{}, 1),
+		title:    paneTitleState{displayFallback: "sh"},
-		rect:       domain.Rect{Width: sz.Cols, Height: sz.Rows},
-		title:      paneTitleState{displayFallback: "sh"},
 	}
 }
-func paneDone(p *pane) <-chan struct{} {
-	if p.ctx == nil {
-		return nil
-	}
-	return p.ctx.Done()
-}

M internal/usecase/daemon/pane_actions.go +15 -11

 ) error {
 	if d.ptys == nil {
 		if ac != nil {
-			d.paint(sess, ac, true)
+			d.invalidateRender(sess, ac, true, "pane_actions.go")
 		}
 		return nil
 	}
 		tb.mu.Unlock()
 		return fmt.Errorf("daemon: generating pane identity: %w", err)
 	}
-	pty, err := d.ptys.Open(d.shell, d.shellArgs, d.childEnv(name, tabStableID, paneStableID, term), cwd, rectSize(newRect))
+	pty, err := d.ptys.Open(sess.ctx, d.shell, d.shellArgs, d.childEnv(name, tabStableID, paneStableID, term), cwd, rectSize(newRect))
 	if err != nil {
 		d.log.Warn("pty spawn failed", "err", err, "session", name, "pane", newID, "kind", "pane")
 		tb.mu.Lock()
 	d.startPaneGoroutines(sess, tb, p)
 	markSnapshotDirty(sess)
 	if ac != nil {
-		d.paint(sess, ac, true)
+		d.invalidateRender(sess, ac, true, "pane_actions.go")
 	}
 	return nil
 }
 	placements, ok := layout.Solve(tb.tree.Root, domain.Rect{Width: tb.size.Cols, Height: tb.size.Rows})
 	if !ok {
 		for _, p := range tb.panes {
-			d.resizePane(p, domain.Rect{Width: tb.size.Cols, Height: tb.size.Rows})
+			d.applyPaneResize(p, domain.Rect{Width: tb.size.Cols, Height: tb.size.Rows})
 		}
 		return
 	}
 			continue
 		}
 		if p := tb.panes[pl.ID]; p != nil {
-			d.resizePane(p, pl.Content)
+			d.applyPaneResize(p, pl.Content)
 		}
 	}
 }
-func (d *Daemon) resizePane(p *pane, r domain.Rect) {
+func (d *Daemon) applyPaneResize(p *pane, r domain.Rect) {
 	if p == nil {
 		return
 	}
 func (d *Daemon) toggleStack(sess *session, ac *attachedClient) error {
 	if d.ptys == nil {
 		if ac != nil {
-			d.paint(sess, ac, true)
+			d.invalidateRender(sess, ac, true, "pane_actions.go")
 		}
 		return nil
 	}
 	if err == nil {
 		markSnapshotDirty(sess)
 		if ac != nil {
-			d.paint(sess, ac, true)
+			d.invalidateRender(sess, ac, true, "pane_actions.go")
 		}
 	}
 	return err
 func (d *Daemon) closeFocusedPane(sess *session, ac *attachedClient) error {
 	if d.ptys == nil {
 		if ac != nil {
-			d.paint(sess, ac, true)
+			d.invalidateRender(sess, ac, true, "pane_actions.go")
 		}
 		return nil
 	}
 	if ac != nil {
 		ac.overlays.clearCopyModeForPane(p)
+		ac.pruneCaptureFrames(p)
 	}
+	if rc := sess.renderCoordinator(); rc != nil {
+		rc.noteSyncPaneRemoved(p)
+	}
 	if p.cancel != nil {
 		p.cancel()
 	}
 	markSnapshotDirty(sess)
 	if repaint {
 		if ac != nil {
-			d.paint(sess, ac, true)
+			d.invalidateRender(sess, ac, true, "pane_actions.go")
 		}
 	}
 	return nil
 				d.refreshPaneTitleOnFocus(sess, newFocus)
 			}
 		}
-		d.paint(sess, ac, true)
+		d.invalidateRender(sess, ac, true, "pane_actions.go")
 	}
 	if errors.Is(err, layout.ErrNoPane) {
 		return nil

M internal/usecase/daemon/pane_actions_test.go +9 -9


@@ -39,7 +39,7 @@ func TestSplitPaneCreatesFocusedShellInRequestedPosition(t *testing.T) {
39		newPTY := portsmocks.NewMockPTY(t)	39		newPTY := portsmocks.NewMockPTY(t)
40		oldPTY.EXPECT().Resize(rectSize(tt.wantOld)).Return(nil).Once()	40		oldPTY.EXPECT().Resize(rectSize(tt.wantOld)).Return(nil).Once()
41		newPTY.EXPECT().Read(mock.Anything).RunAndReturn(blockingRead(t)).Maybe()	41		newPTY.EXPECT().Read(mock.Anything).RunAndReturn(blockingRead(t)).Maybe()
42	-	factory.EXPECT().Open("/bin/sh", []string(nil), mock.Anything, "/work", rectSize(tt.wantNew)).Return(newPTY, nil).Once()	42	+	factory.EXPECT().Open(mock.Anything, "/bin/sh", []string(nil), mock.Anything, "/work", rectSize(tt.wantNew)).Return(newPTY, nil).Once()
43			43
44		require.NoError(t, d.splitPane(sess, nil, tt.dir))	44		require.NoError(t, d.splitPane(sess, nil, tt.dir))
45			45

@@ -68,7 +68,7 @@ func TestSplitPaneRightFromStackSplitsWholeStack(t *testing.T) {
68		newPTY := portsmocks.NewMockPTY(t)	68		newPTY := portsmocks.NewMockPTY(t)
69		stackPTY.EXPECT().Resize(domain.Size{Cols: 20, Rows: 2}).Return(nil).Once()	69		stackPTY.EXPECT().Resize(domain.Size{Cols: 20, Rows: 2}).Return(nil).Once()
70		newPTY.EXPECT().Read(mock.Anything).RunAndReturn(blockingRead(t)).Maybe()	70		newPTY.EXPECT().Read(mock.Anything).RunAndReturn(blockingRead(t)).Maybe()
71	-	factory.EXPECT().Open("/bin/sh", []string(nil), mock.Anything, "/work", domain.Size{Cols: 20, Rows: 4}).Return(newPTY, nil).Once()	71	+	factory.EXPECT().Open(mock.Anything, "/bin/sh", []string(nil), mock.Anything, "/work", domain.Size{Cols: 20, Rows: 4}).Return(newPTY, nil).Once()
72			72
73		require.NoError(t, d.splitPane(sess, nil, layout.Right))	73		require.NoError(t, d.splitPane(sess, nil, layout.Right))
74			74

@@ -84,7 +84,7 @@ func TestSplitPaneRightFromStackSplitsWholeStack(t *testing.T) {
84			84
85		func TestSplitPaneOpenErrorRollsBackTreeAndPaneMap(t *testing.T) {	85		func TestSplitPaneOpenErrorRollsBackTreeAndPaneMap(t *testing.T) {
86		d, sess, _, factory := newSplitTestDaemon(t, domain.Size{Cols: 41, Rows: 10})	86		d, sess, _, factory := newSplitTestDaemon(t, domain.Size{Cols: 41, Rows: 10})
87	-	factory.EXPECT().Open("/bin/sh", []string(nil), mock.Anything, "/work", domain.Size{Cols: 20, Rows: 10}).Return(nil, errors.New("open failed")).Once()	87	+	factory.EXPECT().Open(mock.Anything, "/bin/sh", []string(nil), mock.Anything, "/work", domain.Size{Cols: 20, Rows: 10}).Return(nil, errors.New("open failed")).Once()
88			88
89		tb := sess.activeTab()	89		tb := sess.activeTab()
90		tb.mu.Lock()	90		tb.mu.Lock()

@@ -231,7 +231,7 @@ func TestStackPaneCreatesStackAndToggleRestoresSplit(t *testing.T) {
231		d, sess, oldPTY, factory := newSplitTestDaemon(t, domain.Size{Cols: 20, Rows: 5})	231		d, sess, oldPTY, factory := newSplitTestDaemon(t, domain.Size{Cols: 20, Rows: 5})
232		newPTY := portsmocks.NewMockPTY(t)	232		newPTY := portsmocks.NewMockPTY(t)
233		newPTY.EXPECT().Read(mock.Anything).RunAndReturn(blockingRead(t)).Maybe()	233		newPTY.EXPECT().Read(mock.Anything).RunAndReturn(blockingRead(t)).Maybe()
234	-	factory.EXPECT().Open("/bin/sh", []string(nil), mock.Anything, "/work", domain.Size{Cols: 20, Rows: 3}).Return(newPTY, nil).Once()	234	+	factory.EXPECT().Open(mock.Anything, "/bin/sh", []string(nil), mock.Anything, "/work", domain.Size{Cols: 20, Rows: 3}).Return(newPTY, nil).Once()
235			235
236		require.NoError(t, d.stackPane(sess, nil))	236		require.NoError(t, d.stackPane(sess, nil))
237		tb := sess.activeTab()	237		tb := sess.activeTab()

@@ -251,7 +251,7 @@ func TestStackFocusWalkExpandsAndOverflowRefuses(t *testing.T) {
251		d, sess, oldPTY, factory := newSplitTestDaemon(t, domain.Size{Cols: 20, Rows: 3})	251		d, sess, oldPTY, factory := newSplitTestDaemon(t, domain.Size{Cols: 20, Rows: 3})
252		newPTY := portsmocks.NewMockPTY(t)	252		newPTY := portsmocks.NewMockPTY(t)
253		newPTY.EXPECT().Read(mock.Anything).RunAndReturn(blockingRead(t)).Maybe()	253		newPTY.EXPECT().Read(mock.Anything).RunAndReturn(blockingRead(t)).Maybe()
254	-	factory.EXPECT().Open("/bin/sh", []string(nil), mock.Anything, "/work", domain.Size{Cols: 20, Rows: 1}).Return(newPTY, nil).Once()	254	+	factory.EXPECT().Open(mock.Anything, "/bin/sh", []string(nil), mock.Anything, "/work", domain.Size{Cols: 20, Rows: 1}).Return(newPTY, nil).Once()
255		require.NoError(t, d.stackPane(sess, nil))	255		require.NoError(t, d.stackPane(sess, nil))
256		oldPTY.EXPECT().Resize(domain.Size{Cols: 20, Rows: 1}).Return(nil).Once()	256		oldPTY.EXPECT().Resize(domain.Size{Cols: 20, Rows: 1}).Return(nil).Once()
257			257

@@ -261,7 +261,7 @@ func TestStackFocusWalkExpandsAndOverflowRefuses(t *testing.T) {
261		require.Equal(t, layout.PaneID("pane-1"), tb.tree.Focus)	261		require.Equal(t, layout.PaneID("pane-1"), tb.tree.Focus)
262		require.Equal(t, layout.PaneID("pane-1"), tb.tree.Root.Expanded)	262		require.Equal(t, layout.PaneID("pane-1"), tb.tree.Root.Expanded)
263		thirdPTY := portsmocks.NewMockPTY(t)	263		thirdPTY := portsmocks.NewMockPTY(t)
264	-	factory.EXPECT().Open("/bin/sh", []string(nil), mock.Anything, "/work", mock.Anything).Return(thirdPTY, nil).Maybe()	264	+	factory.EXPECT().Open(mock.Anything, "/bin/sh", []string(nil), mock.Anything, "/work", mock.Anything).Return(thirdPTY, nil).Maybe()
265		require.ErrorIs(t, d.stackPane(sess, nil), layout.ErrTooSmall)	265		require.ErrorIs(t, d.stackPane(sess, nil), layout.ErrTooSmall)
266		require.Len(t, tb.panes, 2)	266		require.Len(t, tb.panes, 2)
267		}	267		}

@@ -291,7 +291,7 @@ func TestCloseOriginalPaneLeavesSurvivorFunctional(t *testing.T) {
291		defer releaseNew()	291		defer releaseNew()
292		oldPTY.EXPECT().Resize(domain.Size{Cols: 20, Rows: 10}).Return(nil).Once()	292		oldPTY.EXPECT().Resize(domain.Size{Cols: 20, Rows: 10}).Return(nil).Once()
293		newPTY.EXPECT().Read(mock.Anything).RunAndReturn(blockingRead(t)).Maybe()	293		newPTY.EXPECT().Read(mock.Anything).RunAndReturn(blockingRead(t)).Maybe()
294	-	factory.EXPECT().Open("/bin/sh", []string(nil), mock.Anything, "/work", domain.Size{Cols: 20, Rows: 10}).Return(newPTY, nil).Once()	294	+	factory.EXPECT().Open(mock.Anything, "/bin/sh", []string(nil), mock.Anything, "/work", domain.Size{Cols: 20, Rows: 10}).Return(newPTY, nil).Once()
295			295
296		require.NoError(t, d.splitPane(sess, nil, layout.Right))	296		require.NoError(t, d.splitPane(sess, nil, layout.Right))
297		tb := sess.activeTab()	297		tb := sess.activeTab()

@@ -350,8 +350,8 @@ func TestSplitPaneInheritsTerminalEnv(t *testing.T) {
350		oldPTY.EXPECT().Resize(domain.Size{Cols: 20, Rows: 10}).Return(nil).Once()	350		oldPTY.EXPECT().Resize(domain.Size{Cols: 20, Rows: 10}).Return(nil).Once()
351		newPTY.EXPECT().Read(mock.Anything).RunAndReturn(blockingRead(t)).Maybe()	351		newPTY.EXPECT().Read(mock.Anything).RunAndReturn(blockingRead(t)).Maybe()
352		var gotEnv []string	352		var gotEnv []string
353	-	factory.EXPECT().Open("/bin/sh", []string(nil), mock.Anything, "/work", domain.Size{Cols: 20, Rows: 10}).RunAndReturn(	353	+	factory.EXPECT().Open(mock.Anything, "/bin/sh", []string(nil), mock.Anything, "/work", domain.Size{Cols: 20, Rows: 10}).RunAndReturn(
354	-	func(_ string, _ []string, env []string, _ string, _ domain.Size) (ports.PTY, error) {	354	+	func(_ context.Context, _ string, _ []string, env []string, _ string, _ domain.Size) (ports.PTY, error) {
355		gotEnv = append([]string(nil), env...)	355		gotEnv = append([]string(nil), env...)
356		return newPTY, nil	356		return newPTY, nil
357		},	357		},

M internal/usecase/daemon/performance_benchmark_test.go +650 -54


@@ -5,6 +5,7 @@ import (
5		"io"	5		"io"
6		"sync"	6		"sync"
7		"testing"	7		"testing"
			8	+	"time"
8			9
9		"github.com/stretchr/testify/require"	10		"github.com/stretchr/testify/require"
10			11

@@ -49,8 +50,15 @@ func TestPerformanceFixtureCounters(t *testing.T) {
49		fixture.paintLive()	50		fixture.paintLive()
50		metrics := fixture.metrics()	51		metrics := fixture.metrics()
51		require.Equal(t, uint64(2), metrics.outputFrames)	52		require.Equal(t, uint64(2), metrics.outputFrames)
			53	+	require.Equal(t, uint64(2), metrics.renderCaptures, "capture is one immutable state snapshot per live render request")
			54	+	require.Equal(t, uint64(2), metrics.renderCompositions, "composition is one frame built from each captured state")
			55	+	require.Equal(t, uint64(2), metrics.renderEmissions, "emission is one accepted output frame")
52		require.Positive(t, metrics.outputBytes)	56		require.Positive(t, metrics.outputBytes)
53		require.Positive(t, metrics.outputPayloadBytes)	57		require.Positive(t, metrics.outputPayloadBytes)
			58	+	require.Equal(t, uint64(2), metrics.coordinatorInvalidations)
			59	+	require.Equal(t, uint64(2), metrics.coordinatorWakes)
			60	+	require.Equal(t, uint64(2), metrics.coordinatorCoalesced)
			61	+	require.Positive(t, metrics.coordinatorWakes, "coordinator metric ratios require a nonzero denominator")
54			62
55		fixture.captureSnapshot()	63		fixture.captureSnapshot()
56		metrics = fixture.metrics()	64		metrics = fixture.metrics()

@@ -64,14 +72,70 @@ func TestPerformanceFixtureCounters(t *testing.T) {
64		require.True(t, fixture.resized())	72		require.True(t, fixture.resized())
65		}	73		}
66			74
67	-	func TestPerformanceFixtureResizeAlternatesRealDimensions(t *testing.T) {	75	+	func TestRenderStageHooksCountProductionBoundariesOnFailedSend(t *testing.T) {
68	-	fixture := newPerformanceFixture(t, performanceConfig{size: domain.Size{Cols: 120, Rows: 40}})	76	+	d, sess, ac, sends := newManualSessionWithPTYs(t, nil)
			77	+	d.paint(sess, ac, true, nil)
			78	+	_ = awaitFrame(t, sends, ports.MsgOutput)
			79	+	var captures, compositions, emissions int
			80	+	ac.renderStages = renderStageHooks{
			81	+	capture: func() { captures++ },
			82	+	compose: func() { compositions++ },
			83	+	emit: func() { emissions++ },
			84	+	}
			85	+	ac.replaceTransport(cacheFailTransport{})
			86	+	p := sess.tabs[0].focusedPane()
			87	+	p.mu.Lock()
			88	+	p.screen.Write([]byte("hook"))
			89	+	p.mu.Unlock()
			90	+	d.paint(sess, ac, false, nil)
			91	+	require.Equal(t, 1, captures, "a completed capture counts even when its emission fails")
			92	+	require.Equal(t, 1, compositions, "a completed composition counts even when its emission fails")
			93	+	require.Zero(t, emissions, "emit counts only prepare and transport success")
			94	+	}
			95	+
			96	+	func TestPerformanceFixtureResize(t *testing.T) {
			97	+	fixture := newPerformanceFixture(t, performanceConfig{size: domain.Size{Cols: 80, Rows: 24}})
			98	+	sequence := []domain.Size{{Cols: 100, Rows: 30}, {Cols: 120, Rows: 40}, {Cols: 160, Rows: 50}, {Cols: 80, Rows: 24}}
69			99
70	-	for _, want := range []domain.Size{{Cols: 100, Rows: 30}, {Cols: 120, Rows: 40}} {	100	+	for _, want := range sequence {
71	-	fixture.resize()	101	+	fixture.resizeTo(want)
72		require.Equal(t, want, fixture.ac.size)	102		require.Equal(t, want, fixture.ac.size)
73		require.True(t, fixture.resized())	103		require.True(t, fixture.resized())
74		}	104		}
			105	+	metrics := fixture.metrics()
			106	+	require.Equal(t, uint64(len(sequence)), metrics.resizeRequests)
			107	+	require.Equal(t, uint64(len(sequence)), metrics.resizeCommits)
			108	+	require.Equal(t, uint64(len(sequence)), metrics.outputFrames, "one full frame per accepted epoch")
			109	+	require.Zero(t, metrics.skippedEpochs)
			110	+	require.Zero(t, metrics.ptyFailures)
			111	+	require.Zero(t, metrics.frameGapEpochs)
			112	+	require.NotZero(t, metrics.resizeCommitNanos)
			113	+	require.Equal(t, []domain.Size{{Cols: 50, Rows: 28}, {Cols: 60, Rows: 38}, {Cols: 80, Rows: 48}, {Cols: 40, Rows: 22}}, fixture.pty.requested())
			114	+	}
			115	+
			116	+	func TestTransactionalResizeMetricDefinitionsAreZeroSafe(t *testing.T) {
			117	+	metrics := performanceMetrics{}
			118	+	require.Zero(t, coordinatorCoalescingRatio(metrics))
			119	+	require.Zero(t, metrics.skippedEpochs, "no accepted request means no skipped epoch")
			120	+	require.Zero(t, metrics.frameGapEpochs, "no committed epoch means no frame gap")
			121	+	}
			122	+
			123	+	func TestTransactionalResizeMetrics(t *testing.T) {
			124	+	fixture := newPerformanceFixture(t, performanceConfig{size: domain.Size{Cols: 80, Rows: 24}, ptyErrors: []error{fmt.Errorf("scripted failure"), nil}})
			125	+	fixture.resizeTo(domain.Size{Cols: 100, Rows: 30})
			126	+	// The fixture deliberately invokes the coordinator's committed retry path:
			127	+	// a retry never republishes geometry and its reset produces one extra frame.
			128	+	fixture.retryLatest()
			129	+
			130	+	metrics := fixture.metrics()
			131	+	require.Equal(t, uint64(1), metrics.resizeRequests)
			132	+	require.Equal(t, uint64(1), metrics.resizeCommits)
			133	+	require.Equal(t, uint64(1), metrics.ptyFailures)
			134	+	require.Equal(t, uint64(1), metrics.ptyRetries)
			135	+	require.Equal(t, uint64(2), metrics.outputFrames, "commit plus successful full-reset retry")
			136	+	require.Zero(t, metrics.frameGapEpochs, "every committed epoch emitted its required frame")
			137	+	require.Equal(t, []domain.Size{{Cols: 50, Rows: 28}, {Cols: 50, Rows: 28}}, fixture.pty.requested())
			138	+	require.Equal(t, domain.Size{Cols: 100, Rows: 30}, fixture.ac.size)
75		}	139		}
76			140
77		func TestPerformanceFixtureLargeHistoryTopology(t *testing.T) {	141		func TestPerformanceFixtureLargeHistoryTopology(t *testing.T) {

@@ -81,45 +145,137 @@ func TestPerformanceFixtureLargeHistoryTopology(t *testing.T) {
81		for _, tb := range fixture.sess.tabs {	145		for _, tb := range fixture.sess.tabs {
82		require.Len(t, tb.panes, 4)	146		require.Len(t, tb.panes, 4)
83		for _, p := range tb.panes {	147		for _, p := range tb.panes {
84	-	require.Equal(t, 10_000, p.scrollback.Len())	148	+	require.Equal(t, 10_000, p.history.Len())
85		}	149		}
86		}	150		}
87		}	151		}
88			152
			153	+	func TestPerformanceFixtureSnapshotCaptureRetainsSealedHistory(t *testing.T) {
			154	+	fixture := newPerformanceFixture(t, performanceConfig{size: domain.Size{Cols: 120, Rows: 40}, historyRows: 10_000})
			155	+
			156	+	first, ok := fixture.d.captureSnapshotState(fixture.sess, 1)
			157	+	require.True(t, ok)
			158	+	require.Len(t, first.tabs, 1)
			159	+	require.Len(t, first.tabs[0].panes, 2)
			160	+	require.Equal(t, 10_000, first.tabs[0].panes[0].history.Len())
			161	+	require.Equal(t, 40, first.tabs[0].panes[0].history.ChunkCount())
			162	+
			163	+	second, ok := fixture.d.captureSnapshotState(fixture.sess, 2)
			164	+	require.True(t, ok)
			165	+	secondPane := capturePaneByID(second, first.tabs[0].panes[0].id)
			166	+	if secondPane == nil \|\| first.tabs[0].panes[0].history.Chunk(0) != secondPane.history.Chunk(0) {
			167	+	t.Fatal("unchanged sealed history chunk was not reused across captures")
			168	+	}
			169	+	}
			170	+
			171	+	func TestPerformanceFixtureCoordinatorMetricsForCanonicalTopologies(t *testing.T) {
			172	+	for _, topology := range daemonHistoryTopologies {
			173	+	t.Run(topology.name, func(t *testing.T) {
			174	+	fixture := newPerformanceFixture(t, performanceConfig{
			175	+	size: domain.Size{Cols: 120, Rows: 40}, tabs: topology.tabs, panes: topology.panes, historyRows: 10_000,
			176	+	})
			177	+	fixture.paintLive()
			178	+	metrics := fixture.metrics()
			179	+	require.Equal(t, uint64(1), metrics.coordinatorInvalidations)
			180	+	require.Equal(t, uint64(1), metrics.coordinatorWakes)
			181	+	require.Equal(t, uint64(1), metrics.coordinatorCoalesced)
			182	+	require.Positive(t, metrics.coordinatorWakes, "coalescing ratio denominator")
			183	+	})
			184	+	}
			185	+	}
			186	+
			187	+	type daemonHistoryTopology struct {
			188	+	name string
			189	+	tabs, panes int
			190	+	}
			191	+
			192	+	var daemonHistoryTopologies = []daemonHistoryTopology{
			193	+	{name: "1tab-1pane-control", tabs: 1, panes: 1},
			194	+	{name: "1tab-4panes", tabs: 1, panes: 4},
			195	+	{name: "4tabs-1pane", tabs: 4, panes: 1},
			196	+	{name: "4tabs-4panes", tabs: 4, panes: 4},
			197	+	{name: "8tabs-1pane", tabs: 8, panes: 1},
			198	+	}
			199	+
			200	+	// v3 bounds a decoded snapshot to 256 MiB. One 120-column 10k-row tab is
			201	+	// representative and valid; multi-tab fixtures exceed that durable budget, so
			202	+	// snapshot capture/encode benchmark only the valid one-tab topologies.
			203	+	var daemonSnapshotTopologies = []daemonHistoryTopology{
			204	+	{name: "1tab-1pane-control", tabs: 1, panes: 1},
			205	+	{name: "1tab-4panes", tabs: 1, panes: 4},
			206	+	}
			207	+
89		func BenchmarkDaemonHistoryLivePaint(b *testing.B) {	208		func BenchmarkDaemonHistoryLivePaint(b *testing.B) {
90		benchmarkDaemonLargeHistory(b, "live-paint", func(f *performanceFixture) { f.paintLive() })	209		benchmarkDaemonLargeHistory(b, "live-paint", func(f *performanceFixture) { f.paintLive() })
91		}	210		}
			211	+
			212	+	// BenchmarkDaemonHistorySnapshotCapture measures the synchronous immutable
			213	+	// capture boundary only. Encoding and the bounded worker queue are deliberately
			214	+	// excluded so scheduler latency cannot affect this benchmark.
92		func BenchmarkDaemonHistorySnapshotCapture(b *testing.B) {	215		func BenchmarkDaemonHistorySnapshotCapture(b *testing.B) {
93	-	benchmarkDaemonLargeHistory(b, "snapshot-capture", func(f *performanceFixture) { f.captureSnapshot() })	216	+	benchmarkDaemonSnapshotCapture(b)
			217	+	}
			218	+
			219	+	// BenchmarkDaemonHistorySnapshotEncode measures v3 encoding from an already
			220	+	// immutable capture, without queueing or persistence.
			221	+	func BenchmarkDaemonHistorySnapshotEncode(b *testing.B) {
			222	+	benchmarkDaemonSnapshotEncode(b)
			223	+	}
			224	+
			225	+	// BenchmarkDaemonHistorySnapshotRestore measures terminal-state restoration
			226	+	// from a validated v3 pane manifest without spawning PTY or session workers.
			227	+	func BenchmarkDaemonHistorySnapshotRestore(b *testing.B) {
			228	+	benchmarkDaemonSnapshotRestore(b)
94		}	229		}
			230	+
			231	+	// BenchmarkDaemonHistoryCopyEnter uses the parent benchmark's repeated-entry
			232	+	// semantics: each operation replaces the active copy document. Exit cleanup is
			233	+	// intentionally outside this benchmark, so its allocations cannot be reported
			234	+	// as copy-entry work.
95		func BenchmarkDaemonHistoryCopyEnter(b *testing.B) {	235		func BenchmarkDaemonHistoryCopyEnter(b *testing.B) {
96		benchmarkDaemonLargeHistory(b, "copy-enter", func(f *performanceFixture) { f.d.enterCopyMode(f.sess, f.ac) })	236		benchmarkDaemonLargeHistory(b, "copy-enter", func(f *performanceFixture) { f.d.enterCopyMode(f.sess, f.ac) })
97		}	237		}
98		func BenchmarkDaemonHistoryCopySearch(b *testing.B) {	238		func BenchmarkDaemonHistoryCopySearch(b *testing.B) {
99		benchmarkDaemonLargeHistory(b, "copy-search", func(f *performanceFixture) { f.d.handleInput(f.sess, f.ac, []byte("/needle\r")) })	239		benchmarkDaemonLargeHistory(b, "copy-search", func(f *performanceFixture) { f.d.handleInput(f.sess, f.ac, []byte("/needle\r")) })
100		}	240		}
101	-	func BenchmarkDaemonHistoryResize(b *testing.B) {	241	+
102	-	benchmarkDaemonLargeHistory(b, "resize", func(f *performanceFixture) { f.resize() })	242	+	// BenchmarkDaemonHistoryResizeSweep alternates deterministic grow/shrink
			243	+	// geometry. Metrics are fixture counters; no benchmark result depends on wall
			244	+	// time or scheduler delivery.
			245	+	func BenchmarkDaemonHistoryResizeSweep(b *testing.B) {
			246	+	sequence := []domain.Size{{Cols: 80, Rows: 24}, {Cols: 100, Rows: 30}, {Cols: 120, Rows: 40}, {Cols: 160, Rows: 50}, {Cols: 120, Rows: 40}, {Cols: 100, Rows: 30}}
			247	+	benchmarkDaemonLargeHistory(b, "resize-sweep", func(f *performanceFixture) {
			248	+	f.resizeTo(sequence[f.resizes%len(sequence)])
			249	+	})
			250	+	}
			251	+
			252	+	// BenchmarkDaemonHistoryResizeRetry measures one deterministic failed apply
			253	+	// followed by its successful full-reset retry for each operation.
			254	+	func BenchmarkDaemonHistoryResizeRetry(b *testing.B) {
			255	+	benchmarkDaemonLargeHistory(b, "resize-retry", func(f *performanceFixture) {
			256	+	f.pty.setErrors([]error{fmt.Errorf("scripted resize failure"), nil})
			257	+	size := domain.Size{Cols: 100, Rows: 30}
			258	+	if f.resizes%2 != 0 {
			259	+	size = domain.Size{Cols: 120, Rows: 40}
			260	+	}
			261	+	f.resizeTo(size)
			262	+	f.retryLatest()
			263	+	})
103		}	264		}
104			265
105		func benchmarkDaemonLargeHistory(b testing.B, workload string, run func(performanceFixture)) {	266		func benchmarkDaemonLargeHistory(b testing.B, workload string, run func(performanceFixture)) {
106		b.Helper()	267		b.Helper()
107	-	for _, topology := range []struct {	268	+	for _, topology := range daemonHistoryTopologies {
108	-	name string
109	-	tabs, panes int
110	-	}{
111	-	{name: "1tab-1pane-control", tabs: 1, panes: 1},
112	-	{name: "1tab-4panes", tabs: 1, panes: 4},
113	-	{name: "4tabs-1pane", tabs: 4, panes: 1},
114	-	{name: "4tabs-4panes", tabs: 4, panes: 4},
115	-	{name: "8tabs-1pane", tabs: 8, panes: 1},
116	-	} {
117		b.Run(topology.name, func(b *testing.B) {	269		b.Run(topology.name, func(b *testing.B) {
118		fixture := newPerformanceFixture(b, performanceConfig{	270		fixture := newPerformanceFixture(b, performanceConfig{
119		size: domain.Size{Cols: 120, Rows: 40}, tabs: topology.tabs, panes: topology.panes, historyRows: 10_000,	271		size: domain.Size{Cols: 120, Rows: 40}, tabs: topology.tabs, panes: topology.panes, historyRows: 10_000,
120		})	272		})
			273	+	if !fixture.hasHistoryTopology(topology.tabs, topology.panes, 10_000) {
			274	+	b.Fatal("invalid daemon history fixture")
			275	+	}
121		if workload == "copy-search" {	276		if workload == "copy-search" {
122	-	fixture.d.enterCopyMode(fixture.sess, fixture.ac)	277	+	benchmarkDaemonCopySearch(b, fixture, run)
			278	+	return
123		}	279		}
124		b.ReportAllocs()	280		b.ReportAllocs()
125		fixture.resetMetrics()	281		fixture.resetMetrics()

@@ -133,26 +289,180 @@ func benchmarkDaemonLargeHistory(b testing.B, workload string, run func(perfor
133		b.Fatalf("decode last output: %v", err)	289		b.Fatalf("decode last output: %v", err)
134		}	290		}
135		}	291		}
136	-	if workload == "snapshot-capture" {
137	-	payload := fixture.snaps.lastPayload()
138	-	if payload == nil {
139	-	b.Fatal("snapshot capture retained no snapshot payload")
140	-	}
141	-	if _, err := snapcodec.Unmarshal(payload); err != nil {
142	-	b.Fatalf("decode last snapshot: %v", err)
143	-	}
144	-	}
145		if workload == "live-paint" && metrics.outputFrames != uint64(b.N) {	292		if workload == "live-paint" && metrics.outputFrames != uint64(b.N) {
146		b.Fatalf("live paint emitted %d frames for %d operations", metrics.outputFrames, b.N)	293		b.Fatalf("live paint emitted %d frames for %d operations", metrics.outputFrames, b.N)
147		}	294		}
148	-	if workload == "snapshot-capture" && metrics.snapshotWrites != uint64(b.N) {
149	-	b.Fatalf("snapshot capture wrote %d snapshots for %d operations", metrics.snapshotWrites, b.N)
150	-	}
151		benchmarkReportMetrics(b, metrics, b.N, 10_000)	295		benchmarkReportMetrics(b, metrics, b.N, 10_000)
152		})	296		})
153		}	297		}
154		}	298		}
155			299
			300	+	func benchmarkDaemonCopySearch(b testing.B, fixture performanceFixture, run func(*performanceFixture)) {
			301	+	b.Helper()
			302	+	fixture.d.enterCopyMode(fixture.sess, fixture.ac)
			303	+	fixture.ac.ackOutputState(fixture.ac.output.next)
			304	+	b.ReportAllocs()
			305	+	b.ResetTimer()
			306	+	for range b.N {
			307	+	run(fixture)
			308	+	fixture.ac.ackOutputState(fixture.ac.output.next)
			309	+	}
			310	+	b.StopTimer()
			311	+	if fixture.searchMatches() == 0 {
			312	+	b.Fatal("copy search produced no deterministic matches")
			313	+	}
			314	+	b.ReportMetric(1, "copyoperations/op")
			315	+	}
			316	+
			317	+	func benchmarkDaemonSnapshotCapture(b *testing.B) {
			318	+	for _, topology := range daemonSnapshotTopologies {
			319	+	b.Run(topology.name, func(b *testing.B) {
			320	+	fixture := newPerformanceFixture(b, performanceConfig{size: domain.Size{Cols: 120, Rows: 40}, tabs: topology.tabs, panes: topology.panes, historyRows: 10_000})
			321	+	if !fixture.hasHistoryTopology(topology.tabs, topology.panes, 10_000) {
			322	+	b.Fatal("invalid daemon snapshot capture fixture")
			323	+	}
			324	+	first, ok := fixture.d.captureSnapshotState(fixture.sess, 1)
			325	+	if !ok \|\| !validBenchmarkCapture(first) {
			326	+	b.Fatal("invalid snapshot capture fixture")
			327	+	}
			328	+	firstPane := first.tabs[0].panes[0]
			329	+	firstChunk := firstPane.history.Chunk(0)
			330	+	captures := 0
			331	+	var last *snapshotCapture
			332	+	b.ReportAllocs()
			333	+	b.ResetTimer()
			334	+	for range b.N {
			335	+	last, ok = fixture.d.captureSnapshotState(fixture.sess, uint64(captures+2))
			336	+	if !ok {
			337	+	b.Fatal("snapshot capture unexpectedly rejected fixture")
			338	+	}
			339	+	captures++
			340	+	}
			341	+	b.StopTimer()
			342	+	lastPane := capturePaneByID(last, firstPane.id)
			343	+	if captures != b.N \|\| !validBenchmarkCapture(last) \|\| lastPane == nil \|\| lastPane.history.Chunk(0) != firstChunk {
			344	+	b.Fatal("snapshot capture did not preserve immutable sealed history")
			345	+	}
			346	+	b.ReportMetric(float64(capturePaneCount(last)), "capturepanes/op")
			347	+	})
			348	+	}
			349	+	}
			350	+
			351	+	func benchmarkDaemonSnapshotEncode(b *testing.B) {
			352	+	for _, topology := range daemonSnapshotTopologies {
			353	+	b.Run(topology.name, func(b *testing.B) {
			354	+	fixture := newPerformanceFixture(b, performanceConfig{size: domain.Size{Cols: 120, Rows: 40}, tabs: topology.tabs, panes: topology.panes, historyRows: 10_000})
			355	+	if !fixture.hasHistoryTopology(topology.tabs, topology.panes, 10_000) {
			356	+	b.Fatal("invalid daemon snapshot encode fixture")
			357	+	}
			358	+	capture, ok := fixture.d.captureSnapshotState(fixture.sess, 1)
			359	+	if !ok \|\| !validBenchmarkCapture(capture) {
			360	+	b.Fatal("invalid snapshot encode fixture")
			361	+	}
			362	+	encoded, err := fixture.d.encodeSnapshotCapture(capture)
			363	+	if err != nil {
			364	+	b.Fatal(err)
			365	+	}
			366	+	if _, err := snapcodec.Unmarshal(encoded); err != nil {
			367	+	b.Fatalf("invalid v3 snapshot encode fixture: %v", err)
			368	+	}
			369	+	var last []byte
			370	+	b.ReportAllocs()
			371	+	b.SetBytes(int64(len(encoded)))
			372	+	b.ResetTimer()
			373	+	for range b.N {
			374	+	last, err = fixture.d.encodeSnapshotCapture(capture)
			375	+	if err != nil {
			376	+	b.Fatal(err)
			377	+	}
			378	+	}
			379	+	b.StopTimer()
			380	+	if _, err := snapcodec.Unmarshal(last); err != nil {
			381	+	b.Fatalf("snapshot encoder produced invalid v3 data: %v", err)
			382	+	}
			383	+	b.ReportMetric(float64(len(last)), "snapshotbytes/op")
			384	+	})
			385	+	}
			386	+	}
			387	+
			388	+	func benchmarkDaemonSnapshotRestore(b *testing.B) {
			389	+	fixture := newPerformanceFixture(b, performanceConfig{size: domain.Size{Cols: 120, Rows: 40}, panes: 1, historyRows: 10_000})
			390	+	if !fixture.hasHistoryTopology(1, 1, 10_000) {
			391	+	b.Fatal("invalid daemon snapshot restore fixture")
			392	+	}
			393	+	capture, ok := fixture.d.captureSnapshotState(fixture.sess, 1)
			394	+	if !ok \|\| !validBenchmarkCapture(capture) {
			395	+	b.Fatal("invalid snapshot restore fixture")
			396	+	}
			397	+	encoded, err := fixture.d.encodeSnapshotCapture(capture)
			398	+	if err != nil {
			399	+	b.Fatal(err)
			400	+	}
			401	+	snapshot, err := snapcodec.Unmarshal(encoded)
			402	+	if err != nil \|\| len(snapshot.Tabs) != 1 \|\| len(snapshot.Tabs[0].Panes) == 0 {
			403	+	b.Fatalf("invalid v3 snapshot restore fixture: %v", err)
			404	+	}
			405	+	paneSnapshot := snapshot.Tabs[0].Panes[0]
			406	+	b.ReportAllocs()
			407	+	b.SetBytes(int64(len(encoded)))
			408	+	b.ResetTimer()
			409	+	for range b.N {
			410	+	if err := restorePaneTerminal(fixture.activePane, paneSnapshot); err != nil {
			411	+	b.Fatal(err)
			412	+	}
			413	+	}
			414	+	b.StopTimer()
			415	+	fixture.activePane.mu.Lock()
			416	+	restoredRows := fixture.activePane.history.Len()
			417	+	fixture.activePane.mu.Unlock()
			418	+	if restoredRows != 10_000 {
			419	+	b.Fatalf("snapshot restore retained %d history rows, want 10000", restoredRows)
			420	+	}
			421	+	b.ReportMetric(float64(restoredRows), "historyrows/restore")
			422	+	}
			423	+
			424	+	func validBenchmarkCapture(capture *snapshotCapture) bool {
			425	+	return capture != nil && len(capture.tabs) > 0 && len(capture.tabs[0].panes) > 0 && capture.tabs[0].panes[0].history.Len() == 10_000 && capture.tabs[0].panes[0].history.ChunkCount() > 0
			426	+	}
			427	+
			428	+	func capturePaneCount(capture *snapshotCapture) int {
			429	+	count := 0
			430	+	for _, tab := range capture.tabs {
			431	+	count += len(tab.panes)
			432	+	}
			433	+	return count
			434	+	}
			435	+
			436	+	func capturePaneByID(capture snapshotCapture, id layout.PaneID) snapshotCapturePane {
			437	+	if capture == nil {
			438	+	return nil
			439	+	}
			440	+	for tabIndex := range capture.tabs {
			441	+	for paneIndex := range capture.tabs[tabIndex].panes {
			442	+	pane := &capture.tabs[tabIndex].panes[paneIndex]
			443	+	if pane.id == id {
			444	+	return pane
			445	+	}
			446	+	}
			447	+	}
			448	+	return nil
			449	+	}
			450	+
			451	+	func TestCoordinatorCoalescingRatioReportsInvalidationsPerWake(t *testing.T) {
			452	+	require.Equal(t, float64(3), coordinatorCoalescingRatio(performanceMetrics{
			453	+	coordinatorWakes: 1,
			454	+	coordinatorCoalesced: 3,
			455	+	}))
			456	+	require.Zero(t, coordinatorCoalescingRatio(performanceMetrics{}))
			457	+	}
			458	+
			459	+	func coordinatorCoalescingRatio(metrics performanceMetrics) float64 {
			460	+	if metrics.coordinatorWakes == 0 {
			461	+	return 0
			462	+	}
			463	+	return float64(metrics.coordinatorCoalesced) / float64(metrics.coordinatorWakes)
			464	+	}
			465	+
156		func benchmarkReportMetrics(b *testing.B, metrics performanceMetrics, operations, historyRows int) {	466		func benchmarkReportMetrics(b *testing.B, metrics performanceMetrics, operations, historyRows int) {
157		b.Helper()	467		b.Helper()
158		if operations == 0 {	468		if operations == 0 {

@@ -164,9 +474,89 @@ func benchmarkReportMetrics(b *testing.B, metrics performanceMetrics, operations
164		b.ReportMetric(float64(metrics.outputPayloadBytes)/perOperation, "framepayloadbytes/op")	474		b.ReportMetric(float64(metrics.outputPayloadBytes)/perOperation, "framepayloadbytes/op")
165		b.ReportMetric(float64(metrics.snapshotWrites)/perOperation, "snapshotwrites/op")	475		b.ReportMetric(float64(metrics.snapshotWrites)/perOperation, "snapshotwrites/op")
166		b.ReportMetric(float64(metrics.snapshotBytes)/perOperation, "snapshotbytes/op")	476		b.ReportMetric(float64(metrics.snapshotBytes)/perOperation, "snapshotbytes/op")
			477	+	b.ReportMetric(float64(metrics.coordinatorInvalidations)/perOperation, "coordinatorinvalidations/op")
			478	+	b.ReportMetric(float64(metrics.coordinatorWakes)/perOperation, "coordinatorwakes/op")
			479	+	b.ReportMetric(float64(metrics.coordinatorCoalesced)/perOperation, "coordinatorcoalesced/op")
			480	+	// Split counters have precise fixture definitions: capture is an immutable
			481	+	// state snapshot requested for a live paint; composition builds the frame
			482	+	// from that capture; emission accepts its output frame. Keep bytes/frames
			483	+	// above as transport observables rather than folding them into these rates.
			484	+	b.ReportMetric(float64(metrics.renderCaptures)/perOperation, "rendercaptures/op")
			485	+	b.ReportMetric(float64(metrics.renderCompositions)/perOperation, "rendercompositions/op")
			486	+	b.ReportMetric(float64(metrics.renderEmissions)/perOperation, "renderemissions/op")
			487	+	// resizecommitns is the injected-clock elapsed time from accepted request
			488	+	// to committed epoch. The remaining values are counter deltas: skipped is
			489	+	// accepted minus committed (never negative), framegap is committed epochs
			490	+	// lacking their required output frame (never negative), and retries/failures
			491	+	// are PTY Resize outcomes. All divisions are guarded above.
			492	+	b.ReportMetric(float64(metrics.resizeCommitNanos)/perOperation, "resizecommitns/op")
			493	+	b.ReportMetric(float64(metrics.skippedEpochs)/perOperation, "skippedepochs/op")
			494	+	b.ReportMetric(float64(metrics.ptyFailures)/perOperation, "ptyfailures/op")
			495	+	b.ReportMetric(float64(metrics.ptyRetries)/perOperation, "ptyretries/op")
			496	+	b.ReportMetric(float64(metrics.frameGapEpochs)/perOperation, "framegapepochs/op")
			497	+	b.ReportMetric(coordinatorCoalescingRatio(metrics), "coordinatorcoalescingratio")
167		b.ReportMetric(float64(historyRows), "historyrows/pane")	498		b.ReportMetric(float64(historyRows), "historyrows/pane")
168		}	499		}
169			500
			501	+	// Root-cause hypothesis (verified by this guard before the fix): S2 allocates
			502	+	// one visible VT-frame copy during capture plus a new composed frame and its
			503	+	// base-frame clone for every live paint. Those client-sized copies, rather
			504	+	// than history, account for the ~1 MiB/op regression in the pinned benchmark.
			505	+	func TestLivePaintAllocationBudget(t *testing.T) {
			506	+	for _, tt := range []struct {
			507	+	name string
			508	+	panes int
			509	+	}{
			510	+	{name: "1tab-1pane", panes: 1},
			511	+	{name: "1tab-4panes", panes: 4},
			512	+	} {
			513	+	t.Run(tt.name, func(t *testing.T) {
			514	+	fixture := newPerformanceFixture(t, performanceConfig{size: domain.Size{Cols: 120, Rows: 40}, panes: tt.panes, historyRows: 10_000})
			515	+	allocs := testing.AllocsPerRun(20, func() {
			516	+	fixture.paintLive()
			517	+	fixture.ac.ackOutputState(fixture.ac.output.next)
			518	+	})
			519	+	require.LessOrEqual(t, allocs, float64(38), "live paint must reuse attachment-owned render scratch across warm paints")
			520	+	})
			521	+	}
			522	+	}
			523	+
			524	+	// TestCopyEnterAllocationBudget protects the parent baseline plus 10%. Copy
			525	+	// rendering borrows sealed VT history rows; allocating a copy per viewport row
			526	+	// is a production regression even though the capture itself remains immutable.
			527	+	func TestCopyEnterAllocationBudget(t *testing.T) {
			528	+	for _, tt := range []struct {
			529	+	name string
			530	+	tabs, panes, max int
			531	+	}{
			532	+	{name: "1tab-1pane", tabs: 1, panes: 1, max: 38},
			533	+	{name: "1tab-4panes", tabs: 1, panes: 4, max: 43},
			534	+	{name: "4tabs-1pane", tabs: 4, panes: 1, max: 42},
			535	+	{name: "4tabs-4panes", tabs: 4, panes: 4, max: 48},
			536	+	{name: "8tabs-1pane", tabs: 8, panes: 1, max: 50},
			537	+	} {
			538	+	t.Run(tt.name, func(t *testing.T) {
			539	+	fixture := newPerformanceFixture(t, performanceConfig{size: domain.Size{Cols: 120, Rows: 40}, tabs: tt.tabs, panes: tt.panes, historyRows: 10_000})
			540	+	run := func() {
			541	+	fixture.d.enterCopyMode(fixture.sess, fixture.ac)
			542	+	fixture.ac.ackOutputState(fixture.ac.output.next)
			543	+	}
			544	+
			545	+	// Always exercise and validate copy entry. The race detector adds
			546	+	// instrumentation allocations which are not reported by the parent
			547	+	// benchmark, so its allocation count cannot represent this budget.
			548	+	run()
			549	+	require.True(t, fixture.copyModeActive(), "copy entry must install a history-backed mode")
			550	+	if !copyEnterAllocationBudgetEnabled {
			551	+	return
			552	+	}
			553	+
			554	+	allocs := testing.AllocsPerRun(20, run)
			555	+	require.LessOrEqual(t, allocs, float64(tt.max), "copy enter must stay within 10%% of the parent allocation baseline")
			556	+	})
			557	+	}
			558	+	}
			559	+
170		func TestPerformanceFixturePaintLiveUsesPrecomputedAlternatingWrites(t *testing.T) {	560		func TestPerformanceFixturePaintLiveUsesPrecomputedAlternatingWrites(t *testing.T) {
171		fixture := newPerformanceFixture(t, performanceConfig{})	561		fixture := newPerformanceFixture(t, performanceConfig{})
172			562

@@ -189,29 +579,119 @@ type performanceConfig struct {
189		tabs int	579		tabs int
190		panes int	580		panes int
191		historyRows int	581		historyRows int
			582	+	ptyErrors []error
192		}	583		}
193			584
194		type performanceMetrics struct {	585		type performanceMetrics struct {
195	-	outputFrames uint64	586	+	outputFrames uint64
196	-	outputBytes uint64	587	+	outputBytes uint64
197	-	outputPayloadBytes uint64	588	+	outputPayloadBytes uint64
198	-	snapshotWrites uint64	589	+	snapshotWrites uint64
199	-	snapshotBytes uint64	590	+	snapshotBytes uint64
			591	+	coordinatorInvalidations uint64
			592	+	coordinatorWakes uint64
			593	+	coordinatorCoalesced uint64
			594	+	renderCaptures uint64
			595	+	renderCompositions uint64
			596	+	renderEmissions uint64
			597	+	resizeRequests uint64
			598	+	resizeCommits uint64
			599	+	resizeCommitNanos uint64
			600	+	skippedEpochs uint64
			601	+	ptyFailures uint64
			602	+	ptyRetries uint64
			603	+	frameGapEpochs uint64
			604	+	}
			605	+
			606	+	// performanceClock makes transition latency a reproducible counter rather
			607	+	// than a wall-clock observation. A nil timer channel deliberately executes
			608	+	// coordinator work synchronously.
			609	+	type performanceClock struct {
			610	+	mu sync.Mutex
			611	+	now time.Time
			612	+	}
			613	+
			614	+	func (c *performanceClock) Now() time.Time {
			615	+	c.mu.Lock()
			616	+	defer c.mu.Unlock()
			617	+	c.now = c.now.Add(time.Microsecond)
			618	+	return c.now
			619	+	}
			620	+	func (*performanceClock) NewTimer(time.Duration) ports.Timer { return performanceTimer{} }
			621	+
			622	+	type performanceTimer struct{}
			623	+
			624	+	func (performanceTimer) C() <-chan time.Time { return nil }
			625	+	func (performanceTimer) Reset(time.Duration) bool { return true }
			626	+	func (performanceTimer) Stop() bool { return true }
			627	+
			628	+	// scriptedPerformancePTY records every requested terminal size and returns a
			629	+	// caller-provided error script. It has no reads or scheduling side effects.
			630	+	type scriptedPerformancePTY struct {
			631	+	mu sync.Mutex
			632	+	sizes []domain.Size
			633	+	errs []error
			634	+	fails uint64
			635	+	}
			636	+
			637	+	func (p *scriptedPerformancePTY) Resize(size domain.Size) error {
			638	+	p.mu.Lock()
			639	+	defer p.mu.Unlock()
			640	+	p.sizes = append(p.sizes, size)
			641	+	index := len(p.sizes) - 1
			642	+	if index < len(p.errs) && p.errs[index] != nil {
			643	+	p.fails++
			644	+	return p.errs[index]
			645	+	}
			646	+	return nil
			647	+	}
			648	+	func (*scriptedPerformancePTY) Read([]byte) (int, error) { return 0, io.EOF }
			649	+	func (*scriptedPerformancePTY) Write(b []byte) (int, error) { return len(b), nil }
			650	+	func (*scriptedPerformancePTY) Close() error { return nil }
			651	+	func (*scriptedPerformancePTY) Pid() int { return 0 }
			652	+	func (*scriptedPerformancePTY) ForegroundPgid() (int, error) { return 0, nil }
			653	+	func (p *scriptedPerformancePTY) requested() []domain.Size {
			654	+	p.mu.Lock()
			655	+	defer p.mu.Unlock()
			656	+	return append([]domain.Size(nil), p.sizes...)
			657	+	}
			658	+	func (p *scriptedPerformancePTY) metrics() (uint64, uint64) {
			659	+	p.mu.Lock()
			660	+	defer p.mu.Unlock()
			661	+	return p.fails, uint64(len(p.sizes))
			662	+	}
			663	+	func (p *scriptedPerformancePTY) setErrors(errs []error) {
			664	+	p.mu.Lock()
			665	+	defer p.mu.Unlock()
			666	+	p.errs = append([]error(nil), errs...)
			667	+	p.sizes = nil
			668	+	p.fails = 0
200		}	669		}
201			670
202		type performanceFixture struct {	671		type performanceFixture struct {
203	-	t testing.TB	672	+	t testing.TB
204	-	d *Daemon	673	+	d *Daemon
205	-	sess *session	674	+	sess *session
206	-	ac *attachedClient	675	+	ac *attachedClient
207	-	output *countingOutputTransport	676	+	output *countingOutputTransport
208	-	snaps *countingSnapshotStore	677	+	snaps *countingSnapshotStore
209	-	activePane *pane	678	+	activePane *pane
210	-	liveWrites [][]byte	679	+	liveWrites [][]byte
211	-	paints int	680	+	paints int
212	-	resizeSizes [2]domain.Size	681	+	resizeSizes [2]domain.Size
213	-	resizes int	682	+	resizes int
214	-	resizedSize domain.Size	683	+	resizedSize domain.Size
			684	+	pty *scriptedPerformancePTY
			685	+	clock *performanceClock
			686	+	resizeRequests uint64
			687	+	resizeCommits uint64
			688	+	resizeCommitNanos uint64
			689	+	ptyFailures uint64
			690	+	ptyRetries uint64
			691	+	coordinatorBaseline renderCoordinatorBurstMetricsSnapshot
			692	+	renderCaptures uint64
			693	+	renderCompositions uint64
			694	+	renderEmissions uint64
215		}	695		}
216			696
217		func newPerformanceFixture(t testing.TB, config performanceConfig) *performanceFixture {	697		func newPerformanceFixture(t testing.TB, config performanceConfig) *performanceFixture {

@@ -228,8 +708,12 @@ func newPerformanceFixture(t testing.TB, config performanceConfig) *performanceF
228		if config.historyRows == 0 {	708		if config.historyRows == 0 {
229		config.historyRows = 1	709		config.historyRows = 1
230		}	710		}
			711	+	pty := &scriptedPerformancePTY{}
231		ptys := make([]ports.PTY, config.tabs)	712		ptys := make([]ports.PTY, config.tabs)
			713	+	ptys[0] = pty
232		d, sess, ac, _ := newManualSessionWithPTYs(t, ptys...)	714		d, sess, ac, _ := newManualSessionWithPTYs(t, ptys...)
			715	+	clock := &performanceClock{}
			716	+	d.clock = clock
233		output := &countingOutputTransport{}	717		output := &countingOutputTransport{}
234		ac.tr = output	718		ac.tr = output
235		ac.size = config.size	719		ac.size = config.size

@@ -244,18 +728,34 @@ func newPerformanceFixture(t testing.TB, config performanceConfig) *performanceF
244		ac: ac,	728		ac: ac,
245		output: output,	729		output: output,
246		snaps: &countingSnapshotStore{},	730		snaps: &countingSnapshotStore{},
			731	+	pty: pty,
			732	+	clock: clock,
247		liveWrites: [][]byte{[]byte("\x1b[1;1HA\x1b[2;2HA"), []byte("\x1b[1;1HB\x1b[2;2HB")},	733		liveWrites: [][]byte{[]byte("\x1b[1;1HA\x1b[2;2HA"), []byte("\x1b[1;1HB\x1b[2;2HB")},
248		resizeSizes: [2]domain.Size{{Cols: 100, Rows: 30}, config.size},	734		resizeSizes: [2]domain.Size{{Cols: 100, Rows: 30}, config.size},
249		}	735		}
250		WithSnapshotStore(fixture.snaps)(d)	736		WithSnapshotStore(fixture.snaps)(d)
			737	+	d.startSnapshotEncodeWorker()
			738	+	t.Cleanup(d.stopSnapshotEncodeWorker)
251		for tabIndex, tb := range sess.tabs {	739		for tabIndex, tb := range sess.tabs {
252		fixture.configureTab(tb, tabIndex, config.panes, config.historyRows, tabSize(config.size))	740		fixture.configureTab(tb, tabIndex, config.panes, config.historyRows, tabSize(config.size))
253		}	741		}
254		fixture.activePane = fixture.findActivePane()	742		fixture.activePane = fixture.findActivePane()
			743	+	// Fixture construction may solve the initial layout; benchmark scripts begin
			744	+	// after that setup work and therefore have an empty call history.
			745	+	pty.setErrors(config.ptyErrors)
			746	+
			747	+	// The fixture uses the production coordinator. performanceClock's nil timer
			748	+	// channel completes the coordinator path synchronously and deterministically.
			749	+	ac.renderStages = renderStageHooks{
			750	+	capture: func() { fixture.renderCaptures++ },
			751	+	compose: func() { fixture.renderCompositions++ },
			752	+	emit: func() { fixture.renderEmissions++ },
			753	+	}
			754	+	d.attachCoordinator(sess, nil, ac, true)
255			755
256		// Prime the real renderer shadow before measurements. Subsequent paints use	756		// Prime the real renderer shadow before measurements. Subsequent paints use
257		// actual production diffs rather than the initial full frame.	757		// actual production diffs rather than the initial full frame.
258	-	d.paint(sess, ac, true)	758	+	d.paint(sess, ac, true, nil)
259		fixture.resetMetrics()	759		fixture.resetMetrics()
260		return fixture	760		return fixture
261		}	761		}

@@ -278,7 +778,7 @@ func (f performanceFixture) configureTab(tb tab, tabIndex, paneCount, historyR
278		for _, p := range tb.panes {	778		for _, p := range tb.panes {
279		p.mu.Lock()	779		p.mu.Lock()
280		width := p.screen.Frame.Width	780		width := p.screen.Frame.Width
281	-	for row := 0; p.scrollback.Len() < historyRows; row++ {	781	+	for row := 0; p.history.Len() < historyRows; row++ {
282		p.screen.Write([]byte(performanceFullWidthRow(width, tabIndex, row) + "\r\n"))	782		p.screen.Write([]byte(performanceFullWidthRow(width, tabIndex, row) + "\r\n"))
283		}	783		}
284		p.mu.Unlock()	784		p.mu.Unlock()

@@ -311,13 +811,20 @@ func (f *performanceFixture) paintLive() {
311		p.mu.Lock()	811		p.mu.Lock()
312		p.screen.Write(f.liveWrites[f.paints%len(f.liveWrites)])	812		p.screen.Write(f.liveWrites[f.paints%len(f.liveWrites)])
313		p.mu.Unlock()	813		p.mu.Unlock()
314	-	f.d.paint(f.sess, f.ac, false)	814	+	if rc := f.sess.renderCoordinator(); rc != nil {
			815	+	rc.invalidate(renderInvalidation{class: invalidateOutput, producer: "performance_benchmark_test.go"})
			816	+	} else {
			817	+	f.d.paint(f.sess, f.ac, false, nil)
			818	+	}
315		f.paints++	819		f.paints++
316		}	820		}
317			821
318		func (f *performanceFixture) captureSnapshot() {	822		func (f *performanceFixture) captureSnapshot() {
319		f.t.Helper()	823		f.t.Helper()
320		require.True(f.t, f.d.captureSession(f.sess))	824		require.True(f.t, f.d.captureSession(f.sess))
			825	+	// Worker finalization clears snapshotPending and signals snapshotChanged on
			826	+	// every success and failure path after persistence has returned.
			827	+	awaitSnapshotIdle(f.t, f.sess)
321		}	828		}
322			829
323		func (f *performanceFixture) enterCopySearch() {	830		func (f *performanceFixture) enterCopySearch() {

@@ -326,6 +833,30 @@ func (f *performanceFixture) enterCopySearch() {
326		f.d.handleInput(f.sess, f.ac, []byte("/needle\r"))	833		f.d.handleInput(f.sess, f.ac, []byte("/needle\r"))
327		}	834		}
328			835
			836	+	func (f *performanceFixture) hasHistoryTopology(tabs, panes, historyRows int) bool {
			837	+	if len(f.sess.tabs) != tabs {
			838	+	return false
			839	+	}
			840	+	for _, tb := range f.sess.tabs {
			841	+	tb.mu.Lock()
			842	+	if len(tb.panes) != panes {
			843	+	tb.mu.Unlock()
			844	+	return false
			845	+	}
			846	+	for _, p := range tb.panes {
			847	+	p.mu.Lock()
			848	+	rows := p.history.Len()
			849	+	p.mu.Unlock()
			850	+	if rows != historyRows {
			851	+	tb.mu.Unlock()
			852	+	return false
			853	+	}
			854	+	}
			855	+	tb.mu.Unlock()
			856	+	}
			857	+	return true
			858	+	}
			859	+
329		func (f *performanceFixture) searchMatches() int {	860		func (f *performanceFixture) searchMatches() int {
330		f.ac.overlays.copyMu.Lock()	861		f.ac.overlays.copyMu.Lock()
331		defer f.ac.overlays.copyMu.Unlock()	862		defer f.ac.overlays.copyMu.Unlock()

@@ -335,10 +866,48 @@ func (f *performanceFixture) searchMatches() int {
335		return len(f.ac.overlays.copyMode.Searches)	866		return len(f.ac.overlays.copyMode.Searches)
336		}	867		}
337			868
			869	+	func (f *performanceFixture) copyModeActive() bool {
			870	+	f.ac.overlays.copyMu.Lock()
			871	+	defer f.ac.overlays.copyMu.Unlock()
			872	+	return f.ac.overlays.copyMode != nil && f.ac.overlays.copySnapshot != nil && f.ac.overlays.copySnapshot.Len() >= 10_000
			873	+	}
			874	+
338		func (f *performanceFixture) resize() {	875		func (f *performanceFixture) resize() {
339	-	f.resizedSize = f.resizeSizes[f.resizes%len(f.resizeSizes)]	876	+	f.resizeTo(f.resizeSizes[f.resizes%len(f.resizeSizes)])
			877	+	}
			878	+
			879	+	func (f *performanceFixture) resizeTo(size domain.Size) {
			880	+	f.resizedSize = size
340		f.resizes++	881		f.resizes++
341	-	f.d.resize(f.sess, f.ac, f.resizedSize)	882	+	before := f.sess.renderCoordinator().resizeSnapshot()
			883	+	start := f.clock.Now()
			884	+	failuresBefore, callsBefore := f.pty.metrics()
			885	+	f.d.resize(f.sess, f.ac, size)
			886	+	after := f.sess.renderCoordinator().resizeSnapshot()
			887	+	if after.epoch > before.epoch {
			888	+	f.resizeRequests += after.epoch - before.epoch
			889	+	}
			890	+	if after.committed > before.committed {
			891	+	f.resizeCommits += after.committed - before.committed
			892	+	f.resizeCommitNanos += uint64(f.clock.Now().Sub(start).Nanoseconds())
			893	+	}
			894	+	failuresAfter, callsAfter := f.pty.metrics()
			895	+	f.ptyFailures += failuresAfter - failuresBefore
			896	+	if callsAfter > callsBefore+1 {
			897	+	f.ptyRetries += callsAfter - callsBefore - 1
			898	+	}
			899	+	}
			900	+
			901	+	func (f *performanceFixture) retryLatest() {
			902	+	epoch := f.sess.renderCoordinator().resizeSnapshot().committed
			903	+	plan := f.d.prepareResize(f.sess, f.ac.size)
			904	+	failuresBefore, callsBefore := f.pty.metrics()
			905	+	f.d.retryResizeMembers(f.sess, f.ac, f.sess.renderCoordinator().attachmentLease(f.ac), epoch, plan.members)
			906	+	failuresAfter, callsAfter := f.pty.metrics()
			907	+	f.ptyFailures += failuresAfter - failuresBefore
			908	+	if callsAfter > callsBefore {
			909	+	f.ptyRetries += callsAfter - callsBefore
			910	+	}
342		}	911		}
343			912
344		func (f *performanceFixture) resized() bool {	913		func (f *performanceFixture) resized() bool {

@@ -359,15 +928,42 @@ func (f *performanceFixture) resized() bool {
359		func (f *performanceFixture) resetMetrics() {	928		func (f *performanceFixture) resetMetrics() {
360		f.output.reset()	929		f.output.reset()
361		f.snaps.reset()	930		f.snaps.reset()
			931	+	f.renderCaptures = 0
			932	+	f.renderCompositions = 0
			933	+	f.renderEmissions = 0
			934	+	f.resizeRequests = 0
			935	+	f.resizeCommits = 0
			936	+	f.resizeCommitNanos = 0
			937	+	f.ptyFailures = 0
			938	+	f.ptyRetries = 0
			939	+	if rc := f.sess.renderCoordinator(); rc != nil {
			940	+	f.coordinatorBaseline = rc.burstMetricsSnapshot()
			941	+	}
362		}	942		}
363			943
364		func (f *performanceFixture) metrics() performanceMetrics {	944		func (f *performanceFixture) metrics() performanceMetrics {
365		output := f.output.metrics()	945		output := f.output.metrics()
366		snapshots := f.snaps.metrics()	946		snapshots := f.snaps.metrics()
367	-	return performanceMetrics{	947	+	metrics := performanceMetrics{
368		outputFrames: output.frames, outputBytes: output.bytes, outputPayloadBytes: output.payloadBytes,	948		outputFrames: output.frames, outputBytes: output.bytes, outputPayloadBytes: output.payloadBytes,
369		snapshotWrites: snapshots.writes, snapshotBytes: snapshots.bytes,	949		snapshotWrites: snapshots.writes, snapshotBytes: snapshots.bytes,
			950	+	renderCaptures: f.renderCaptures, renderCompositions: f.renderCompositions, renderEmissions: f.renderEmissions,
			951	+	resizeRequests: f.resizeRequests, resizeCommits: f.resizeCommits, resizeCommitNanos: f.resizeCommitNanos,
			952	+	ptyFailures: f.ptyFailures, ptyRetries: f.ptyRetries,
			953	+	}
			954	+	if metrics.resizeRequests > metrics.resizeCommits {
			955	+	metrics.skippedEpochs = metrics.resizeRequests - metrics.resizeCommits
			956	+	}
			957	+	if metrics.resizeCommits > metrics.outputFrames {
			958	+	metrics.frameGapEpochs = metrics.resizeCommits - metrics.outputFrames
			959	+	}
			960	+	if rc := f.sess.renderCoordinator(); rc != nil {
			961	+	coordinator := rc.burstMetricsSnapshot()
			962	+	metrics.coordinatorInvalidations = coordinator.invalidations - f.coordinatorBaseline.invalidations
			963	+	metrics.coordinatorWakes = coordinator.wakes - f.coordinatorBaseline.wakes
			964	+	metrics.coordinatorCoalesced = coordinator.coalesced - f.coordinatorBaseline.coalesced
370		}	965		}
			966	+	return metrics
371		}	967		}
372			968
373		const outputPayloadHeaderBytes = 24	969		const outputPayloadHeaderBytes = 24

M internal/usecase/daemon/picker.go +173 -128

@@ -1,27 +1,4 @@
1	-	// Package daemon holds vev's server-side session multiplexer use case: the	1	+	// Package daemon holds vev's server-side session multiplexer use case.
2	-	// accept loop, the ephemeral/named session registry, the per-tab PTY reader
3	-	// and VT screen, and the per-client debounced render scheduler.
4	-	//
5	-	// Concurrency model (sessions own one or more PTY-backed tabs):
6	-	//
7	-	// - Serve runs the accept loop. Each accepted connection is handled by its
8	-	// own goroutine (handleConn): it reads the first frame and routes it to a
9	-	// session create/attach, a list, or a kill.
10	-	// - Per session there are exactly two long-lived goroutines: the PTY reader
11	-	// (drains child output into the VT screen and pokes a cap-1 dirty channel)
12	-	// and the render scheduler (debounces dirties and paints the attached
13	-	// client). Both are tied to the session context and unwind when the
14	-	// session is killed (pty.Close unblocks the reader; ctx cancel stops the
15	-	// scheduler).
16	-	// - The daemon exits (Serve returns) when the last session is removed, or
17	-	// when the parent context is cancelled (graceful shutdown notifies any
18	-	// attached clients with ReasonServerShutdown).
19	-	//
20	-	// Locking: a pane's screen/scrollback and per-client renderer shadow are
21	-	// guarded by pane.mu/tab.mu as appropriate; the attached-client pointer by
22	-	// session.mu; the registry by Daemon.mu. When more than one is held the order
23	-	// is always attachedClient.sendMu > Daemon.mu > session.mu > tab.mu > pane.mu.
24	-	// The PTY reader only ever takes pane.mu, so it never blocks on a slow client.
25		package daemon	2		package daemon
26			3
27		import (	4		import (

@@ -32,7 +9,6 @@ import (
32		"github.com/bnema/vev/internal/usecase/keys"	9		"github.com/bnema/vev/internal/usecase/keys"
33		"github.com/bnema/vev/internal/usecase/layout"	10		"github.com/bnema/vev/internal/usecase/layout"
34		"github.com/bnema/vev/internal/usecase/picker"	11		"github.com/bnema/vev/internal/usecase/picker"
35	-	themeui "github.com/bnema/vev/internal/usecase/theme"
36		"github.com/bnema/vev/internal/usecase/ui"	12		"github.com/bnema/vev/internal/usecase/ui"
37		"github.com/bnema/vev/pkg/renderer"	13		"github.com/bnema/vev/pkg/renderer"
38		)	14		)

@@ -47,7 +23,7 @@ func (d Daemon) enterPicker(sess session, ac *attachedClient) {
47		ac.overlays.pickerPreview = nil	23		ac.overlays.pickerPreview = nil
48		ac.overlays.pickerMu.Unlock()	24		ac.overlays.pickerMu.Unlock()
49		d.registerPreviewForSelection(ac)	25		d.registerPreviewForSelection(ac)
50	-	d.paint(sess, ac, true)	26	+	d.invalidateRender(sess, ac, true, "picker.go")
51		}	27		}
52			28
53		func (d Daemon) pickerViews(cur session) ([]picker.SessionView, int) {	29		func (d Daemon) pickerViews(cur session) ([]picker.SessionView, int) {

@@ -141,7 +117,7 @@ func (d Daemon) handlePickerInput(ac attachedClient, data []byte) {
141		d.killPickerTarget(target)	117		d.killPickerTarget(target)
142		}	118		}
143		d.refreshPicker(ac)	119		d.refreshPicker(ac)
144	-	d.paint(sess, ac, true)	120	+	d.invalidateRender(sess, ac, true, "picker.go")
145		return	121		return
146		case 'j':	122		case 'j':
147		ac.overlays.picker.Down()	123		ac.overlays.picker.Down()

@@ -191,7 +167,7 @@ func (d Daemon) handlePickerInput(ac attachedClient, data []byte) {
191		return	167		return
192		}	168		}
193		if exit \|\| changed {	169		if exit \|\| changed {
194	-	d.paint(sess, ac, true)	170	+	d.invalidateRender(sess, ac, true, "picker.go")
195		}	171		}
196		}	172		}
197			173

@@ -207,11 +183,12 @@ func (d Daemon) retainPickerESCLocked(ac attachedClient) {
207		ac.overlays.pickerESC.timer = nil	183		ac.overlays.pickerESC.timer = nil
208		ac.overlays.pickerESC.done = nil	184		ac.overlays.pickerESC.done = nil
209		ac.overlays.picker = nil	185		ac.overlays.picker = nil
			186	+	generation := ac.overlays.pickerPreviewGeneration
210		ac.overlays.pickerMu.Unlock()	187		ac.overlays.pickerMu.Unlock()
211			188
212	-	d.unregisterPreview(ac)	189	+	d.clearPreviewGeneration(ac, generation)
213		if sess := ac.currentSession(); sess != nil {	190		if sess := ac.currentSession(); sess != nil {
214	-	d.paint(sess, ac, true)	191	+	d.invalidateRender(sess, ac, true, "picker.go")
215		}	192		}
216		})	193		})
217		}	194		}

@@ -239,89 +216,138 @@ func isPickerEscapePrefix(data []byte) bool {
239		}	216		}
240			217
241		func (d Daemon) registerPreviewForSelection(ac attachedClient) {	218		func (d Daemon) registerPreviewForSelection(ac attachedClient) {
			219	+	if ac == nil \|\| ac.overlays == nil {
			220	+	return
			221	+	}
242		ac.overlays.pickerMu.Lock()	222		ac.overlays.pickerMu.Lock()
243		var target picker.Target	223		var target picker.Target
244		var ok bool	224		var ok bool
245		if ac.overlays.picker != nil {	225		if ac.overlays.picker != nil {
246		target, ok = ac.overlays.picker.Selected()	226		target, ok = ac.overlays.picker.Selected()
247		}	227		}
			228	+	previous, previousGeneration := ac.overlays.pickerPreviewSession, ac.overlays.pickerPreviewGeneration
			229	+	ac.overlays.pickerPreviewGeneration++
			230	+	generation := ac.overlays.pickerPreviewGeneration
			231	+	ac.overlays.pickerPreviewSession = nil
			232	+	ac.overlays.pickerPreview = nil
248		ac.overlays.pickerMu.Unlock()	233		ac.overlays.pickerMu.Unlock()
249	-		234	+	d.teardownPreviewSubscription(ac, previous, previousGeneration)
250	-	var next *tab	235	+	if !ok {
251	-	if ok {	236	+	return
252	-	next = d.tabByTarget(target)	237	+	}
			238	+	next := d.tabByTarget(target)
			239	+	targetSess := d.sessionByID(target.Session)
			240	+	if next == nil \|\| targetSess == nil {
			241	+	return
253		}	242		}
254	-
255		ac.overlays.pickerMu.Lock()	243		ac.overlays.pickerMu.Lock()
256	-	old := ac.overlays.pickerPreview	244	+	// Selection may have changed while the target was resolved.
257	-	if ac.overlays.picker == nil {	245	+	selected, stillSelected := ac.overlays.picker.Selected()
258	-	next = nil	246	+	valid := ac.overlays.pickerPreviewGeneration == generation && stillSelected && selected == target
			247	+	if valid {
			248	+	ac.overlays.pickerPreview = next
			249	+	ac.overlays.pickerPreviewSession = targetSess
259		}	250		}
260	-	ac.overlays.pickerPreview = next
261		ac.overlays.pickerMu.Unlock()	251		ac.overlays.pickerMu.Unlock()
262	-		252	+	if !valid {
263	-	if old != nil && old != next {	253	+	return
264	-	old.mu.Lock()	254	+	}
265	-	if old.previewClient == ac {	255	+	// A same-session preview is already invalidated by its own coordinator.
266	-	old.previewClient = nil	256	+	// Cross-session previews subscribe to the target's producer wake instead
			257	+	// of starting a direct paint/timer path.
			258	+	if targetSess == ac.currentSession() {
			259	+	return
			260	+	}
			261	+	rc := d.attachCoordinator(targetSess, nil, nil, false)
			262	+	rc.subscribePreviewFor(ac, generation, func(renderWake) {
			263	+	if pickerPreviewCurrent(ac, targetSess, next, generation) {
			264	+	if owner := ac.currentSession(); owner != nil {
			265	+	d.invalidateRender(owner, ac, false, "picker preview")
			266	+	}
267		}	267		}
268	-	old.mu.Unlock()	268	+	})
			269	+	// subscribePreviewFor is deliberately outside pickerMu. Revalidate after it
			270	+	// returns and remove only this generation if selection changed meanwhile.
			271	+	d.revalidatePreviewSubscription(ac, rc, targetSess, next, generation)
			272	+	}
			273	+
			274	+	func pickerPreviewCurrent(ac attachedClient, targetSess session, next *tab, generation uint64) bool {
			275	+	ac.overlays.pickerMu.Lock()
			276	+	defer ac.overlays.pickerMu.Unlock()
			277	+	return ac.overlays.pickerPreviewGeneration == generation && ac.overlays.pickerPreviewSession == targetSess && ac.overlays.pickerPreview == next
			278	+	}
			279	+
			280	+	// revalidatePreviewSubscription removes a subscription that lost the picker
			281	+	// selection race after it was installed.
			282	+	func (d Daemon) revalidatePreviewSubscription(ac attachedClient, rc renderCoordinator, targetSess session, next *tab, generation uint64) {
			283	+	if !pickerPreviewCurrent(ac, targetSess, next, generation) {
			284	+	rc.teardownPreviewFor(ac, generation)
269		}	285		}
270	-	if next != nil {	286	+	}
271	-	next.mu.Lock()
272	-	next.previewClient = ac
273	-	next.mu.Unlock()
274			287
275	-	ac.overlays.pickerMu.Lock()	288	+	func (d Daemon) teardownPreviewSubscription(ac attachedClient, previous *session, generation uint64) {
276	-	keep := ac.overlays.pickerPreview == next	289	+	if previous != nil {
277	-	ac.overlays.pickerMu.Unlock()	290	+	if rc := previous.renderCoordinator(); rc != nil {
278	-	if !keep {	291	+	rc.teardownPreviewFor(ac, generation)
279	-	next.mu.Lock()
280	-	if next.previewClient == ac {
281	-	next.previewClient = nil
282	-	}
283	-	next.mu.Unlock()
284		}	292		}
285		}	293		}
286		}	294		}
287			295
288		func (d Daemon) unregisterPreview(ac attachedClient) {	296		func (d Daemon) unregisterPreview(ac attachedClient) {
			297	+	if ac == nil \|\| ac.overlays == nil {
			298	+	return
			299	+	}
289		ac.overlays.pickerMu.Lock()	300		ac.overlays.pickerMu.Lock()
290	-	old := ac.overlays.pickerPreview	301	+	generation := ac.overlays.pickerPreviewGeneration
291	-	ac.overlays.pickerPreview = nil
292		ac.overlays.pickerMu.Unlock()	302		ac.overlays.pickerMu.Unlock()
			303	+	d.clearPreviewGeneration(ac, generation)
			304	+	}
293			305
294	-	if old != nil {	306	+	// clearPreviewGeneration removes the preview only if it is still the observed
295	-	old.mu.Lock()	307	+	// selection. Teardown paths may race a new picker selection.
296	-	if old.previewClient == ac {	308	+	func (d Daemon) clearPreviewGeneration(ac attachedClient, generation uint64) bool {
297	-	old.previewClient = nil	309	+	ac.overlays.pickerMu.Lock()
298	-	}	310	+	if ac.overlays.pickerPreviewGeneration != generation {
299	-	old.mu.Unlock()	311	+	ac.overlays.pickerMu.Unlock()
			312	+	return false
300		}	313		}
			314	+	previous := ac.overlays.pickerPreviewSession
			315	+	ac.overlays.pickerPreviewGeneration++
			316	+	ac.overlays.pickerPreviewSession = nil
			317	+	ac.overlays.pickerPreview = nil
			318	+	ac.overlays.pickerMu.Unlock()
			319	+	d.teardownPreviewSubscription(ac, previous, generation)
			320	+	return true
301		}	321		}
302			322
			323	+	// clearDestroyedTabPreview clears coordinator-owned picker state for clients
			324	+	// that were previewing a removed tab. Tabs do not route render ownership.
303		func (d Daemon) clearDestroyedTabPreview(tb tab) {	325		func (d Daemon) clearDestroyedTabPreview(tb tab) {
304	-	tb.mu.Lock()	326	+	if d == nil \|\| d.sessions == nil {
305	-	previewer := tb.previewClient
306	-	tb.mu.Unlock()
307	-	if previewer == nil {
308		return	327		return
309		}	328		}
310	-	cleared := false	329	+	d.mu.Lock()
311	-	previewer.overlays.pickerMu.Lock()	330	+	var clients []*attachedClient
312	-	if previewer.overlays.pickerPreview == tb {	331	+	for _, sess := range d.sessions {
313	-	previewer.overlays.pickerPreview = nil	332	+	sess.mu.Lock()
314	-	cleared = true	333	+	if sess.client != nil {
315	-	}	334	+	clients = append(clients, sess.client)
316	-	previewer.overlays.pickerMu.Unlock()	335	+	}
317	-	tb.mu.Lock()	336	+	sess.mu.Unlock()
318	-	if tb.previewClient == previewer {
319	-	tb.previewClient = nil
320		}	337		}
321	-	tb.mu.Unlock()	338	+	d.mu.Unlock()
322	-	if cleared {	339	+	for _, ac := range clients {
323	-	if sess := previewer.currentSession(); sess != nil {	340	+	if ac == nil \|\| ac.overlays == nil {
324	-	d.paint(sess, previewer, true)	341	+	continue
			342	+	}
			343	+	ac.overlays.pickerMu.Lock()
			344	+	generation := ac.overlays.pickerPreviewGeneration
			345	+	observed := ac.overlays.pickerPreview == tb
			346	+	ac.overlays.pickerMu.Unlock()
			347	+	if observed && d.clearPreviewGeneration(ac, generation) {
			348	+	if sess := ac.currentSession(); sess != nil {
			349	+	d.invalidateRender(sess, ac, true, "picker.go")
			350	+	}
325		}	351		}
326		}	352		}
327		}	353		}

@@ -331,8 +357,9 @@ func (d Daemon) closePicker(ac attachedClient) {
331		ac.overlays.picker = nil	357		ac.overlays.picker = nil
332		ac.overlays.pickerPending = nil	358		ac.overlays.pickerPending = nil
333		d.stopPickerPendingTimerLocked(ac)	359		d.stopPickerPendingTimerLocked(ac)
			360	+	generation := ac.overlays.pickerPreviewGeneration
334		ac.overlays.pickerMu.Unlock()	361		ac.overlays.pickerMu.Unlock()
335	-	d.unregisterPreview(ac)	362	+	d.clearPreviewGeneration(ac, generation)
336		}	363		}
337			364
338		func (d Daemon) tabByTarget(target picker.Target) tab {	365		func (d Daemon) tabByTarget(target picker.Target) tab {

@@ -362,19 +389,19 @@ func (d Daemon) switchToTarget(sess session, ac *attachedClient, target picker
362		}	389		}
363		targetSess := d.sessionByID(target.Session)	390		targetSess := d.sessionByID(target.Session)
364		if targetSess == nil {	391		if targetSess == nil {
365	-	d.paint(sess, ac, true)	392	+	d.invalidateRender(sess, ac, true, "picker.go")
366		return false	393		return false
367		}	394		}
368		if targetSess == sess {	395		if targetSess == sess {
369		if sess.switchTab(target.TabIndex) {	396		if sess.switchTab(target.TabIndex) {
370		d.activateTab(sess, sess.activeTab())	397		d.activateTab(sess, sess.activeTab())
371		}	398		}
372	-	d.paint(sess, ac, true)	399	+	d.invalidateRender(sess, ac, true, "picker.go")
373		return true	400		return true
374		}	401		}
375		old := d.stealClientForTarget(sess, ac, targetSess, target)	402		old := d.stealClientForTarget(sess, ac, targetSess, target)
376		if ac.currentSession() != targetSess {	403		if ac.currentSession() != targetSess {
377	-	d.paint(sess, ac, true)	404	+	d.invalidateRender(sess, ac, true, "picker.go")
378		return false	405		return false
379		}	406		}
380		if old != nil && old != ac {	407		if old != nil && old != ac {

@@ -407,16 +434,22 @@ func (d Daemon) stealClientForTarget(from session, ac *attachedClient, targetS
407		targetSess.mu.Lock()	434		targetSess.mu.Lock()
408		old := targetSess.client	435		old := targetSess.client
409		targetSess.terminal = term	436		targetSess.terminal = term
410	-	targetSess.client = ac
411		if target.TabIndex >= 0 && target.TabIndex < len(targetSess.tabs) {	437		if target.TabIndex >= 0 && target.TabIndex < len(targetSess.tabs) {
412		targetSess.active = target.TabIndex	438		targetSess.active = target.TabIndex
413		}	439		}
414		targetSess.mu.Unlock()	440		targetSess.mu.Unlock()
415			441
416	-	d.touchMRU(targetSess)	442	+	// Do not expose targetSess.client until the target coordinator has claimed
			443	+	// this attachment and the output dependency chain has been rebased.
			444	+	cleanups := d.handoffCoordinator(from, targetSess, old, ac)
417		ac.setSession(targetSess)	445		ac.setSession(targetSess)
			446	+	targetSess.mu.Lock()
			447	+	targetSess.client = ac
			448	+	targetSess.mu.Unlock()
			449	+	d.touchMRU(targetSess)
418		ac.recordPreviousSession(from)	450		ac.recordPreviousSession(from)
419		d.mu.Unlock()	451		d.mu.Unlock()
			452	+	finishRenderLifecycleCleanups(cleanups)
420		return old	453		return old
421		}	454		}
422			455

@@ -425,7 +458,7 @@ func (d Daemon) resumeStoppedAndSwitch(from session, ac *attachedClient, targe
425		stopped, ok := d.stopped[target.Name]	458		stopped, ok := d.stopped[target.Name]
426		if !ok \|\| stopped.purging {	459		if !ok \|\| stopped.purging {
427		d.mu.Unlock()	460		d.mu.Unlock()
428	-	d.paint(from, ac, true)	461	+	d.invalidateRender(from, ac, true, "picker.go")
429		return false	462		return false
430		}	463		}
431		from.mu.Lock()	464		from.mu.Lock()

@@ -441,20 +474,22 @@ func (d Daemon) resumeStoppedAndSwitch(from session, ac *attachedClient, targe
441		from.mu.Unlock()	474		from.mu.Unlock()
442		d.mu.Unlock()	475		d.mu.Unlock()
443		d.log.Warn("resuming stopped session failed", "err", err, "session", target.Name)	476		d.log.Warn("resuming stopped session failed", "err", err, "session", target.Name)
444	-	d.paint(from, ac, true)	477	+	d.invalidateRender(from, ac, true, "picker.go")
445		return false	478		return false
446		}	479		}
447		from.client = nil	480		from.client = nil
448		ac.setSession(nil)	481		ac.setSession(nil)
449		from.mu.Unlock()	482		from.mu.Unlock()
			483	+
			484	+	cleanups := d.handoffCoordinator(from, targetSess, nil, ac)
			485	+	ac.setSession(targetSess)
450		targetSess.mu.Lock()	486		targetSess.mu.Lock()
451		targetSess.client = ac	487		targetSess.client = ac
452		targetSess.mu.Unlock()	488		targetSess.mu.Unlock()
453	-
454		d.touchMRU(targetSess)	489		d.touchMRU(targetSess)
455	-	ac.setSession(targetSess)
456		ac.recordPreviousSession(from)	490		ac.recordPreviousSession(from)
457		d.mu.Unlock()	491		d.mu.Unlock()
			492	+	finishRenderLifecycleCleanups(cleanups)
458		d.firstPaint(targetSess, ac, ac.size)	493		d.firstPaint(targetSess, ac, ac.size)
459		return true	494		return true
460		}	495		}

@@ -496,49 +531,59 @@ func (d Daemon) refreshPicker(ac attachedClient) {
496		d.registerPreviewForSelection(ac)	531		d.registerPreviewForSelection(ac)
497		}	532		}
498			533
499	-	func composePickerClientFrame(model *picker.Model, preview picker.Preview, base renderer.Frame, styles ...themeStyles) (renderer.Frame, []renderer.Damage) {
500	-	styleSet := resolveThemeStyles(styles)
501	-	renderStyles := picker.RenderStyles{
502	-	Selection: styleSet.selection,
503	-	SelectionName: styleSet.pickerSelectionName,
504	-	SelectionMuted: styleSet.pickerSelectionMuted,
505	-	Name: styleSet.pickerName,
506	-	Detail: styleSet.paletteDesc,
507	-	Base: renderer.DefaultStyle(),
508	-	}
509	-	return composeModalClientFrame(base, pickerModal, styleSet, styleSet.selection, func(size domain.Size, _ ...renderer.Style) renderer.Frame {
510	-	return model.Render(size, preview, renderStyles)
511	-	})
512	-	}
513	-
514		func snapshotPickerPreview(tb *tab) picker.Preview {	534		func snapshotPickerPreview(tb *tab) picker.Preview {
515		if tb == nil {	535		if tb == nil {
516		return picker.Preview{}	536		return picker.Preview{}
517		}	537		}
518		tb.mu.Lock()	538		tb.mu.Lock()
519	-	defer tb.mu.Unlock()	539	+	layoutSnap := solveTabLayoutLocked(tb)
520	-	return pickerPreviewFromLockedTab(tb)	540	+	if !layoutSnap.ok {
521	-	}	541	+	p := tb.focusedPane()
522	-		542	+	if p == nil {
523	-	func pickerPreviewFromLockedTab(tb *tab) picker.Preview {	543	+	tb.mu.Unlock()
524	-	p := tb.focusedPane()	544	+	return picker.Preview{}
525	-	if p == nil {	545	+	}
526	-	return picker.Preview{}
527	-	}
528	-	if tb.tree == nil \|\| tb.tree.Root == nil \|\| tb.tree.Root.Kind == layout.Leaf {
529		p.mu.Lock()	546		p.mu.Lock()
530	-	defer p.mu.Unlock()	547	+	preview := pickerPreviewFromLockedPane(p)
531	-	return pickerPreviewFromLockedPane(p)	548	+	p.mu.Unlock()
			549	+	tb.mu.Unlock()
			550	+	return preview
			551	+	}
			552	+	state := capturedRenderState{layout: capturedTabLayout{
			553	+	root: tb.tree.Clone().Root, area: layoutSnap.area, focus: layoutSnap.focus,
			554	+	placements: append([]layout.Placement(nil), layoutSnap.placements...),
			555	+	fingerprint: layoutSnap.fingerprint, valid: true,
			556	+	}}
			557	+	for _, placement := range layoutSnap.placements {
			558	+	p := tb.panes[placement.ID]
			559	+	if p == nil {
			560	+	continue
			561	+	}
			562	+	visible := placement.Content
			563	+	if placement.Collapsed {
			564	+	visible = domain.Rect{}
			565	+	}
			566	+	p.mu.Lock()
			567	+	captured := capturePaneRenderStateLocked(p, visible)
			568	+	p.mu.Unlock()
			569	+	captured.placement = placement
			570	+	captured.focused = placement.ID == layoutSnap.focus
			571	+	state.panes = append(state.panes, captured)
532		}	572		}
			573	+	tb.mu.Unlock()
			574	+	return pickerPreviewFromCapturedRender(state)
			575	+	}
533			576
534	-	area := domain.Rect{Width: tb.size.Cols, Height: tb.size.Rows}	577	+	func pickerPreviewFromCapturedRender(state capturedRenderState) picker.Preview {
535	-	if area.Width <= 0 \|\| area.Height <= 0 {	578	+	composed := composeFrame(state, composeCacheInput{}, composeCacheInput{}).frame
536	-	p.mu.Lock()	579	+	if composed.Height < 2 {
537	-	defer p.mu.Unlock()	580	+	return pickerPreviewFromFrame(composed)
538	-	return pickerPreviewFromLockedPane(p)	581	+	}
			582	+	rows := make([][]renderer.Cell, composed.Height-2)
			583	+	for y := range rows {
			584	+	rows[y] = append([]renderer.Cell(nil), composed.Row(y+1)...)
539		}	585		}
540	-	frame, _ := composeTabFrame(tb, area, themeui.Theme{})	586	+	return picker.Preview{Rows: rows, Width: composed.Width, Height: len(rows)}
541	-	return pickerPreviewFromFrame(frame)
542		}	587		}
543			588
544		func pickerPreviewFromLockedPane(p *pane) picker.Preview {	589		func pickerPreviewFromLockedPane(p *pane) picker.Preview {

M internal/usecase/daemon/picker_test.go +151 -126

 import (
 	"context"
-	"sync"
+	"strings"
 	"testing"
 	"time"
 	"github.com/bnema/vev/internal/usecase/keys"
 	"github.com/bnema/vev/internal/usecase/layout"
 	"github.com/bnema/vev/internal/usecase/picker"
+	"github.com/bnema/vev/pkg/vt"
 )
 // --- test doubles -----------------------------------------------------------
 	require.Eventually(t, func() bool { return !ac.overlays.pickerActive() }, time.Second, 5*time.Millisecond)
 }
+func TestBackSessionFirstResetDoesNotReuseSamePaneIDCapture(t *testing.T) {
+	p1, releasePTY1 := newBlockingPTY(t)
+	p2, releasePTY2 := newBlockingPTY(t)
+	defer releasePTY1()
+	defer releasePTY2()
+	d, source, ac, sends := newManualSessionWithPTYs(t, p1)
+	source.id, source.name = "source", "source"
+	delete(d.sessions, domain.SessionID("manual"))
+	d.sessions[source.id] = source
+	target := &session{
+		id: "target", name: "target", ctx: source.ctx, cancel: func() {},
+		tabs: []*tab{newTab(p2, domain.Size{Cols: 80, Rows: 22})},
+	}
+	d.sessions[target.id] = target
+	sourcePane := source.tabs[0].focusedPane()
+	targetPane := target.tabs[0].focusedPane()
+	require.Equal(t, layout.PaneID("pane-1"), sourcePane.id, "source uses reusable pane-1")
+	require.Equal(t, sourcePane.id, targetPane.id, "target deliberately reuses pane-1")
+	sourcePane.screen.Write([]byte("SOURCE"))
+	client := vt.NewScreen(80, 25)
+	d.firstPaint(source, ac, ac.size)
+	sourceOutput := mustApplyOutput(t, client, awaitFrame(t, sends, ports.MsgOutput))
+	ac.ackOutputState(sourceOutput.NewStateNum)
+	targetPane.screen.Write([]byte("TARGET"))
+	targetPane.screen.ClearDamage() // TARGET is already rendered and has no pending VT damage.
+	require.Empty(t, targetPane.screen.Damage(), "target deliberately has no pending VT damage")
+	ac.previousSession.Set(target)
+	// Exercise the user-facing previous-session route, which delegates to the
+	// real switchToTarget hand-off and immediately emits its required reset.
+	d.backSession(source, ac)
+	firstReset := awaitFrame(t, sends, ports.MsgOutput)
+	out := mustApplyOutput(t, client, firstReset)
+	require.Zero(t, out.BaseStateNum, "the first target frame must be the reset, not an eventual repair")
+	frame := strings.Join(frameRows(client.Frame), "\n")
+	require.NotContains(t, frame, "SOURCE", "first target reset must not reuse source capture")
+	require.Contains(t, frame, "TARGET", "first target reset must immediately show clean target VT state")
+}
 func TestPickerCrossSessionSwitchDetachesExistingClient(t *testing.T) {
 	p1, releasePTY1 := newBlockingPTY(t)
 	p2, releasePTY2 := newBlockingPTY(t)
 	awaitFrame(t, sends1, ports.MsgOutput)
 }
+func TestPickerDisplacementCancelsSupersededResize(t *testing.T) {
+	p1, releasePTY1 := newBlockingPTY(t)
+	p2, releasePTY2 := newBlockingPTY(t)
+	defer releasePTY1()
+	defer releasePTY2()
+	clock := &signalClock{timers: make(chan *signalTimer, 2)}
+	d := newTestDaemon(t, nil, clock)
+	tr1, _ := newCapturingTransport(t)
+	tr2, _ := newCapturingTransport(t)
+	ac1 := &attachedClient{tr: tr1, output: newOutputStateStream(), size: domain.Size{Cols: 80, Rows: 24}}
+	ac2 := &attachedClient{tr: tr2, output: newOutputStateStream(), size: domain.Size{Cols: 80, Rows: 24}}
+	sctx1, cancel1 := context.WithCancel(d.serveCtx)
+	sctx2, cancel2 := context.WithCancel(d.serveCtx)
+	defer cancel1()
+	defer cancel2()
+	sess1 := &session{id: "s1", name: "alpha", ctx: sctx1, cancel: cancel1, tabs: []*tab{newTestTabWithContext(p1, sctx1, cancel1)}, client: ac1}
+	sess2 := &session{id: "s2", name: "beta", ctx: sctx2, cancel: cancel2, tabs: []*tab{newTestTabWithContext(p2, sctx2, cancel2)}, client: ac2}
+	ac1.setSession(sess1)
+	ac2.setSession(sess2)
+	d.sessions[sess1.id], d.sessions[sess2.id] = sess1, sess2
+	d.resize(sess2, ac2, domain.Size{Cols: 100, Rows: 24})
+	timer := <-clock.timers
+	done := captureResizeCallbackDone(t, sess2.renderCoordinator())
+	before := sess2.renderCoordinator().resizeSnapshot().epoch
+	require.Same(t, ac2, d.stealClientForTarget(sess1, ac1, sess2, picker.Target{Session: sess2.id}))
+	// The target coordinator invalidates its scheduled epoch during handoff;
+	// no attachment-local timer survives the transfer.
+	require.Equal(t, before, sess2.renderCoordinator().resizeSnapshot().epoch)
+	timer.ch <- time.Time{}
+	awaitTestCompletion(t, done, "superseded resize callback did not complete")
+}
 func TestPickerStalePaintAfterSessionSwitchSendsNoFrame(t *testing.T) {
 	p1, releasePTY1 := newBlockingPTY(t)
 	p2, releasePTY2 := newBlockingPTY(t)
 	oldPane.screen.Write([]byte("stale-damage"))
 	require.NotEmpty(t, oldPane.screen.Damage())
-	d.paint(sess1, ac1, false)
+	d.paint(sess1, ac1, false, nil)
 	require.Zero(t, len(sends1), "stale paint from old session sent a frame")
 	require.NotEmpty(t, oldPane.screen.Damage(), "stale paint from old session consumed damage")
 	sess1.tabs[0].focusedPane().screen.Write([]byte("paint while switching"))
 	paintDone := make(chan struct{})
 	go func() {
-		d.paint(sess1, ac, false)
+		d.paint(sess1, ac, false, nil)
 		close(paintDone)
 	}()
 	<-enteredSend
 	defer releasePTY3()
 	var openedEnv []string
 	f := portsmocks.NewMockPTYFactory(t)
-	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything).RunAndReturn(
+	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything).RunAndReturn(
-		func(_ string, _ []string, env []string, _ string, sz domain.Size) (ports.PTY, error) {
+		func(_ context.Context, _ string, _ []string, env []string, _ string, sz domain.Size) (ports.PTY, error) {
 			if sz != (domain.Size{Cols: 80, Rows: 22}) {
 				return newQuietPTY(), nil
 			}
 	require.NotContains(t, openedEnv, "TERM=xterm-direct")
 }
+func TestPickerPreviewGenerationRejectsStaleSubscriptionReplacement(t *testing.T) {
+	rc := newRenderCoordinator(renderCoordinatorOptions{})
+	viewer := &attachedClient{}
+	newer := func(renderWake) {}
+	require.True(t, rc.subscribePreviewFor(viewer, 2, newer))
+	require.False(t, rc.subscribePreviewFor(viewer, 1, func(renderWake) {}))
+	rc.teardownPreviewFor(viewer, 1)
+	rc.mu.Lock()
+	subscription, ok := rc.previewWakes[viewer]
+	rc.mu.Unlock()
+	require.True(t, ok)
+	require.Equal(t, uint64(2), subscription.generation)
+	require.NotNil(t, subscription.fn)
+}
+func TestPickerPreviewPostSubscribeRevalidationKeepsNewerSubscription(t *testing.T) {
+	d := newTestDaemon(t, nil, stubClock{})
+	viewer := &attachedClient{}
+	viewer.initOverlays()
+	target := &session{}
+	tab := &tab{}
+	rc := newRenderCoordinator(renderCoordinatorOptions{})
+	viewer.overlays.pickerMu.Lock()
+	viewer.overlays.pickerPreviewGeneration = 2
+	viewer.overlays.pickerPreviewSession = target
+	viewer.overlays.pickerPreview = tab
+	viewer.overlays.pickerMu.Unlock()
+	require.True(t, rc.subscribePreviewFor(viewer, 2, func(renderWake) {}))
+	// This is the exact post-subscribe check from registerPreviewForSelection.
+	// It represents a generation-1 subscriber resuming after generation 2 won.
+	d.revalidatePreviewSubscription(viewer, rc, target, tab, 1)
+	rc.mu.Lock()
+	subscription, ok := rc.previewWakes[viewer]
+	rc.mu.Unlock()
+	require.True(t, ok)
+	require.Equal(t, uint64(2), subscription.generation)
+}
+func TestPickerDestroyedTabCleanupDoesNotClearNewerGeneration(t *testing.T) {
+	d := newTestDaemon(t, nil, stubClock{})
+	viewer := &attachedClient{}
+	viewer.initOverlays()
+	newerSession := &session{}
+	newerTab := &tab{}
+	rc := newRenderCoordinator(renderCoordinatorOptions{})
+	viewer.overlays.pickerMu.Lock()
+	viewer.overlays.pickerPreviewGeneration = 2
+	viewer.overlays.pickerPreviewSession = newerSession
+	viewer.overlays.pickerPreview = newerTab
+	viewer.overlays.pickerMu.Unlock()
+	require.True(t, rc.subscribePreviewFor(viewer, 2, func(renderWake) {}))
+	require.False(t, d.clearPreviewGeneration(viewer, 1))
+	require.True(t, pickerPreviewCurrent(viewer, newerSession, newerTab, 2))
+	rc.mu.Lock()
+	_, subscribed := rc.previewWakes[viewer]
+	rc.mu.Unlock()
+	require.True(t, subscribed)
+}
 func TestPickerPreviewSinglePaneSnapshotsFocusedPane(t *testing.T) {
 	tb := newTab(nil, domain.Size{Cols: 10, Rows: 3})
 	p := tb.focusedPane()
 	require.Equal(t, 'R', preview.Rows[2][21].Rune, "expanded stacked pane content should be included")
 }
-func TestPickerLivePreviewRepaintsInactiveTab(t *testing.T) {
-	d, sess, ac, sends, releases := newManualTabSession(t, 2)
-	defer func() {
-		for _, release := range releases {
-			release()
-		}
-	}()
-	d.enterPicker(sess, ac)
-	awaitFrame(t, sends, ports.MsgOutput)
-	d.handleInput(sess, ac, []byte("j"))
-	awaitFrame(t, sends, ports.MsgOutput)
-	previewTab := sess.tabs[1]
-	previewTab.mu.Lock()
-	previewTab.focusedPane().screen.Write([]byte("inactive-preview-live"))
-	previewTab.mu.Unlock()
-	d.render(sess, previewTab, previewTab.focusedPane())
-	previewOut := awaitFrame(t, sends, ports.MsgOutput)
-	previewMsg, err := ports.UnmarshalOutput(previewOut.Payload)
-	require.NoError(t, err)
-	require.Contains(t, string(previewMsg.Data), "inactive-preview-live")
-}
-func TestPickerLivePreviewRepaintsCrossSessionTab(t *testing.T) {
-	p1, releasePTY1 := newBlockingPTY(t)
-	p2, releasePTY2 := newBlockingPTY(t)
-	defer releasePTY1()
-	defer releasePTY2()
-	d := newTestDaemon(t, nil, stubClock{})
-	tr1, sends1 := newCapturingTransport(t)
-	tr2, _ := newCapturingTransport(t)
-	ac1 := &attachedClient{tr: tr1, output: newOutputStateStream(), size: domain.Size{Cols: 80, Rows: 24}}
-	ac1.initOverlays()
-	ac2 := &attachedClient{tr: tr2, output: newOutputStateStream(), size: domain.Size{Cols: 80, Rows: 24}}
-	ac2.initOverlays()
-	sctx1, cancel1 := context.WithCancel(d.serveCtx)
-	sctx2, cancel2 := context.WithCancel(d.serveCtx)
-	defer cancel1()
-	defer cancel2()
-	sess1 := &session{id: "s1", name: "alpha", ctx: sctx1, cancel: cancel1, tabs: []*tab{newTestTabWithContext(p1, sctx1, cancel1)}, client: ac1}
-	sess2 := &session{id: "s2", name: "beta", ctx: sctx2, cancel: cancel2, tabs: []*tab{newTestTabWithContext(p2, sctx2, cancel2)}, client: ac2}
-	ac1.setSession(sess1)
-	ac2.setSession(sess2)
-	ac1.keys = keys.NewRouter(d.clock, daemonKeyHandler{d: d, ac: ac1}, nil)
-	ac2.keys = keys.NewRouter(d.clock, daemonKeyHandler{d: d, ac: ac2}, nil)
-	d.sessions[sess1.id] = sess1
-	d.sessions[sess2.id] = sess2
-	d.enterPicker(sess1, ac1)
-	awaitFrame(t, sends1, ports.MsgOutput)
-	d.handleInput(sess1, ac1, []byte("j"))
-	awaitFrame(t, sends1, ports.MsgOutput)
-	previewTab := sess2.tabs[0]
-	previewTab.mu.Lock()
-	previewTab.focusedPane().screen.Write([]byte("cross-session-preview-live"))
-	previewTab.mu.Unlock()
-	d.render(sess2, previewTab, previewTab.focusedPane())
-	previewOut := awaitFrame(t, sends1, ports.MsgOutput)
-	previewMsg, err := ports.UnmarshalOutput(previewOut.Payload)
-	require.NoError(t, err)
-	require.Contains(t, string(previewMsg.Data), "cross-session-preview-live")
-}
-func TestPickerOpenCloseNavigationConcurrentWithRenderRace(t *testing.T) {
-	d, sess, ac, sends, releases := newManualTabSession(t, 3)
-	defer func() {
-		for _, release := range releases {
-			release()
-		}
-	}()
-	done := make(chan struct{})
-	var drain sync.WaitGroup
-	drain.Go(func() {
-		for {
-			select {
-			case <-sends:
-			case <-done:
-				return
-			}
-		}
-	})
-	var wg sync.WaitGroup
-	for i := range 100 {
-		wg.Go(func() {
-			d.enterPicker(sess, ac)
-		})
-		wg.Go(func() {
-			d.handlePickerInput(ac, []byte("j"))
-			d.handlePickerInput(ac, []byte("k"))
-		})
-		wg.Go(func() {
-			d.closePicker(ac)
-		})
-		tb := sess.tabs[i%len(sess.tabs)]
-		wg.Go(func() {
-			tb.mu.Lock()
-			p := tb.focusedPane()
-			tb.mu.Unlock()
-			if p != nil {
-				p.mu.Lock()
-				p.screen.Write([]byte("render-race"))
-				p.mu.Unlock()
-			}
-			d.render(sess, tb, p)
-		})
-	}
-	wg.Wait()
-	close(done)
-	drain.Wait()
-}
 func TestPickerModalGeometry(t *testing.T) {
 	base := domain.Size{Cols: 100, Rows: 40}
 	d.resize(sess, ac, domain.Size{Cols: 100, Rows: 30})
 	// Picker recomposition is the relevant event here; do not depend on a real
 	// resize-idle timer in this synchronous rendering test.
-	d.paint(sess, ac, false)
+	d.paint(sess, ac, false, nil)
 	out := awaitFrame(t, sends, ports.MsgOutput)
 	msg, err := ports.UnmarshalOutput(out.Payload)
 	var opens [][]string
 	f := portsmocks.NewMockPTYFactory(t)
 	normalSize := domain.Size{Cols: sz.Cols, Rows: sz.Rows - 2}
-	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, normalSize).RunAndReturn(
+	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything, normalSize).RunAndReturn(
-		func(_ string, _ []string, env []string, _ string, _ domain.Size) (ports.PTY, error) {
+		func(_ context.Context, _ string, _ []string, env []string, _ string, _ domain.Size) (ports.PTY, error) {
 			opens = append(opens, append([]string(nil), env...))
 			if len(opens) == 1 {
 				return p1, nil
 		},
 	).Twice()
 	floating := newQuietPTY()
-	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.MatchedBy(func(got domain.Size) bool {
+	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.MatchedBy(func(got domain.Size) bool {
 		return got != normalSize && got.Valid()
 	})).Return(floating, nil).Once()
 	d := newTestDaemon(t, f, stubClock{})

M internal/usecase/daemon/prompt.go +5 -11

 	"github.com/bnema/vev/internal/domain"
 	promptui "github.com/bnema/vev/internal/usecase/prompt"
 	"github.com/bnema/vev/internal/usecase/ui"
-	"github.com/bnema/vev/pkg/renderer"
 )
 func promptModalFor(title string) ui.Modal {
 	ac.overlays.promptSubmit = submit
 	ac.overlays.promptPending = nil
 	ac.overlays.promptMu.Unlock()
-	d.paint(sess, ac, true)
+	d.invalidateRender(sess, ac, true, "prompt.go")
 }
 func (d *Daemon) closePrompt(ac *attachedClient) {
 				ac.overlays.prompt.SetError(err.Error())
 			}
 			ac.overlays.promptMu.Unlock()
-			d.paint(sess, ac, true)
+			d.invalidateRender(sess, ac, true, "prompt.go")
 			return
 		}
 		d.closePrompt(ac)
 		if current := ac.currentSession(); current != nil {
-			d.paint(current, ac, true)
+			d.invalidateRender(current, ac, true, "prompt.go")
 		}
 		return
 	}
 	if exit {
 		d.closePrompt(ac)
-		d.paint(sess, ac, true)
+		d.invalidateRender(sess, ac, true, "prompt.go")
 		return
 	}
 	if changed {
-		d.paint(sess, ac, true)
+		d.invalidateRender(sess, ac, true, "prompt.go")
 	}
 }
-func composePromptClientFrame(model *promptui.Model, base renderer.Frame, styles ...themeStyles) (renderer.Frame, []renderer.Damage) {
-	styleSet := resolveThemeStyles(styles)
-	return composeModalClientFrame(base, promptModalFor(model.Title()), styleSet, styleSet.accent, model.Render)
-}

M internal/usecase/daemon/render.go +230 -688

@@ -1,49 +1,85 @@
1	-	// Package daemon holds vev's server-side session multiplexer use case: the	1	+	// Package daemon holds vev's server-side session multiplexer use case.
2	-	// accept loop, the ephemeral/named session registry, the per-tab PTY reader
3	-	// and VT screen, and the per-client debounced render scheduler.
4	-	//
5	-	// Concurrency model (sessions own one or more PTY-backed tabs):
6	-	//
7	-	// - Serve runs the accept loop. Each accepted connection is handled by its
8	-	// own goroutine (handleConn): it reads the first frame and routes it to a
9	-	// session create/attach, a list, or a kill.
10	-	// - Per session there are exactly two long-lived goroutines: the PTY reader
11	-	// (drains child output into the VT screen and pokes a cap-1 dirty channel)
12	-	// and the render scheduler (debounces dirties and paints the attached
13	-	// client). Both are tied to the session context and unwind when the
14	-	// session is killed (pty.Close unblocks the reader; ctx cancel stops the
15	-	// scheduler).
16	-	// - The daemon exits (Serve returns) when the last session is removed, or
17	-	// when the parent context is cancelled (graceful shutdown notifies any
18	-	// attached clients with ReasonServerShutdown).
19	-	//
20	-	// Locking: a pane's screen/scrollback and per-client renderer shadow are
21	-	// guarded by pane.mu/tab.mu as appropriate; the attached-client pointer by
22	-	// session.mu; the registry by Daemon.mu. When more than one is held the order
23	-	// is always attachedClient.sendMu > Daemon.mu > session.mu > tab.mu > pane.mu.
24	-	// The PTY reader only ever takes pane.mu, so it never blocks on a slow client.
25		package daemon	2		package daemon
26			3
27		import (	4		import (
28		"bytes"	5		"bytes"
29	-	"errors"
30	-	"strconv"
31	-	"time"
32			6
33		"github.com/bnema/vev/internal/domain"	7		"github.com/bnema/vev/internal/domain"
34		"github.com/bnema/vev/internal/ports"	8		"github.com/bnema/vev/internal/ports"
35		"github.com/bnema/vev/internal/usecase/layout"	9		"github.com/bnema/vev/internal/usecase/layout"
36		themeui "github.com/bnema/vev/internal/usecase/theme"	10		themeui "github.com/bnema/vev/internal/usecase/theme"
37	-	"github.com/bnema/vev/internal/usecase/ui"
38		"github.com/bnema/vev/pkg/renderer"	11		"github.com/bnema/vev/pkg/renderer"
39	-	"github.com/bnema/vev/pkg/vt"
40		)	12		)
41			13
42	-	func signal(ch chan struct{}) {	14	+	// completedSynchronizedUpdate reports a full DEC 2026 batch contained in one
43	-	select {	15	+	// PTY read. The VT's final state is inactive in this case, so it cannot be
44	-	case ch <- struct{}{}:	16	+	// inferred from before/after SyncUpdateActive alone.
45	-	default:	17	+	func completedSynchronizedUpdate(data []byte) bool {
			18	+	_, tail, found := bytes.Cut(data, []byte(renderer.SyncStartCSI))
			19	+	return found && bytes.Contains(tail, []byte(renderer.SyncEndCSI))
			20	+	}
			21	+
			22	+	// paneRenderable reports dynamic composition visibility. Callers must not
			23	+	// hold coordinator locks: this acquires daemon/session/tab/pane-adjacent locks.
			24	+	func (d Daemon) paneRenderable(sess session, tb tab, p pane) bool {
			25	+	if sess == nil \|\| tb == nil \|\| p == nil {
			26	+	return false
			27	+	}
			28	+	sess.mu.Lock()
			29	+	active := sess.active >= 0 && sess.active < len(sess.tabs) && sess.tabs[sess.active] == tb
			30	+	attached := sess.client != nil
			31	+	sess.mu.Unlock()
			32	+
			33	+	// The normal attached render path needs no cross-session picker lookup.
			34	+	// Only inactive or headless tabs can be renderable as a picker preview.
			35	+	if (!active \|\| !attached) && !d.tabIsPickerPreview(tb) {
			36	+	return false
			37	+	}
			38	+	tb.mu.Lock()
			39	+	defer tb.mu.Unlock()
			40	+	if tb.floating.pane == p {
			41	+	return active && attached && tb.floating.state == floatingVisible
			42	+	}
			43	+	placements, ok := solvedPlacementsLocked(tb)
			44	+	if !ok {
			45	+	return false
			46	+	}
			47	+	for _, placement := range placements {
			48	+	if placement.ID == p.id && !placement.Collapsed && placement.Content.Width > 0 && placement.Content.Height > 0 {
			49	+	return true
			50	+	}
46		}	51		}
			52	+	return false
			53	+	}
			54	+
			55	+	// tabIsPickerPreview reports whether any attached client currently composes
			56	+	// tb as a picker preview. It snapshots daemon ownership before taking overlay
			57	+	// locks, preserving the Daemon -> session -> tab -> pane lock order.
			58	+	func (d Daemon) tabIsPickerPreview(tb tab) bool {
			59	+	if d == nil \|\| tb == nil {
			60	+	return false
			61	+	}
			62	+	d.mu.Lock()
			63	+	sessions := make([]*session, 0, len(d.sessions))
			64	+	for _, sess := range d.sessions {
			65	+	sessions = append(sessions, sess)
			66	+	}
			67	+	d.mu.Unlock()
			68	+	for _, sess := range sessions {
			69	+	sess.mu.Lock()
			70	+	ac := sess.client
			71	+	sess.mu.Unlock()
			72	+	if ac == nil \|\| ac.overlays == nil {
			73	+	continue
			74	+	}
			75	+	ac.overlays.pickerMu.Lock()
			76	+	preview := ac.overlays.pickerPreview == tb
			77	+	ac.overlays.pickerMu.Unlock()
			78	+	if preview {
			79	+	return true
			80	+	}
			81	+	}
			82	+	return false
47		}	83		}
48			84
49		func (d Daemon) ptyReader(sess session, tb tab, p pane) {	85		func (d Daemon) ptyReader(sess session, tb tab, p pane) {

@@ -52,61 +88,17 @@ func (d Daemon) ptyReader(sess session, tb tab, p pane) {
52		return	88		return
53		}	89		}
54		buf := make([]byte, ptyReadBufSize)	90		buf := make([]byte, ptyReadBufSize)
55	-	var resp []byte
56	-	var clipboards []string
57	-	attentionCh := make(chan struct{}, 1)
58		p.mu.Lock()	91		p.mu.Lock()
59	-	p.screen.OnResponse = func(b []byte) { resp = append(resp, b...) }	92	+	p.screen.OnResponse = func(b []byte) { p.ptyResponses = append(p.ptyResponses, b...) }
60	-	p.screen.OnBell = func() { signal(attentionCh) }	93	+	p.screen.OnBell = func() { p.ptyAttention = true }
61	-	p.screen.OnNotify = func(string, string) { signal(attentionCh) }	94	+	p.screen.OnNotify = func(string, string) { p.ptyAttention = true }
62	-	p.screen.OnProgress = func(bool) { signal(attentionCh) }	95	+	p.screen.OnProgress = func(bool) { p.ptyAttention = true }
63	-	p.screen.OnClipboard = func(b64 string) { clipboards = append(clipboards, b64) }	96	+	p.screen.OnClipboard = func(b64 string) { p.ptyClipboards = append(p.ptyClipboards, b64) }
64		p.mu.Unlock()	97		p.mu.Unlock()
65		for {	98		for {
66		n, err := p.pty.Read(buf)	99		n, err := p.pty.Read(buf)
67		if n > 0 {	100		if n > 0 {
68	-	data := buf[:n]	101	+	d.processPTYData(sess, tb, p, buf[:n], true)
69	-	p.mu.Lock()
70	-	wasSyncing := p.screen.SyncUpdateActive()
71	-	p.screen.Write(data)
72	-	p.refreshTerminalTitleLocked()
73	-	isSyncing := p.screen.SyncUpdateActive()
74	-	p.mu.Unlock()
75	-	markSnapshotDirty(sess)
76	-	select {
77	-	case <-attentionCh:
78	-	d.noteAttention(sess, tb)
79	-	default:
80	-	}
81	-	if len(clipboards) > 0 {
82	-	for _, b64 := range clipboards {
83	-	d.forwardClipboardAsync(sess, b64)
84	-	}
85	-	clipboards = clipboards[:0]
86	-	}
87	-	if len(resp) > 0 {
88	-	if _, writeErr := p.pty.Write(resp); writeErr != nil {
89	-	d.log.Warn("pty response write failed", "err", writeErr, "session", sess.name)
90	-	}
91	-	resp = resp[:0]
92	-	}
93	-	if wasSyncing != isSyncing {
94	-	p.mu.Lock()
95	-	p.syncGen++
96	-	gen := p.syncGen
97	-	p.mu.Unlock()
98	-	if isSyncing {
99	-	go d.syncWatchdog(p, gen)
100	-	}
101	-	}
102	-	if (wasSyncing && !isSyncing) \|\| (!isSyncing && syncUpdateEndIn(data)) {
103	-	signal(p.flush)
104	-	continue
105	-	}
106	-	if isSyncing {
107	-	continue
108	-	}
109	-	signal(p.dirty)
110		}	102		}
111		if err != nil {	103		if err != nil {
112		if p.onExit != nil {	104		if p.onExit != nil {

@@ -119,250 +111,182 @@ func (d Daemon) ptyReader(sess session, tb tab, p pane) {
119		}	111		}
120		}	112		}
121			113
122	-	// scheduler debounces dirty signals. The first dirty opens a short tab;	114	+	// processPTYData is the sole VT parsing path. While a resize apply owns the
123	-	// sustained floods progressively widen that tab, while isolated updates	115	+	// pane, reader calls append exact bytes and returns immediately; it never waits
124	-	// return to the minimum delay for interactive latency.	116	+	// for PTY.Resize or loses a short read.
125	-	func (d Daemon) scheduler(sess session, tb tab, p pane) {	117	+	func (d Daemon) processPTYData(sess session, tb tab, p pane, data []byte, bufferDuringApply bool) {
126	-	defer d.sessWg.Done()	118	+	rc := sess.renderCoordinator()
127	-	if p == nil {	119	+	p.mu.Lock()
			120	+	if bufferDuringApply && p.resizeApplying {
			121	+	p.resizePending = append(p.resizePending, data...)
			122	+	p.mu.Unlock()
128		return	123		return
129		}	124		}
130	-	delay := minDebounceInterval	125	+	wasSyncing := p.screen.SyncUpdateActive()
131	-	lastRender := d.clock.Now()	126	+	p.screen.Write(data)
132	-	outer:	127	+	p.refreshTerminalTitleLocked()
133	-	for {	128	+	isSyncing := p.screen.SyncUpdateActive()
134	-	select {	129	+	completeSyncRead := !wasSyncing && !isSyncing && completedSynchronizedUpdate(data)
135	-	case <-sess.ctx.Done():	130	+	var syncGen uint64
136	-	return	131	+	syncEnded := false
137	-	case <-paneDone(p):	132	+	if wasSyncing != isSyncing {
138	-	return	133	+	if isSyncing {
139	-	case <-p.flush:	134	+	syncGen = sess.syncGen.Add(1)
140	-	d.render(sess, tb, p)	135	+	p.syncGen = syncGen
141	-	lastRender = d.clock.Now()	136	+	if rc != nil {
142	-	continue	137	+	rc.noteSyncBeginWithRenderability(p, syncGen, func() bool { return d.paneRenderable(sess, tb, p) }, func() {
143	-	case <-p.dirty:	138	+	p.mu.Lock()
144	-	if d.clock.Now().Sub(lastRender) >= maxDebounceInterval {	139	+	if p.syncGen == syncGen && p.screen.SyncUpdateActive() {
145	-	delay = minDebounceInterval	140	+	p.screen.ForceSyncEnd()
			141	+	}
			142	+	p.mu.Unlock()
			143	+	})
146		}	144		}
			145	+	} else {
			146	+	syncGen = p.syncGen
			147	+	syncEnded = rc != nil && rc.removeSyncEnd(p, syncGen, true)
147		}	148		}
148	-		149	+	}
149	-	coalesced := 0	150	+	p.mu.Unlock()
150	-	timer := d.clock.NewTimer(delay)	151	+	renderable := d.paneRenderable(sess, tb, p)
151	-	absorb:	152	+	markSnapshotDirty(sess)
152	-	for {	153	+	d.flushPTYEffects(sess, tb, p)
153	-	select {	154	+	if rc != nil {
154	-	case <-sess.ctx.Done():	155	+	if wasSyncing != isSyncing {
155	-	timer.Stop()	156	+	if isSyncing {
156	-	return	157	+	if renderable {
157	-	case <-paneDone(p):	158	+	rc.invalidate(renderInvalidation{class: invalidateOutput, producer: "render.go"})
158	-	timer.Stop()
159	-	return
160	-	case <-p.flush:
161	-	if !timer.Stop() {
162	-	select {
163	-	case <-timer.C():
164	-	default:
165	-	}
166		}	159		}
167	-	d.render(sess, tb, p)	160	+	} else if syncEnded && renderable {
168	-	lastRender = d.clock.Now()	161	+	rc.invalidate(renderInvalidation{class: invalidateUrgent, producer: "render.go"})
169	-	continue outer
170	-	case <-p.dirty:
171	-	coalesced++
172	-	case <-timer.C():
173	-	break absorb
174		}	162		}
175		}	163		}
176	-	delay = nextDebounceDelay(delay, coalesced)	164	+	if completeSyncRead && renderable {
177	-	d.render(sess, tb, p)	165	+	rc.invalidate(renderInvalidation{class: invalidateUrgent, producer: "render.go"})
178	-	lastRender = d.clock.Now()	166	+	} else if renderable && !isSyncing && wasSyncing == isSyncing {
179	-	}	167	+	rc.invalidate(renderInvalidation{class: invalidateOutput, producer: "render.go"})
180	-	}	168	+	}
181	-
182	-	func nextDebounceDelay(delay time.Duration, coalesced int) time.Duration {
183	-	if coalesced == 0 {
184	-	return minDebounceInterval
185	-	}
186	-	if delay >= maxDebounceInterval {
187	-	return maxDebounceInterval
188	-	}
189	-	delay += debounceStep
190	-	if delay > maxDebounceInterval {
191	-	return maxDebounceInterval
192		}	169		}
193	-	return delay
194		}	170		}
195			171
196	-	func (d Daemon) syncWatchdog(p pane, gen uint64) {	172	+	// flushPTYEffects preserves the normal response/title/attention/clipboard
197	-	timer := d.clock.NewTimer(maxSyncUpdateDuration)	173	+	// behavior for both direct reads and resize-buffer replay.
198	-	select {	174	+	func (d Daemon) flushPTYEffects(sess session, tb tab, p pane) {
199	-	case <-paneDone(p):
200	-	timer.Stop()
201	-	return
202	-	case <-timer.C():
203	-	}
204	-
205		p.mu.Lock()	175		p.mu.Lock()
206	-	if p.syncGen != gen \|\| !p.screen.SyncUpdateActive() {	176	+	attention := p.ptyAttention
207	-	p.mu.Unlock()	177	+	p.ptyAttention = false
208	-	return	178	+	responses := append([]byte(nil), p.ptyResponses...)
209	-	}	179	+	p.ptyResponses = p.ptyResponses[:0]
210	-	p.screen.ForceSyncEnd()	180	+	clipboards := append([]string(nil), p.ptyClipboards...)
			181	+	p.ptyClipboards = p.ptyClipboards[:0]
211		p.mu.Unlock()	182		p.mu.Unlock()
212	-	signal(p.flush)	183	+	if attention {
213	-	}	184	+	d.noteAttention(sess, tb)
214	-
215	-	func syncUpdateEndIn(data []byte) bool {
216	-	return bytes.Contains(data, []byte("\x1b[?2026l"))
217	-	}
218	-
219	-	// render paints the current client, or (when detached) just clears accumulated
220	-	// damage so it never grows unbounded while headless.
221	-	func (d Daemon) render(sess session, tb tab, p pane) {
222	-	if p == nil {
223	-	return
224		}	185		}
225	-	p.mu.Lock()	186	+	for _, b64 := range clipboards {
226	-	if p.screen.SyncUpdateActive() {	187	+	d.forwardClipboardAsync(sess, b64)
227	-	p.mu.Unlock()
228	-	return
229		}	188		}
230	-	p.mu.Unlock()	189	+	if len(responses) > 0 {
231	-		190	+	if _, err := p.pty.Write(responses); err != nil {
232	-	// A retained floating pane continues to drain its PTY while hidden. Its	191	+	d.log.Warn("pty response write failed", "err", err, "session", sess.name)
233	-	// scheduler must not route through picker preview (or the active client),
234	-	// but it must consume damage so a long-hidden command cannot retain it.
235	-	tb.mu.Lock()
236	-	hiddenFloating := tb.floating.pane == p && tb.floating.state != floatingVisible
237	-	previewer := tb.previewClient
238	-	tb.mu.Unlock()
239	-	if hiddenFloating {
240	-	p.mu.Lock()
241	-	p.screen.ClearDamage()
242	-	p.mu.Unlock()
243	-	return
244	-	}
245	-
246	-	sess.mu.Lock()
247	-	ac := sess.client
248	-	active := sess.active >= 0 && sess.active < len(sess.tabs) && sess.tabs[sess.active] == tb
249	-	sess.mu.Unlock()
250	-
251	-	if previewer != nil {
252	-	if previewSess := previewer.currentSession(); previewSess != nil {
253	-	d.paint(previewSess, previewer, false)
254		}	192		}
255	-	if ac != nil && active && ac != previewer {
256	-	d.paint(sess, ac, false)
257	-	return
258	-	}
259	-	p.mu.Lock()
260	-	p.screen.ClearDamage()
261	-	p.mu.Unlock()
262	-	return
263	-	}
264	-
265	-	if ac == nil \|\| !active {
266	-	p.mu.Lock()
267	-	p.screen.ClearDamage()
268	-	p.mu.Unlock()
269	-	return
270		}	193		}
271	-	d.refreshBarScriptsIfDue(sess, d.clock.Now(), false)
272	-	d.paint(sess, ac, false)
273		}	194		}
274			195
275		// paint draws the composed client frame (active tab plus status bar) and	196		// paint draws the composed client frame (active tab plus status bar) and
276		// sends the resulting bytes. The renderer shadow is reset on explicit invalidations	197		// sends the resulting bytes. The renderer shadow is reset on explicit invalidations
277		// such as switch/create/close/rename/resize so the repaint is complete.	198		// such as switch/create/close/rename/resize so the repaint is complete.
278	-	// copyTargetRectLocked maps a captured copy source into the already-composed	199	+	func (d *Daemon) observeRuntime(kind ports.RuntimeMarkKind, bytes uint64, valid bool) {
279	-	// client frame. The caller holds tb.mu, preserving the layout/floating	200	+	if d != nil && d.runtimeObserver != nil {
280	-	// snapshot while the rectangle is chosen.	201	+	d.runtimeObserver.ObserveRuntime(ports.NewRuntimeMark("daemon", kind, bytes, valid))
281	-	func copyTargetRectLocked(layoutSnap tabLayoutSnapshot, contentArea domain.Rect, p, floating *pane, hasFloating bool, floatingFrameGeometry floatingGeometry) domain.Rect {
282	-	if p == nil {
283	-	return domain.Rect{}
284	-	}
285	-	if hasFloating && p == floating {
286	-	return floatingFrameGeometry.Inner
287	-	}
288	-	for _, placement := range layoutSnap.placements {
289	-	if placement.ID == p.id && !placement.Collapsed {
290	-	r := placement.Content
291	-	r.X += contentArea.X
292	-	r.Y += contentArea.Y
293	-	return r
294	-	}
295		}	202		}
296	-	p.mu.Lock()
297	-	width, height := p.screen.Frame.Width, p.screen.Frame.Height
298	-	p.mu.Unlock()
299	-	return domain.Rect{X: contentArea.X, Y: contentArea.Y, Width: min(width, contentArea.Width), Height: min(height, contentArea.Height)}
300		}	203		}
301			204
302	-	func titleBarPaneIDs(placements []layout.Placement, ok bool) []layout.PaneID {	205	+	// runtimeMarkBatch captures immutable runtime marks during a render transaction.
303	-	if !ok {	206	+	// paint owns attachment sendMu while it captures, composes, and emits; JSONL
304	-	return nil	207	+	// observer I/O must therefore happen only after that ownership is released.
305	-	}	208	+	type runtimeMarkBatch struct {
306	-	ids := make([]layout.PaneID, 0, len(placements))	209	+	observer ports.RuntimeObserver
307	-	for _, pl := range placements {	210	+	marks []ports.RuntimeMark
308	-	if pl.TitleBar.Height > 0 {	211	+	}
309	-	ids = append(ids, pl.ID)	212	+
310	-	}	213	+	func (d *Daemon) newRuntimeMarkBatch() runtimeMarkBatch {
			214	+	if d == nil {
			215	+	return runtimeMarkBatch{}
311		}	216		}
312	-	return ids	217	+	return runtimeMarkBatch{observer: d.runtimeObserver}
313		}	218		}
314			219
315	-	func (d Daemon) scheduleResizePaintLocked(sess session, ac *attachedClient) {	220	+	func (b *runtimeMarkBatch) span(start, end ports.RuntimeMarkKind, bytes uint64) func(uint64, bool) {
316	-	ac.resizePaint.stop()	221	+	if b == nil \|\| b.observer == nil {
317	-	ac.resizePaintGeneration++	222	+	return func(uint64, bool) {}
318	-	generation := ac.resizePaintGeneration	223	+	}
319	-	ac.resizePaintPending = true	224	+	correlation := ports.NewRuntimeCorrelation()
320	-	ac.resizePaint.retain(d.clock, maxDebounceInterval, func(ports.Timer) {	225	+	b.marks = append(b.marks, ports.NewRuntimeMarkWithCorrelation("daemon", correlation, start, bytes, true))
321	-	d.paintForResizeGeneration(sess, ac, true, generation)	226	+	return func(endBytes uint64, valid bool) {
322	-	})	227	+	b.marks = append(b.marks, ports.NewRuntimeMarkWithCorrelation("daemon", correlation, end, endBytes, valid))
			228	+	}
323		}	229		}
324			230
325	-	func (d Daemon) paint(sess session, ac *attachedClient, reset bool) {	231	+	func (b *runtimeMarkBatch) flush() {
326	-	d.paintForResizeGeneration(sess, ac, reset, 0)	232	+	if b == nil \|\| b.observer == nil {
			233	+	return
			234	+	}
			235	+	for _, mark := range b.marks {
			236	+	b.observer.ObserveRuntime(mark)
			237	+	}
327		}	238		}
328			239
329	-	func (d Daemon) paintForResizeGeneration(sess session, ac *attachedClient, reset bool, resizeGeneration uint64) {	240	+	func (d Daemon) paint(sess session, ac attachedClient, reset bool, lease attachmentLease) {
			241	+	marks := d.newRuntimeMarkBatch()
330		tb := sess.activeTab()	242		tb := sess.activeTab()
331		if tb == nil {	243		if tb == nil {
332		return	244		return
333		}	245		}
334			246
335	-	ac.initOverlays()
336		ac.sendMu.Lock()	247		ac.sendMu.Lock()
337	-	if ac.currentSession() != sess {	248	+	// sendMu is the attachment ownership boundary. Check the session's
338	-	ac.sendMu.Unlock()	249	+	// published identity while holding it so a deadline captured before an
339	-	return	250	+	// attach/replace cannot emit on either the old or new output chain.
340	-	}	251	+	sess.mu.Lock()
341	-	if resizeGeneration != 0 && (!ac.resizePaintPending \|\| ac.resizePaintGeneration != resizeGeneration) {	252	+	owned := sess.client == ac
			253	+	sess.mu.Unlock()
			254	+	if !owned \|\| ac.currentSession() != sess {
342		ac.sendMu.Unlock()	255		ac.sendMu.Unlock()
343		return	256		return
344		}	257		}
345	-	if ac.resizePaintPending {	258	+	if lease != nil {
346	-	ac.resizePaint.stop()	259	+	rc := sess.renderCoordinator()
347	-	ac.resizePaintPending = false	260	+	if rc == nil \|\| lease.attachment != ac \|\| !rc.leaseCurrent(lease, true) {
348	-	reset = true	261	+	ac.sendMu.Unlock()
			262	+	return
			263	+	}
349		}	264		}
350	-	// Ack-gated coalescing: while the client is at least maxUnackedOutputStates	265	+	// Capacity is checked before any destructive capture. The coordinator is
351	-	// behind, skip composing another diff and mark the paint deferred. The	266	+	// the normal gate, but this ownership check also protects direct resize and
352	-	// MsgOutputAck handler flushes one cumulative paint once acks catch up. A	267	+	// test paint paths from consuming damage that cannot be emitted.
353	-	// deferred reset stays sticky so the cumulative paint is dependency-free.	268	+	if ac.output != nil && ac.output.atCapacity() {
354	-	if ac.output.deferIfAtCapacity(reset) {	269	+	// The coordinator owns the blocked interval; this guard only protects
			270	+	// direct test/resize paint calls from destructive capture.
355		ac.sendMu.Unlock()	271		ac.sendMu.Unlock()
356		return	272		return
357		}	273		}
			274	+	// Composition owns attachment sendMu; initialize its lazy overlay state
			275	+	// under that same ownership so concurrent fallback paints cannot observe a
			276	+	// partially published runtime.
			277	+	ac.initOverlays()
358		overlays := ac.overlays.SnapshotForRender()	278		overlays := ac.overlays.SnapshotForRender()
359		repaintAttachedClients := false	279		repaintAttachedClients := false
360		defer func() {	280		defer func() {
			281	+	overlays.Unlock()
			282	+	// Every return below has released sendMu before reaching this defer.
			283	+	// Emit the captured sequence before any follow-up repaint can introduce
			284	+	// another boundary, while no attachment, session, tab, or pane lock is held.
			285	+	marks.flush()
361		if repaintAttachedClients {	286		if repaintAttachedClients {
362		d.repaintAllAttachedClients()	287		d.repaintAllAttachedClients()
363		}	288		}
364		}()	289		}()
365	-	defer overlays.Unlock()
366		preview := snapshotPickerPreview(nil)	290		preview := snapshotPickerPreview(nil)
367		if overlays.previewTab != tb {	291		if overlays.previewTab != tb {
368		preview = snapshotPickerPreview(overlays.previewTab)	292		preview = snapshotPickerPreview(overlays.previewTab)

@@ -375,11 +299,16 @@ func (d Daemon) paintForResizeGeneration(sess session, ac *attachedClient, res
375		attentionVisible := pulseVisible(bars.attentionFrame)	299		attentionVisible := pulseVisible(bars.attentionFrame)
376		repaintAttachedClients = sess.ackAttention(tb, attentionVisible)	300		repaintAttachedClients = sess.ackAttention(tb, attentionVisible)
377			301
378	-	styles := newThemeStyles(ac.getTheme())
379		paletteCfg := d.currentPaletteConfig()	302		paletteCfg := d.currentPaletteConfig()
			303	+	floatingCfg := d.currentFloatingConfig()
			304	+	// Title refresh may inspect process state, so it remains before the capture
			305	+	// boundary and outside tab/pane ownership locks.
380		tb.mu.Lock()	306		tb.mu.Lock()
381	-	layoutSnap := solveTabLayoutLocked(tb)	307	+	titleIDs := ac.renderScratch.titleIDs[:0]
382	-	titleIDs := titleBarPaneIDs(layoutSnap.placements, layoutSnap.ok)	308	+	if tb.tree != nil {
			309	+	titleIDs = appendStackPaneIDs(titleIDs, tb.tree.Root)
			310	+	}
			311	+	ac.renderScratch.titleIDs = titleIDs
383		floating := tb.floating.pane	312		floating := tb.floating.pane
384		hasFloating := tb.floating.state == floatingVisible && floating != nil	313		hasFloating := tb.floating.state == floatingVisible && floating != nil
385		tb.mu.Unlock()	314		tb.mu.Unlock()

@@ -390,189 +319,35 @@ func (d Daemon) paintForResizeGeneration(sess session, ac *attachedClient, res
390		d.refreshPaneDisplayTitle(sess, floating, false)	319		d.refreshPaneDisplayTitle(sess, floating, false)
391		}	320		}
392			321
393	-	tb.mu.Lock()	322	+	capturedOverlays := capturedOverlayRenderState{
394	-	if !layoutSnap.matchesLocked(tb) {	323	+	copyActive: overlays.copyActive, copySearchActive: overlays.copySearchModel != nil,
395	-	layoutSnap = solveTabLayoutLocked(tb)	324	+	pickerActive: overlays.pickerActive, paletteActive: overlays.paletteActive, promptActive: overlays.promptActive,
			325	+	copyFeedback: overlays.copyFeedback,
396		}	326		}
397	-	p := tb.focusedPane()	327	+	endCapture := marks.span(ports.RuntimeCaptureStart, ports.RuntimeCaptureEnd, 0)
398	-	if p == nil {	328	+	state, ok := capturePrimaryRenderState(sess, ac, bars, capturedOverlays, preview, floatingCfg, reset, lease)
399	-	tb.mu.Unlock()	329	+	endCapture(0, ok)
400	-	ac.sendMu.Unlock()	330	+	if !ok {
401	-	return
402	-	}
403	-	// Recheck the slot after refreshing its title outside the tab lock. One
404	-	// paint uses one immutable floating config snapshot across composition,
405	-	// copy targeting, and cursor placement.
406	-	floating = tb.floating.pane
407	-	hasFloating = tb.floating.state == floatingVisible && floating != nil
408	-	floatingCfg := d.currentFloatingConfig()
409	-	overlayActive := overlays.copyActive \|\| overlays.copySearchModel != nil \|\| overlays.pickerActive \|\| overlays.paletteActive \|\| overlays.promptActive
410	-	if reset \|\| overlayActive {
411	-	ac.bars.Reset()
412	-	ac.composed.invalidate()
413	-	}
414	-	if reset \|\| overlays.pickerActive \|\| overlays.paletteActive \|\| overlays.promptActive {
415	-	ac.lastCursor.valid = false
416	-	}
417	-	var frame renderer.Frame
418	-	var damage []renderer.Damage
419	-	if overlayActive {
420	-	frame, damage = composeClientFrameWithLayoutCachedConsumeDamage(bars, tb, reset, layoutSnap, &ac.bars, nil)
421	-	} else {
422	-	frame, damage = composeClientFrameWithLayoutCachedConsumeDamage(bars, tb, reset, layoutSnap, &ac.bars, &ac.composed)
423	-	}
424	-	contentArea := domain.Rect{Y: 1, Width: frame.Width, Height: max(0, frame.Height-2)}
425	-	floatingFrameGeometry := floatingGeometry{}
426	-	if hasFloating {
427	-	desiredFloatingGeometry := calculateContentFloatingGeometry(domain.Size{Cols: contentArea.Width, Rows: contentArea.Height}, floatingCfg)
428	-	frame, damage, floatingFrameGeometry = composeFloatingFrame(frame, damage, floating, tb.floating.generation, contentArea, desiredFloatingGeometry, layoutSnap, bars.theme, &ac.composed, reset \|\| overlayActive)
429	-	}
430	-	if overlays.copyActive {
431	-	copyPane := overlays.copyPane
432	-	if copyPane == nil {
433	-	copyPane = p
434	-	}
435	-	copyTarget := copyTargetRectLocked(layoutSnap, contentArea, copyPane, floating, hasFloating, floatingFrameGeometry)
436	-	frame, damage = composeCopyClientFrame(overlays.copyMode, overlays.copySnapshot, copyTarget, frame, bars)
437	-	}
438	-	// A palette above normal/copy content dims that composed content. When a
439	-	// floating pane is present its own backdrop already dims normal pane
440	-	// contents; applying the palette backdrop here would also dim the popup.
441	-	if overlays.paletteActive && !hasFloating {
442	-	(overlayBackdrop{DimPaneContents: true}).apply(frame, contentArea, layoutSnap, bars.theme)
443	-	}
444	-	if overlays.copySearchModel != nil {
445	-	frame, damage = composeCopySearchClientFrame(overlays.copySearchModel, frame, styles)
446	-	}
447	-	if overlays.pickerActive {
448	-	if overlays.previewTab == tb {
449	-	if layoutSnap.ok && tb.tree != nil && tb.tree.Root != nil && tb.tree.Root.Kind != layout.Leaf {
450	-	previewFrame, _ := composeTabFrameWithLayout(tb, layoutSnap.area, themeui.Theme{}, layoutSnap)
451	-	preview = pickerPreviewFromFrame(previewFrame)
452	-	} else {
453	-	preview = pickerPreviewFromLockedTab(tb)
454	-	}
455	-	}
456	-	frame, damage = composePickerClientFrame(overlays.pickerModel, preview, frame, styles)
457	-	}
458	-	if overlays.paletteActive {
459	-	guidance := ""
460	-	if overlays.paletteHints != nil {
461	-	guidance = overlays.paletteHints.Feedback
462	-	}
463	-	frame, damage = composePaletteClientFrame(overlays.paletteModel, frame, paletteCfg, guidance, styles)
464	-	}
465	-	if overlays.promptActive {
466	-	frame, damage = composePromptClientFrame(overlays.promptModel, frame, styles)
467	-	}
468	-	overlays.Unlock()
469	-	cursorPane := p
470	-	cursorContent, cursorVisible := focusedPaneContentRect(layoutSnap, p.id)
471	-	cursorContent.X += contentArea.X
472	-	cursorContent.Y += contentArea.Y
473	-	if hasFloating {
474	-	cursorPane = floating
475	-	cursorContent, cursorVisible = floatingFrameGeometry.Inner, floatingFrameGeometry.Inner.Width > 0 && floatingFrameGeometry.Inner.Height > 0
476	-	}
477	-	cursorPane.mu.Lock()
478	-	desiredCursor := desiredCursorOut(cursorPane.screen, cursorContent, !cursorVisible \|\| overlays.copyActive \|\| overlays.copySearchModel != nil \|\| overlays.pickerActive \|\| overlays.paletteActive \|\| overlays.promptActive)
479	-	cursorPane.mu.Unlock()
480	-	ac.sess.mu.Lock()
481	-	if ac.sess.v != sess {
482	-	ac.sess.mu.Unlock()
483	-	tb.mu.Unlock()
484		ac.sendMu.Unlock()	331		ac.sendMu.Unlock()
485		return	332		return
486		}	333		}
487	-	prepared, err := ac.output.prepare(frame, damage, reset \|\| overlayActive)	334	+	if ac.renderStages.capture != nil {
488	-	var data []byte	335	+	ac.renderStages.capture()
489	-	if err == nil {
490	-	data = prepared.data
491	-	}
492	-	var cursorTail []byte
493	-	if err == nil {
494	-	cursorTail = ac.encodeCursorTail(desiredCursor, len(data) > 0)
495	-	}
496	-	tb.mu.Unlock()
497	-
498	-	var serr error
499	-	var sendTr ports.Transport
500	-	if err == nil {
501	-	data = append(data, cursorTail...)
502	-	if len(data) > 0 {
503	-	sendTr = ac.transport()
504	-	if sendTr == nil {
505	-	serr = errors.New("client transport is nil")
506	-	} else {
507	-	send := sendTr.Send
508	-	if async, ok := sendTr.(ports.AsyncTransport); ok {
509	-	send = async.SendAsync
510	-	}
511	-	serr = prepared.send(data, ac.echoAck.Load(), send)
512	-	}
513	-	}
514	-	}
515	-	ac.sess.mu.Unlock()
516	-	ac.sendMu.Unlock()
517	-
518	-	if err != nil {
519	-	d.log.Error("render draw failed", "err", err, "session", sess.name)
520	-	return
521	-	}
522	-	if serr != nil {
523	-	d.detachOnSendError(sess, ac, sendTr)
524		}	336		}
525	-	}	337	+	if overlays.pickerActive && overlays.previewTab == tb {
526	-		338	+	previewState := *state
527	-	func focusedPaneContentRect(layoutSnap tabLayoutSnapshot, id layout.PaneID) (domain.Rect, bool) {	339	+	previewState.overlays = capturedOverlayRenderState{}
528	-	if !layoutSnap.ok {	340	+	previewState.reset = true
529	-	return layoutSnap.area, layoutSnap.area.Width > 0 && layoutSnap.area.Height > 0	341	+	state.preview = pickerPreviewFromCapturedRender(previewState)
530		}	342		}
531	-	for _, pl := range layoutSnap.placements {	343	+	captureOverlayLayers(state, overlays, paletteCfg)
532	-	if pl.ID == id && !pl.Collapsed && pl.Content.Width > 0 && pl.Content.Height > 0 {	344	+	endCompose := marks.span(ports.RuntimeComposeStart, ports.RuntimeComposeEnd, 0)
533	-	return pl.Content, true	345	+	composed := composeFrame(*state, ac.pipelineCache, ac.pipelineScratch)
534	-	}	346	+	endCompose(0, true)
			347	+	if ac.renderStages.compose != nil {
			348	+	ac.renderStages.compose()
535		}	349		}
536	-	return domain.Rect{}, false	350	+	d.emitFrame(sess, ac, state, composed, &marks)
537	-	}
538	-
539	-	// desiredCursorOut computes the terminal cursor state that should be shown to
540	-	// the client for the current pane placement and overlay mode.
541	-	func desiredCursorOut(s *vt.Screen, content domain.Rect, hide bool) cursorOut {
542	-	if hide \|\| !s.CursorVisible() {
543	-	return cursorOut{hidden: true}
544	-	}
545	-	style, ok := s.CursorStyle()
546	-	if !ok {
547	-	style = 1
548	-	}
549	-	return cursorOut{row: content.Y + s.CursorRow(), col: content.X + s.CursorCol(), style: style, hasStyle: true}
550	-	}
551	-
552	-	func (ac *attachedClient) encodeCursorTail(desired cursorOut, force bool) []byte {
553	-	changed := force \|\| !ac.lastCursor.valid \|\| ac.lastCursor.hidden != desired.hidden \|\| ac.lastCursor.row != desired.row \|\| ac.lastCursor.col != desired.col \|\| ac.lastCursor.style != desired.style \|\| ac.lastCursor.hasStyle != desired.hasStyle
554	-	if !changed {
555	-	return nil
556	-	}
557	-	prev := ac.lastCursor
558	-	ac.lastCursor = desired
559	-	ac.lastCursor.valid = true
560	-	if desired.hidden {
561	-	return []byte("\x1b[?25l")
562	-	}
563	-	var b []byte
564	-	b = append(b, "\x1b["...)
565	-	b = strconv.AppendInt(b, int64(desired.row+1), 10)
566	-	b = append(b, ';')
567	-	b = strconv.AppendInt(b, int64(desired.col+1), 10)
568	-	b = append(b, 'H')
569	-	if !prev.valid \|\| prev.hidden \|\| prev.style != desired.style \|\| prev.hasStyle != desired.hasStyle {
570	-	b = append(b, "\x1b["...)
571	-	b = strconv.AppendInt(b, int64(desired.style), 10)
572	-	b = append(b, " q"...)
573	-	}
574	-	b = append(b, "\x1b[?25h"...)
575	-	return b
576		}	351		}
577			352
578		type themeStyles struct {	353		type themeStyles struct {

@@ -638,211 +413,6 @@ func (c *barCache) Reset() {
638		c.bottom = nil	413		c.bottom = nil
639		}	414		}
640			415
641	-	type composedFrameCache struct {
642	-	frame renderer.Frame
643	-	valid bool
644	-	layoutSnap tabLayoutSnapshot
645	-	titleGenerations map[layout.PaneID]uint64
646	-	floating *pane
647	-	floatingFrame renderer.Frame
648	-	floatingGeneration uint64
649	-	floatingGeometry floatingGeometry
650	-	floatingTitleGeneration uint64
651	-	}
652	-
653	-	func (c *composedFrameCache) invalidate() {
654	-	c.valid = false
655	-	c.floating = nil
656	-	c.floatingGeneration = 0
657	-	c.floatingGeometry = floatingGeometry{}
658	-	c.floatingTitleGeneration = 0
659	-	}
660	-
661	-	func composeClientFrame(sess session, tb tab, full bool, rightStatus string, caches ...*barCache) (renderer.Frame, []renderer.Damage) {
662	-	return composeClientFrameWithState(barState{status: sess.statusSegments(true), copyFeedback: rightStatus}, tb, full, caches...)
663	-	}
664	-
665	-	func composeClientFrameWithState(bars barState, tb tab, full bool, caches ...barCache) (renderer.Frame, []renderer.Damage) {
666	-	return composeClientFrameWithLayout(bars, tb, full, solveTabLayoutLocked(tb), caches...)
667	-	}
668	-
669	-	func composeClientFrameWithLayout(bars barState, tb tab, full bool, layoutSnap tabLayoutSnapshot, caches ...barCache) (renderer.Frame, []renderer.Damage) {
670	-	var cache *barCache
671	-	if len(caches) > 0 {
672	-	cache = caches[0]
673	-	}
674	-	return composeClientFrameWithLayoutCached(bars, tb, full, layoutSnap, cache, nil)
675	-	}
676	-
677	-	func composeClientFrameWithLayoutCached(bars barState, tb tab, full bool, layoutSnap tabLayoutSnapshot, cache barCache, composed *composedFrameCache) (renderer.Frame, []renderer.Damage) {
678	-	return composeClientFrameWithLayoutCachedOptions(bars, tb, full, layoutSnap, cache, composed, false)
679	-	}
680	-
681	-	func composeClientFrameWithLayoutCachedConsumeDamage(bars barState, tb tab, full bool, layoutSnap tabLayoutSnapshot, cache barCache, composed *composedFrameCache) (renderer.Frame, []renderer.Damage) {
682	-	return composeClientFrameWithLayoutCachedOptions(bars, tb, full, layoutSnap, cache, composed, true)
683	-	}
684	-
685	-	func composeClientFrameWithLayoutCachedOptions(bars barState, tb tab, full bool, layoutSnap tabLayoutSnapshot, cache barCache, composed *composedFrameCache, consumeDamage bool) (renderer.Frame, []renderer.Damage) {
686	-	p := tb.focusedPane()
687	-	styles := newThemeStyles(bars.theme)
688	-	if p == nil {
689	-	width, screenRows := tb.size.Cols, tb.size.Rows
690	-	if width <= 0 \|\| screenRows <= 0 {
691	-	return renderer.NewFrame(0, 0), nil
692	-	}
693	-	return renderer.NewFrame(width, screenRows+2), nil
694	-	}
695	-	p.mu.Lock()
696	-	width, screenRows := p.screen.Frame.Width, p.screen.Frame.Height
697	-	p.mu.Unlock()
698	-	if tb.size.Valid() {
699	-	width, screenRows = tb.size.Cols, tb.size.Rows
700	-	}
701	-	cacheValid := composed != nil && composed.valid && composed.frame.Width == width && composed.frame.Height == screenRows+2
702	-	layoutSame := composed == nil \|\| (cacheValid && sameTabLayoutSnapshot(composed.layoutSnap, layoutSnap))
703	-	var frame renderer.Frame
704	-	if cacheValid {
705	-	frame = composed.frame
706	-	} else {
707	-	frame = renderer.NewFrame(width, screenRows+2)
708	-	if composed != nil {
709	-	full = true
710	-	}
711	-	}
712	-	if !layoutSame {
713	-	full = true
714	-	}
715	-	contentArea := domain.Rect{Y: 1, Width: width, Height: screenRows}
716	-	if cacheValid && !layoutSame {
717	-	clearFrameRect(frame, contentArea)
718	-	}
719	-	topBar := frame.Row(0)
720	-	drawTopBarSnapshot(topBar, bars.status, bars.attentionFrame, bars.topRight, styles)
721	-	var titleGenerations map[layout.PaneID]uint64
722	-	if composed != nil {
723	-	if composed.titleGenerations == nil {
724	-	composed.titleGenerations = make(map[layout.PaneID]uint64)
725	-	}
726	-	titleGenerations = composed.titleGenerations
727	-	}
728	-	contentDamage := composeTabFrameIntoWithLayoutOptions(tb, frame, contentArea, bars.theme, layoutSnap, cacheValid && layoutSame, consumeDamage, titleGenerations)
729	-	bottomBar := frame.Row(screenRows + 1)
730	-	drawStatusBarState(bottomBar, bars, styles)
731	-	if composed != nil {
732	-	composed.frame = frame
733	-	composed.valid = true
734	-	composed.layoutSnap = layoutSnap
735	-	}
736	-	if full {
737	-	if cache != nil {
738	-	cache.capture(topBar, bottomBar)
739	-	}
740	-	return frame, []renderer.Damage{renderer.FullRedraw()}
741	-	}
742	-	damage := translateDamage(contentDamage, 0, 1)
743	-	if cache == nil \|\| !sameCells(cache.top, topBar) {
744	-	damage = append(damage, renderer.Damage{Kind: renderer.DamageText, X: 0, Y: 0, Width: width, Height: 1})
745	-	}
746	-	if cache == nil \|\| !sameCells(cache.bottom, bottomBar) {
747	-	damage = append(damage, renderer.Damage{Kind: renderer.DamageText, X: 0, Y: screenRows + 1, Width: width, Height: 1})
748	-	}
749	-	if cache != nil {
750	-	cache.capture(topBar, bottomBar)
751	-	}
752	-	return frame, damage
753	-	}
754	-
755	-	func composeTabFrame(tb *tab, area domain.Rect, theme themeui.Theme) (renderer.Frame, []renderer.Damage) {
756	-	return composeTabFrameWithLayout(tb, area, theme, tabLayoutSnapshot{})
757	-	}
758	-
759	-	func composeTabFrameWithLayout(tb *tab, area domain.Rect, theme themeui.Theme, layoutSnap tabLayoutSnapshot) (renderer.Frame, []renderer.Damage) {
760	-	frame := renderer.NewFrame(area.Width, area.Height)
761	-	damage := composeTabFrameIntoWithLayout(tb, frame, area, theme, layoutSnap, false)
762	-	return frame, damage
763	-	}
764	-
765	-	func composeTabFrameIntoWithLayout(tb *tab, frame renderer.Frame, area domain.Rect, theme themeui.Theme, layoutSnap tabLayoutSnapshot, cacheValid bool) []renderer.Damage {
766	-	return composeTabFrameIntoWithLayoutOptions(tb, frame, area, theme, layoutSnap, cacheValid, false, nil)
767	-	}
768	-
769	-	func composeTabFrameIntoWithLayoutOptions(tb *tab, frame renderer.Frame, area domain.Rect, theme themeui.Theme, layoutSnap tabLayoutSnapshot, cacheValid bool, consumeDamage bool, titleGenerations map[layout.PaneID]uint64) []renderer.Damage {
770	-	contentArea := domain.Rect{Width: area.Width, Height: area.Height}
771	-	root := tb.tree.Root
772	-	placements, ok := layoutSnap.placements, layoutSnap.ok && layoutSnap.root == root && layoutSnap.area == contentArea
773	-	if !ok {
774	-	placements, ok = layout.Solve(root, contentArea)
775	-	}
776	-	var fallback *pane
777	-	if !ok {
778	-	fallback = tb.focusedPane()
779	-	if fallback == nil {
780	-	return nil
781	-	}
782	-	placements = []layout.Placement{{ID: fallback.id, Content: contentArea}}
783	-	}
784	-	// A valid cache has the same layout, so its title IDs remain valid. Reset
785	-	// the existing cache only when rebuilding after layout or frame churn.
786	-	if titleGenerations != nil && !cacheValid {
787	-	clear(titleGenerations)
788	-	}
789	-	if ok && !cacheValid {
790	-	drawDividers(frame, root, area, themeui.DimStyle(newThemeStyles(theme).border, theme))
791	-	}
792	-	var damage []renderer.Damage
793	-	for _, pl := range placements {
794	-	p := tb.panes[pl.ID]
795	-	if p == nil && fallback != nil && pl.ID == fallback.id {
796	-	p = fallback
797	-	}
798	-	if p == nil {
799	-	continue
800	-	}
801	-	focused := tb.tree.Focus == pl.ID
802	-	pl = offsetPlacement(pl, area.X, area.Y)
803	-	if pl.TitleBar.Height > 0 {
804	-	generation := drawPaneTitleBar(frame, pl, p, focused, theme)
805	-	if cacheValid && (titleGenerations == nil \|\| titleGenerations[pl.ID] != generation) {
806	-	titleDamage := pl.TitleBar
807	-	titleDamage.X -= area.X
808	-	titleDamage.Y -= area.Y
809	-	damage = append(damage, renderer.Damage{Kind: renderer.DamageText, X: titleDamage.X, Y: titleDamage.Y, Width: titleDamage.Width, Height: titleDamage.Height})
810	-	}
811	-	if titleGenerations != nil {
812	-	titleGenerations[pl.ID] = generation
813	-	}
814	-	}
815	-	if pl.Collapsed \|\| pl.Content.Width <= 0 \|\| pl.Content.Height <= 0 {
816	-	if consumeDamage {
817	-	p.mu.Lock()
818	-	p.screen.ClearDamage()
819	-	p.mu.Unlock()
820	-	}
821	-	continue
822	-	}
823	-	p.mu.Lock()
824	-	paneDamage := p.screen.Damage()
825	-	if !cacheValid \|\| len(paneDamage) > 0 {
826	-	blitPaneFrame(frame, pl.Content, p.screen.Frame, !focused, theme)
827	-	}
828	-	for _, d := range paneDamage {
829	-	localContent := pl.Content
830	-	localContent.X -= area.X
831	-	localContent.Y -= area.Y
832	-	damage = append(damage, translatePaneDamage(d, localContent, contentArea)...)
833	-	}
834	-	if consumeDamage {
835	-	p.screen.ClearDamage()
836	-	}
837	-	p.mu.Unlock()
838	-	}
839	-	return damage
840	-	}
841	-
842	-	func sameTabLayoutSnapshot(a, b tabLayoutSnapshot) bool {
843	-	return a.ok == b.ok && a.root == b.root && a.fingerprint == b.fingerprint && a.area == b.area && a.focus == b.focus
844	-	}
845	-
846		func offsetPlacement(pl layout.Placement, dx, dy int) layout.Placement {	416		func offsetPlacement(pl layout.Placement, dx, dy int) layout.Placement {
847		pl.Content.X += dx	417		pl.Content.X += dx
848		pl.Content.Y += dy	418		pl.Content.Y += dy

@@ -851,15 +421,6 @@ func offsetPlacement(pl layout.Placement, dx, dy int) layout.Placement {
851		return pl	421		return pl
852		}	422		}
853			423
854	-	func clearFrameRect(frame renderer.Frame, r domain.Rect) {
855	-	blank := renderer.BlankCell()
856	-	for y := r.Y; y < r.Y+r.Height && y < frame.Height; y++ {
857	-	for x := r.X; x < r.X+r.Width && x < frame.Width; x++ {
858	-	frame.Set(x, y, blank)
859	-	}
860	-	}
861	-	}
862	-
863		func blitPaneFrame(dst renderer.Frame, r domain.Rect, src renderer.Frame, dim bool, theme themeui.Theme) {	424		func blitPaneFrame(dst renderer.Frame, r domain.Rect, src renderer.Frame, dim bool, theme themeui.Theme) {
864		rows := min(r.Height, src.Height)	425		rows := min(r.Height, src.Height)
865		cols := min(r.Width, src.Width)	426		cols := min(r.Width, src.Width)

@@ -874,25 +435,6 @@ func blitPaneFrame(dst renderer.Frame, r domain.Rect, src renderer.Frame, dim bo
874		}	435		}
875		}	436		}
876			437
877	-	func drawPaneTitleBar(frame renderer.Frame, pl layout.Placement, p *pane, focused bool, theme themeui.Theme) uint64 {
878	-	styles := newThemeStyles(theme)
879	-	style := styles.border
880	-	if focused {
881	-	style = styles.statusBar
882	-	} else {
883	-	style = themeui.DimStyle(style, theme)
884	-	}
885	-	for x := pl.TitleBar.X; x < pl.TitleBar.X+pl.TitleBar.Width && x < frame.Width; x++ {
886	-	frame.Set(x, pl.TitleBar.Y, renderer.Cell{Rune: ' ', Style: style})
887	-	}
888	-	p.mu.Lock()
889	-	title := p.displayTitleLocked()
890	-	generation := p.title.generation
891	-	p.mu.Unlock()
892	-	ui.DrawText(frame, pl.TitleBar.X, pl.TitleBar.Y, pl.TitleBar.X+pl.TitleBar.Width, title, style)
893	-	return generation
894	-	}
895	-
896		func drawDividers(frame renderer.Frame, n *layout.Node, r domain.Rect, style renderer.Style) {	438		func drawDividers(frame renderer.Frame, n *layout.Node, r domain.Rect, style renderer.Style) {
897		if n == nil \|\| n.Kind != layout.Split \|\| len(n.Children) <= 1 {	439		if n == nil \|\| n.Kind != layout.Split \|\| len(n.Children) <= 1 {
898		return	440		return

A internal/usecase/daemon/render_capture.go +335 -0

@@ -0,0 +1,335 @@
	1	+	package daemon
	2	+
	3	+	import (
	4	+	"github.com/bnema/vev/internal/domain"
	5	+	scopy "github.com/bnema/vev/internal/usecase/copy"
	6	+	"github.com/bnema/vev/internal/usecase/layout"
	7	+	"github.com/bnema/vev/internal/usecase/picker"
	8	+	themeui "github.com/bnema/vev/internal/usecase/theme"
	9	+	"github.com/bnema/vev/internal/usecase/ui"
	10	+	"github.com/bnema/vev/pkg/renderer"
	11	+	)
	12	+
	13	+	// capturedRenderState is the immutable hand-off from authoritative daemon
	14	+	// state to composition. Every field is a value or an owned bounded copy; the
	15	+	// composer must not retain or consult session, tab, pane, or overlay owners.
	16	+	// renderCaptureScratch owns bounded snapshots for one attachment. The complete
	17	+	// capture/compose/emit transaction holds that attachment's sendMu, so these
	18	+	// buffers cannot be observed or reused before emission completes.
	19	+	type renderCaptureScratch struct {
	20	+	state capturedRenderState
	21	+	panes []capturedPaneRenderState
	22	+	placements []layout.Placement
	23	+	statusTabs []statusTab
	24	+	mru []recentSession
	25	+	ranked []rankedRecent
	26	+	titleIDs []layout.PaneID
	27	+	treeNodes []layout.Node
	28	+	treeUsed int
	29	+	receipts []damageReceipt
	30	+	}
	31	+
	32	+	type damageReceipt struct {
	33	+	pane *pane
	34	+	generation uint64
	35	+	}
	36	+
	37	+	type capturedRenderState struct {
	38	+	attachment *attachedClient // identity only; never dereferenced by composition
	39	+	lease *attachmentLease
	40	+	reset bool
	41	+	layout capturedTabLayout
	42	+	panes []capturedPaneRenderState
	43	+	floating capturedFloatingRenderState
	44	+	bars barState
	45	+	theme themeui.Theme
	46	+	overlays capturedOverlayRenderState
	47	+	preview picker.Preview
	48	+	cursor capturedCursorInputs
	49	+	tabGeneration uint64
	50	+	floatingGeneration uint64
	51	+	receipts []damageReceipt // private capture receipts; compose must not inspect these
	52	+	}
	53	+
	54	+	type capturedTabLayout struct {
	55	+	root *layout.Node
	56	+	area domain.Rect
	57	+	focus layout.PaneID
	58	+	placements []layout.Placement
	59	+	fingerprint string
	60	+	valid bool
	61	+	}
	62	+
	63	+	type capturedPaneRenderState struct {
	64	+	id layout.PaneID
	65	+	frame renderer.Frame
	66	+	rawDamage []renderer.Damage
	67	+	damage []renderer.Damage
	68	+	damageGeneration uint64
	69	+	title string
	70	+	titleGeneration uint64
	71	+	placement layout.Placement
	72	+	focused bool
	73	+	}
	74	+
	75	+	type capturedFloatingRenderState struct {
	76	+	visible bool
	77	+	pane capturedPaneRenderState
	78	+	geometry floatingGeometry
	79	+	title string
	80	+	generation uint64
	81	+	titleGeneration uint64
	82	+	}
	83	+
	84	+	type capturedOverlayRenderState struct {
	85	+	copyActive, copySearchActive, pickerActive, paletteActive, promptActive bool
	86	+	copyMode *scopy.Mode
	87	+	copySnapshot *scopy.Snapshot
	88	+	copyPaneID layout.PaneID
	89	+	copyFeedback string
	90	+	paletteGuidance string
	91	+	copySearch, picker, palette, prompt capturedModal
	92	+	}
	93	+
	94	+	type capturedModal struct {
	95	+	modal ui.Modal
	96	+	inner renderer.Frame
	97	+	}
	98	+
	99	+	func (o capturedOverlayRenderState) active() bool {
	100	+	return o.copyActive \|\| o.copySearchActive \|\| o.pickerActive \|\| o.paletteActive \|\| o.promptActive
	101	+	}
	102	+
	103	+	type capturedCursorInputs struct {
	104	+	row, col, style int
	105	+	hasStyle, visible, renderable, hiddenByOverlay bool
	106	+	content domain.Rect
	107	+	}
	108	+
	109	+	// capturePaneRenderStateLocked copies only the visible rectangle required by
	110	+	// composition. It never copies scrollback/history and never lets the mutable
	111	+	// VT frame's Cells or row-offset slices escape pane.mu.
	112	+	func capturePaneRenderStateLocked(p *pane, visible domain.Rect) capturedPaneRenderState {
	113	+	return capturePaneRenderStateLockedInto(p, visible, capturedPaneRenderState{})
	114	+	}
	115	+
	116	+	func capturePaneRenderStateLockedInto(p *pane, visible domain.Rect, out capturedPaneRenderState) capturedPaneRenderState {
	117	+	out.id, out.title, out.titleGeneration = p.id, p.displayTitleLocked(), p.title.generation
	118	+	damage := p.screen.CaptureDamage()
	119	+	out.damageGeneration = damage.Generation
	120	+	out.rawDamage = append(out.rawDamage[:0], damage.Damage...)
	121	+
	122	+	width := min(max(visible.Width, 0), p.screen.Frame.Width)
	123	+	height := min(max(visible.Height, 0), p.screen.Frame.Height)
	124	+	needsFrame := len(damage.Damage) > 0
	125	+	frameChanged := out.frame.Width != width \|\| out.frame.Height != height
	126	+	if frameChanged {
	127	+	out.frame = renderer.NewFrame(width, height)
	128	+	needsFrame = true
	129	+	}
	130	+	// A retained attachment snapshot is already an immutable copy of this
	131	+	// pane. With no VT damage, retain it rather than copying inactive panes.
	132	+	if needsFrame {
	133	+	for y := range height {
	134	+	copy(out.frame.Row(y), p.screen.Frame.Row(y)[:width])
	135	+	}
	136	+	}
	137	+
	138	+	if frameChanged \|\| uncertainDamage(damage.Damage, p.screen.Frame.Width, p.screen.Frame.Height) {
	139	+	// A cache entry with a new frame has no terminal-shadow equivalent.
	140	+	// Redraw it even when the pane has no VT damage, such as an untouched
	141	+	// pane first made visible after a tab or session switch.
	142	+	out.damage = []renderer.Damage{renderer.FullRedraw()}
	143	+	} else {
	144	+	out.damage = append(out.damage[:0], damage.Damage...)
	145	+	}
	146	+	// Consumption is transactional: capture only snapshots damage. emitFrame
	147	+	// acknowledges this generation after preparation and transport success.
	148	+	return out
	149	+	}
	150	+
	151	+	func uncertainDamage(damage []renderer.Damage, width, height int) bool {
	152	+	for _, d := range damage {
	153	+	if d.Kind == renderer.DamageFullRedraw {
	154	+	if d != renderer.FullRedraw() {
	155	+	return true
	156	+	}
	157	+	continue
	158	+	}
	159	+	if d.Kind < renderer.DamageText \|\| d.Kind > renderer.DamageScrollUp \|\| d.X < 0 \|\| d.Y < 0 \|\| d.Width <= 0 \|\| d.Height <= 0 \|\| d.X+d.Width > width \|\| d.Y+d.Height > height {
	160	+	return true
	161	+	}
	162	+	if d.Kind == renderer.DamageScrollUp && (d.Count <= 0 \|\| d.Count > d.Height) {
	163	+	return true
	164	+	}
	165	+	}
	166	+	return false
	167	+	}
	168	+
	169	+	// capturePrimaryRenderState is the ownership boundary for a primary render
	170	+	// transaction. Callers hold attachment sendMu; this function then follows
	171	+	// session -> tab -> pane lock order. ACK-blocked capture returns before touching
	172	+	// VT damage, and every captured pane records a receipt for successful emission.
	173	+	func capturePrimaryRenderState(sess session, ac attachedClient, bars barState, overlays capturedOverlayRenderState, preview picker.Preview, floatingCfg domain.FloatingConfig, reset bool, lease attachmentLease) (capturedRenderState, bool) {
	174	+	if sess == nil \|\| ac == nil \|\| (ac.output != nil && ac.output.atCapacity()) {
	175	+	return nil, false
	176	+	}
	177	+	sess.mu.Lock()
	178	+	if sess.client != ac \|\| sess.active < 0 \|\| sess.active >= len(sess.tabs) {
	179	+	sess.mu.Unlock()
	180	+	return nil, false
	181	+	}
	182	+	tb := sess.tabs[sess.active]
	183	+	sess.mu.Unlock()
	184	+
	185	+	scratch := &ac.renderScratch
	186	+	scratch.statusTabs = append(scratch.statusTabs[:0], bars.status.tabs...)
	187	+	bars.status.tabs = scratch.statusTabs
	188	+	scratch.mru = append(scratch.mru[:0], bars.mru...)
	189	+	bars.mru = scratch.mru
	190	+	if bars.rankedRecent != nil {
	191	+	scratch.ranked = copyRankedRecentInto(scratch.ranked, bars.rankedRecent)
	192	+	bars.rankedRecent = scratch.ranked
	193	+	}
	194	+	if len(preview.Rows) > 0 {
	195	+	preview = clonePickerPreview(preview)
	196	+	} else {
	197	+	preview = picker.Preview{Width: preview.Width, Height: preview.Height}
	198	+	}
	199	+
	200	+	tb.mu.Lock()
	201	+	defer tb.mu.Unlock()
	202	+	layoutSnap := solveTabLayoutLocked(tb)
	203	+	scratch.placements = append(scratch.placements[:0], layoutSnap.placements...)
	204	+	scratch.treeUsed = 0
	205	+	var root *layout.Node
	206	+	if tb.tree != nil {
	207	+	root = cloneLayoutNodeIntoScratch(tb.tree.Root, scratch)
	208	+	}
	209	+	state := &scratch.state
	210	+	*state = capturedRenderState{
	211	+	attachment: ac, lease: lease, reset: reset, bars: bars, theme: bars.theme,
	212	+	overlays: overlays, preview: preview,
	213	+	layout: capturedTabLayout{root: root, area: layoutSnap.area, focus: layoutSnap.focus, placements: scratch.placements, fingerprint: layoutSnap.fingerprint, valid: layoutSnap.ok},
	214	+	floatingGeneration: tb.floating.generation,
	215	+	receipts: scratch.receipts[:0],
	216	+	}
	217	+	state.panes = scratch.panes[:0]
	218	+	state.tabGeneration = uint64(len(layoutSnap.fingerprint))
	219	+	if ac.captureFrames == nil {
	220	+	ac.captureFrames = make(map[*pane]capturedPaneRenderState)
	221	+	}
	222	+	for _, placement := range layoutSnap.placements {
	223	+	p := tb.panes[placement.ID]
	224	+	if p == nil {
	225	+	continue
	226	+	}
	227	+	visible := placement.Content
	228	+	if placement.Collapsed {
	229	+	visible = domain.Rect{}
	230	+	}
	231	+	p.mu.Lock()
	232	+	captured := capturePaneRenderStateLockedInto(p, visible, ac.captureFrames[p])
	233	+	state.receipts = append(state.receipts, damageReceipt{pane: p, generation: captured.damageGeneration})
	234	+	captured.placement, captured.focused = placement, placement.ID == layoutSnap.focus
	235	+	if captured.focused {
	236	+	state.cursor = captureCursorInputsLocked(p, placement.Content, overlays)
	237	+	}
	238	+	p.mu.Unlock()
	239	+	translated := captured.rawDamage[:0]
	240	+	for _, damage := range captured.damage {
	241	+	translated = append(translated, translatePaneDamage(damage, placement.Content, layoutSnap.area)...)
	242	+	}
	243	+	captured.rawDamage, captured.damage = captured.damage, translated
	244	+	ac.captureFrames[p] = captured
	245	+	state.panes = append(state.panes, captured)
	246	+	}
	247	+	scratch.panes = state.panes
	248	+	if tb.floating.state == floatingVisible && tb.floating.pane != nil {
	249	+	p := tb.floating.pane
	250	+	p.mu.Lock()
	251	+	geometry := p.committedFloatingGeometryLocked(calculateContentFloatingGeometry(domain.Size{Cols: layoutSnap.area.Width, Rows: layoutSnap.area.Height}, floatingCfg))
	252	+	captured := capturePaneRenderStateLockedInto(p, geometry.Inner, ac.captureFrames[p])
	253	+	seen := false
	254	+	for _, receipt := range state.receipts {
	255	+	if receipt.pane == p {
	256	+	seen = true
	257	+	break
	258	+	}
	259	+	}
	260	+	if !seen {
	261	+	state.receipts = append(state.receipts, damageReceipt{pane: p, generation: captured.damageGeneration})
	262	+	}
	263	+	ac.captureFrames[p] = captured
	264	+	state.floating = capturedFloatingRenderState{visible: true, pane: captured, geometry: geometry, title: captured.title, generation: tb.floating.generation, titleGeneration: captured.titleGeneration}
	265	+	state.cursor = captureCursorInputsLocked(p, geometry.Inner, overlays)
	266	+	p.mu.Unlock()
	267	+	}
	268	+	scratch.receipts = state.receipts
	269	+	return state, true
	270	+	}
	271	+
	272	+	// copyRankedRecentInto preserves the non-nil empty slice that selects
	273	+	// contextual recent-session mode while retaining scratch capacity for reuse.
	274	+	func copyRankedRecentInto(dst, src []rankedRecent) []rankedRecent {
	275	+	if len(src) == 0 {
	276	+	if dst != nil {
	277	+	return dst[:0]
	278	+	}
	279	+	return []rankedRecent{}
	280	+	}
	281	+	return append(dst[:0], src...)
	282	+	}
	283	+
	284	+	func cloneLayoutNodeIntoScratch(src layout.Node, scratch renderCaptureScratch) *layout.Node {
	285	+	if src == nil {
	286	+	return nil
	287	+	}
	288	+	if scratch.treeUsed == len(scratch.treeNodes) {
	289	+	scratch.treeNodes = append(scratch.treeNodes, layout.Node{})
	290	+	}
	291	+	index := scratch.treeUsed
	292	+	scratch.treeUsed++
	293	+	dst := &scratch.treeNodes[index]
	294	+	children := dst.Children[:0]
	295	+	dst = src
	296	+	dst.Children = children
	297	+	for _, child := range src.Children {
	298	+	dst.Children = append(dst.Children, cloneLayoutNodeIntoScratch(child, scratch))
	299	+	}
	300	+	return dst
	301	+	}
	302	+
	303	+	func captureCursorInputsLocked(p *pane, content domain.Rect, overlays capturedOverlayRenderState) capturedCursorInputs {
	304	+	style, hasStyle := p.screen.CursorStyle()
	305	+	hidden := overlays.copyActive \|\| overlays.copySearchActive \|\| overlays.pickerActive \|\| overlays.paletteActive \|\| overlays.promptActive
	306	+	return capturedCursorInputs{row: p.screen.CursorRow(), col: p.screen.CursorCol(), style: style, hasStyle: hasStyle, visible: p.screen.CursorVisible(), renderable: content.Width > 0 && content.Height > 0, hiddenByOverlay: hidden, content: content}
	307	+	}
	308	+
	309	+	func clonePickerPreview(in picker.Preview) picker.Preview {
	310	+	out := picker.Preview{Width: in.Width, Height: in.Height, Rows: make([][]renderer.Cell, len(in.Rows))}
	311	+	for i := range in.Rows {
	312	+	out.Rows[i] = append([]renderer.Cell(nil), in.Rows[i]...)
	313	+	}
	314	+	return out
	315	+	}
	316	+
	317	+	// appendStackPaneIDs returns exactly the panes whose solved placement can own a
	318	+	// stack title bar. The caller holds tab.mu, and dst is attachment scratch.
	319	+	func appendStackPaneIDs(dst []layout.PaneID, node *layout.Node) []layout.PaneID {
	320	+	if node == nil {
	321	+	return dst
	322	+	}
	323	+	if node.Kind == layout.Stack {
	324	+	for _, child := range node.Children {
	325	+	if child != nil && child.Kind == layout.Leaf {
	326	+	dst = append(dst, child.Leaf)
	327	+	}
	328	+	}
	329	+	return dst
	330	+	}
	331	+	for _, child := range node.Children {
	332	+	dst = appendStackPaneIDs(dst, child)
	333	+	}
	334	+	return dst
	335	+	}

A internal/usecase/daemon/render_coordinator.go +880 -0

@@ -0,0 +1,880 @@
	1	+	package daemon
	2	+
	3	+	import (
	4	+	"runtime"
	5	+	"sync"
	6	+	"time"
	7	+
	8	+	"github.com/bnema/vev/internal/ports"
	9	+	)
	10	+
	11	+	type renderCoordinator struct {
	12	+	mu sync.Mutex
	13	+	opts renderCoordinatorOptions
	14	+
	15	+	// Coalesced pending state: latest wins, reset stays sticky, and at most
	16	+	// one deadline wake may be armed at a time (cap-1).
	17	+	pending bool
	18	+	pendingReset bool
	19	+	pendingUrgent bool
	20	+	// queueMarked spans the coordinator's actual pending interval, from first
	21	+	// enqueue until fire consumes the coalesced work.
	22	+	queueMarked bool
	23	+	queueCorrelation ports.RuntimeCorrelation
	24	+	// pendingPreview is accounted separately from the target primary wake:
	25	+	// viewers may consume a coalesced target snapshot while that target is
	26	+	// still ACK-blocked.
	27	+	pendingPreview bool
	28	+	// ackDeferred records that an expired deadline or watchdog was blocked by
	29	+	// output-window capacity. ACK notifications may flush only this state.
	30	+	ackDeferred bool
	31	+	// ackBlocked retains the first rejected ACK-capacity probe until the
	32	+	// matching consume or lifecycle closure releases that exact interval.
	33	+	ackBlocked *ackBlockedSpan
	34	+	// deadlineDue closes the handoff between a timer worker receiving its tick
	35	+	// and fire observing unavailable ACK capacity.
	36	+	deadlineDue bool
	37	+	coalesced int
	38	+	// outputPressure carries recent bulk-burst pressure between completed
	39	+	// batches. It selects a bounded 8–16ms deadline without letting a busy
	40	+	// batch perpetually reset its own timer.
	41	+	outputPressure int
	42	+
	43	+	// previewWake receives the same coalesced wakes for the legacy single
	44	+	// subscriber. previewWakes tracks picker subscriptions by viewer, so one
	45	+	// inactive session cannot replace another viewer's live preview.
	46	+	previewWake func(renderWake)
	47	+	previewWakes map[*attachedClient]previewSubscription
	48	+
	49	+	// lease is the sole primary attachment lifecycle state. A revoked lease is
	50	+	// retained after detach so headless preview delivery remains distinguishable
	51	+	// from an unbound test coordinator.
	52	+	lease *attachmentLease
	53	+	torndown bool
	54	+
	55	+	resize resizeRequestMetadata
	56	+	// Each lane owns one complete immutable worker token at a time.
	57	+	resizeLane timerLane
	58	+	retryLane timerLane
	59	+	metrics renderCoordinatorBurstMetrics
	60	+	// normalLane owns render invalidation deadlines. The supervisor is only
	61	+	// waited by the session owner after terminal teardown has rejected future
	62	+	// registrations.
	63	+	normalLane timerLane
	64	+	armed bool
	65	+	supervisor timerSupervisor
	66	+	// syncBatches is keyed by the stable pane owner. Each pane owns an
	67	+	// independent watchdog; composition remains gated until every batch ends.
	68	+	syncBatches map[pane]syncBatch
	69	+	// syncRegistryVersion changes for every batch identity mutation. fire
	70	+	// snapshots it before evaluating visibility and retries if a producer
	71	+	// publishes or completes a batch during that unlocked evaluation.
	72	+	syncRegistryVersion uint64
	73	+	}
	74	+
	75	+	// newRenderCoordinator constructs a coordinator bound to opts. It starts no
	76	+	// goroutine; deadline scheduling is armed lazily by invalidate.
	77	+	func newRenderCoordinator(opts renderCoordinatorOptions) *renderCoordinator {
	78	+	return &renderCoordinator{opts: opts}
	79	+	}
	80	+
	81	+	// detachNormalTimerLocked removes the complete current token. The caller may
	82	+	// stop it only after dropping c.mu and must never wait for its worker.
	83	+	func (c renderCoordinator) detachNormalTimerLocked() timerToken {
	84	+	return c.normalLane.detachLocked()
	85	+	}
	86	+
	87	+	type syncBatch struct {
	88	+	generation uint64
	89	+	// renderable is evaluated outside c.mu: tab/session visibility locks may be
	90	+	// acquired by this predicate.
	91	+	renderable func() bool
	92	+	force func()
	93	+	lane timerLane
	94	+	}
	95	+
	96	+	func detachSyncBatchLocked(batch syncBatch) timerToken { return batch.lane.detachLocked() }
	97	+
	98	+	func (c renderCoordinator) detachSyncTimersLocked() []timerToken {
	99	+	workers := make([]*timerToken, 0, len(c.syncBatches))
	100	+	for pane, batch := range c.syncBatches {
	101	+	workers = append(workers, detachSyncBatchLocked(batch))
	102	+	delete(c.syncBatches, pane)
	103	+	c.syncRegistryVersion++
	104	+	}
	105	+	return workers
	106	+	}
	107	+
	108	+	// invalidate publishes a coordinator-owned producer transition. Callers that
	109	+	// carry an attachment identity must use invalidateForAttachment instead.
	110	+	func (c *renderCoordinator) invalidate(inv renderInvalidation) {
	111	+	c.invalidateForAttachment(nil, inv)
	112	+	}
	113	+
	114	+	// invalidateForAttachment publishes only while source is still the bound
	115	+	// attachment. This check shares c.mu with replacement, making a stale
	116	+	// attachment's timer callback unable to enqueue work for its replacement.
	117	+	// A nil source represents an internal pane/session invalidation.
	118	+	func (c renderCoordinator) invalidateForAttachment(source attachedClient, inv renderInvalidation) bool {
	119	+	return c.invalidateForAttachmentAtResizeEpoch(source, 0, inv)
	120	+	}
	121	+
	122	+	// invalidateForLease publishes an attachment-owned transition only for the
	123	+	// exact lifecycle incarnation that captured it.
	124	+	func (c renderCoordinator) invalidateForLease(source attachedClient, lease *attachmentLease, inv renderInvalidation) bool {
	125	+	return c.invalidateForLeaseAtResizeEpoch(source, lease, 0, inv)
	126	+	}
	127	+
	128	+	// invalidateForAttachmentAtResizeEpoch additionally requires epoch to remain
	129	+	// current when nonzero, so a completed resize cannot publish after a newer
	130	+	// request supersedes it.
	131	+	func (c renderCoordinator) invalidateForAttachmentAtResizeEpoch(source attachedClient, epoch uint64, inv renderInvalidation) bool {
	132	+	return c.invalidateForLeaseAtResizeEpoch(source, nil, epoch, inv)
	133	+	}
	134	+
	135	+	// invalidateForLeaseAtResizeEpoch is the attachment-callback variant: it
	136	+	// rejects a same-object replacement whose source pointer still matches but
	137	+	// whose lifecycle lease has been revoked.
	138	+	func (c renderCoordinator) invalidateForLeaseAtResizeEpoch(source attachedClient, lease *attachmentLease, epoch uint64, inv renderInvalidation) bool {
	139	+	c.mu.Lock()
	140	+	// An unbound coordinator is the manual/headless harness state; there is no
	141	+	// published production attachment to protect there. A parked target may
	142	+	// still publish its own PTY/session mutations to picker observers, but no
	143	+	// attachment-owned callback may revive that target's render path.
	144	+	detachedPreviewOnly := c.primaryDetachedLocked() && source == nil && (c.previewWake != nil \|\| len(c.previewWakes) != 0)
	145	+	if c.torndown \|\| (lease != nil && (!c.leaseCurrentLocked(lease, true) \|\| lease.attachment != source)) \|\|
	146	+	(c.primaryDetachedLocked() && !detachedPreviewOnly) \|\| (source != nil && c.lease != nil && c.lease.attachment != source) \|\|
	147	+	(epoch != 0 && (c.resize.epoch != epoch \|\| c.resize.source != source \|\| (lease != nil && c.resize.lease != lease))) {
	148	+	c.mu.Unlock()
	149	+	return false
	150	+	}
	151	+	onInvalidate := c.opts.onInvalidate
	152	+	c.metrics.invalidations.Add(1)
	153	+	wasPending, wasUrgent, wasPreviewPending := c.pending, c.pendingUrgent, c.pendingPreview
	154	+	queueStart := false
	155	+	var queueCorrelation ports.RuntimeCorrelation
	156	+	if !wasPending {
	157	+	c.ackDeferred = false
	158	+	c.deadlineDue = false
	159	+	if c.opts.observer != nil {
	160	+	c.queueMarked = true
	161	+	c.queueCorrelation = ports.NewRuntimeCorrelation()
	162	+	queueCorrelation = c.queueCorrelation
	163	+	queueStart = true
	164	+	}
	165	+	}
	166	+	c.pending = true
	167	+	c.pendingReset = c.pendingReset \|\| inv.reset
	168	+	c.pendingUrgent = c.pendingUrgent \|\| inv.class == invalidateUrgent
	169	+	c.pendingPreview = c.pendingPreview \|\| c.previewWake != nil \|\| len(c.previewWakes) != 0
	170	+	c.coalesced++
	171	+	// A deadline may have expired while synchronized output still gated the
	172	+	// pending work. Completion republishes urgently and must reserve a fresh
	173	+	// deadline rather than leaving that already-fired timer as the only arm.
	174	+	arm := !wasPending \|\| (!wasUrgent && c.pendingUrgent) \|\| (!wasPreviewPending && c.pendingPreview) \|\| c.deadlineDue
	175	+	var old *timerToken
	176	+	if arm {
	177	+	_, old = c.normalLane.replaceLocked()
	178	+	}
	179	+	gen := c.normalLane.generation
	180	+	delay := minOutputRenderDeadline + time.Duration(c.outputPressure)*time.Millisecond
	181	+	if c.pendingUrgent {
	182	+	delay = urgentRenderDeadline
	183	+	}
	184	+	clock := c.opts.clock
	185	+	c.armed = c.armed \|\| arm
	186	+	observer := c.opts.observer
	187	+	c.mu.Unlock()
	188	+	if queueStart && observer != nil {
	189	+	observer.ObserveRuntime(ports.NewRuntimeMarkWithCorrelation("daemon", queueCorrelation, ports.RuntimeQueueEnqueued, 0, true))
	190	+	}
	191	+	stopDetachedTimer(old)
	192	+	if !arm \|\| clock == nil {
	193	+	if onInvalidate != nil {
	194	+	onInvalidate(inv)
	195	+	}
	196	+	return true
	197	+	}
	198	+
	199	+	// NewTimer and C may re-enter the coordinator. Publish only if this
	200	+	// generation is still the reserved deadline after those external calls.
	201	+	timer := clock.NewTimer(delay)
	202	+	timerC := timer.C()
	203	+	c.mu.Lock()
	204	+	valid := !c.torndown && c.pending && c.normalLane.generation == gen
	205	+	var token *timerToken
	206	+	if valid && timerC != nil {
	207	+	token = c.normalLane.publishLocked(gen, timer)
	208	+	valid = token != nil
	209	+	if valid {
	210	+	c.supervisor.startLocked(token, timerC, func() {
	211	+	c.markDeadlineDue(token)
	212	+	c.fireFromTimer(token, false, true)
	213	+	})
	214	+	}
	215	+	}
	216	+	c.mu.Unlock()
	217	+	if !valid {
	218	+	timer.Stop()
	219	+	if onInvalidate != nil {
	220	+	onInvalidate(inv)
	221	+	}
	222	+	return true
	223	+	}
	224	+	if timerC == nil {
	225	+	timer.Stop()
	226	+	c.fire(gen, false, true)
	227	+	if onInvalidate != nil {
	228	+	onInvalidate(inv)
	229	+	}
	230	+	return true
	231	+	}
	232	+	// Test-visible observation follows deadline publication, so a producer
	233	+	// callback cannot observe a successful invalidation before its timer owns
	234	+	// a worker (and race tests cannot finish while a mock expectation remains).
	235	+	if onInvalidate != nil {
	236	+	onInvalidate(inv)
	237	+	}
	238	+	runtime.Gosched()
	239	+	return true
	240	+	}
	241	+
	242	+	// notifyAck reports that the client acknowledged an output state, releasing
	243	+	// at most one deadline/watchdog-deferred wake. It never bypasses an unexpired
	244	+	// normal or urgent deadline.
	245	+	func (c *renderCoordinator) notifyAck() {
	246	+	c.mu.Lock()
	247	+	if (!c.ackDeferred && !c.deadlineDue) \|\| !c.pending {
	248	+	c.mu.Unlock()
	249	+	return
	250	+	}
	251	+	gen := c.normalLane.generation
	252	+	blocked := c.ackBlocked
	253	+	c.ackBlocked = nil
	254	+	c.mu.Unlock()
	255	+	if blocked != nil {
	256	+	blocked.finish(true)
	257	+	}
	258	+	c.fire(gen, false, false)
	259	+	}
	260	+
	261	+	// noteSyncBegin records a synchronized-output batch for its stable pane and
	262	+	// arms that pane's watchdog. Overlapping pane batches are independent.
	263	+	// markDeadlineDue records a received timer tick before the worker probes ACK
	264	+	// readiness, so a concurrent ACK cannot miss an already-expired deadline.
	265	+	func (c renderCoordinator) markDeadlineDue(token timerToken) {
	266	+	c.mu.Lock()
	267	+	if c.fireValidLocked(token, token.generation, false) {
	268	+	c.deadlineDue = true
	269	+	}
	270	+	c.mu.Unlock()
	271	+	}
	272	+
	273	+	func (c renderCoordinator) noteSyncBegin(p pane, gen uint64, force ...func()) {
	274	+	var renderable func() bool
	275	+	if c.opts.syncRenderable != nil {
	276	+	renderable = func() bool { return c.opts.syncRenderable(p) }
	277	+	}
	278	+	c.noteSyncBeginWithRenderability(p, gen, renderable, force...)
	279	+	}
	280	+
	281	+	// noteSyncBeginWithRenderability records lifecycle unconditionally while the
	282	+	// supplied predicate decides dynamically whether this batch gates composition.
	283	+	func (c renderCoordinator) noteSyncBeginWithRenderability(p pane, gen uint64, renderable func() bool, force ...func()) {
	284	+	c.mu.Lock()
	285	+	if c.torndown {
	286	+	c.mu.Unlock()
	287	+	return
	288	+	}
	289	+	if c.syncBatches == nil {
	290	+	c.syncBatches = make(map[pane]syncBatch)
	291	+	}
	292	+	var old *timerToken
	293	+	if previous := c.syncBatches[p]; previous != nil {
	294	+	old = detachSyncBatchLocked(previous)
	295	+	}
	296	+	batch := &syncBatch{generation: gen, renderable: renderable}
	297	+	if len(force) != 0 {
	298	+	batch.force = force[0]
	299	+	}
	300	+	c.syncBatches[p] = batch
	301	+	c.syncRegistryVersion++
	302	+	clock := c.opts.clock
	303	+	c.mu.Unlock()
	304	+	stopDetachedTimer(old)
	305	+	if clock == nil {
	306	+	return
	307	+	}
	308	+
	309	+	// Both timer operations are external. A reentrant callback may replace or
	310	+	// remove batch; identity validation below rejects this stale timer.
	311	+	timer := clock.NewTimer(maxSyncUpdateDuration)
	312	+	timerC := timer.C()
	313	+	c.mu.Lock()
	314	+	valid := !c.torndown && c.syncBatches[p] == batch && batch.generation == gen && batch.lane.generation == 0
	315	+	var token *timerToken
	316	+	if valid && timerC != nil {
	317	+	batch.lane.generation = gen
	318	+	token = batch.lane.publishLocked(gen, timer)
	319	+	valid = token != nil
	320	+	if valid {
	321	+	c.supervisor.startLocked(token, timerC, func() { c.runSyncWatchdog(p, batch, token) })
	322	+	}
	323	+	}
	324	+	c.mu.Unlock()
	325	+	if !valid \|\| timerC == nil {
	326	+	stopDetachedTimer(&timerToken{timer: timer})
	327	+	return
	328	+	}
	329	+	}
	330	+
	331	+	func (c renderCoordinator) runSyncWatchdog(p pane, batch syncBatch, token timerToken) {
	332	+	// Keep the batch registered while force runs. A concurrent render must
	333	+	// continue to observe the synchronized-output gate until the VT has
	334	+	// authoritatively closed the batch.
	335	+	c.mu.Lock()
	336	+	current := c.syncBatches[p]
	337	+	valid := !c.torndown && current == batch && current.generation == token.generation && current.lane.token == token
	338	+	c.mu.Unlock()
	339	+	if !valid {
	340	+	return
	341	+	}
	342	+	if batch.force != nil {
	343	+	batch.force()
	344	+	}
	345	+
	346	+	// force may synchronously end this batch, replace it with a newer
	347	+	// generation, or tear down the coordinator. Remove only the exact
	348	+	// snapshot that expired, then publish one urgent completion wake.
	349	+	c.mu.Lock()
	350	+	current = c.syncBatches[p]
	351	+	valid = !c.torndown && current == batch && current.generation == token.generation && current.lane.token == token
	352	+	var stopped *timerToken
	353	+	if valid {
	354	+	delete(c.syncBatches, p)
	355	+	c.syncRegistryVersion++
	356	+	stopped = detachSyncBatchLocked(batch)
	357	+	if len(c.syncBatches) == 0 && c.pending {
	358	+	c.pendingUrgent = true
	359	+	}
	360	+	}
	361	+	c.mu.Unlock()
	362	+	if valid {
	363	+	stopDetachedTimer(stopped)
	364	+	// Reevaluate the aggregate gate only after forcing this pane.
	365	+	c.fireCurrent(true)
	366	+	}
	367	+	}
	368	+
	369	+	// noteSyncEnd records completion for exactly one pane batch. It only flushes
	370	+	// after the aggregate gate opens.
	371	+	func (c renderCoordinator) noteSyncEnd(p pane, gen uint64) {
	372	+	// Visibility may acquire session/tab locks, so evaluate it outside c.mu.
	373	+	c.mu.Lock()
	374	+	batch := c.syncBatches[p]
	375	+	if batch == nil \|\| batch.generation != gen {
	376	+	c.mu.Unlock()
	377	+	return
	378	+	}
	379	+	renderable := batch.renderable
	380	+	c.mu.Unlock()
	381	+	wasRenderable := renderable == nil \|\| renderable()
	382	+	if c.removeSyncEnd(p, gen, wasRenderable) {
	383	+	c.fireCurrent(false)
	384	+	}
	385	+	}
	386	+
	387	+	// removeSyncEnd removes a batch without invoking a predicate or callback. PTY
	388	+	// processing calls it while pane.mu is still held, so a deadline can never
	389	+	// compose the post-Write partial screen between mutation and publication.
	390	+	// The caller must invoke fireCurrent only after releasing pane.mu.
	391	+	func (c renderCoordinator) removeSyncEnd(p pane, gen uint64, renderable bool) bool {
	392	+	c.mu.Lock()
	393	+	batch := c.syncBatches[p]
	394	+	if batch == nil \|\| batch.generation != gen {
	395	+	c.mu.Unlock()
	396	+	return false
	397	+	}
	398	+	delete(c.syncBatches, p)
	399	+	c.syncRegistryVersion++
	400	+	worker := detachSyncBatchLocked(batch)
	401	+	if renderable && len(c.syncBatches) == 0 && c.pending {
	402	+	c.pendingUrgent = true
	403	+	}
	404	+	c.mu.Unlock()
	405	+	stopDetachedTimer(worker)
	406	+	return renderable
	407	+	}
	408	+
	409	+	// noteSyncPaneRemoved releases a pane watchdog when pane lifecycle ends.
	410	+	func (c renderCoordinator) noteSyncPaneRemoved(p pane) {
	411	+	c.mu.Lock()
	412	+	var worker *timerToken
	413	+	if batch := c.syncBatches[p]; batch != nil {
	414	+	delete(c.syncBatches, p)
	415	+	c.syncRegistryVersion++
	416	+	worker = detachSyncBatchLocked(batch)
	417	+	}
	418	+	c.mu.Unlock()
	419	+	stopDetachedTimer(worker)
	420	+	}
	421	+
	422	+	// subscribePreview installs fn as the preview observer for coalesced wakes.
	423	+	func (c *renderCoordinator) subscribePreview(fn func(renderWake)) {
	424	+	c.mu.Lock()
	425	+	c.previewWake = fn
	426	+	c.mu.Unlock()
	427	+	}
	428	+
	429	+	// teardownPreview removes the legacy preview subscription.
	430	+	func (c *renderCoordinator) teardownPreview() { c.mu.Lock(); c.previewWake = nil; c.mu.Unlock() }
	431	+
	432	+	type previewSubscription struct {
	433	+	generation uint64
	434	+	fn func(renderWake)
	435	+	}
	436	+
	437	+	// subscribePreviewFor installs the dynamic picker observer owned by viewer.
	438	+	// A newer picker generation wins, so a delayed subscription cannot replace it.
	439	+	func (c renderCoordinator) subscribePreviewFor(viewer attachedClient, generation uint64, fn func(renderWake)) bool {
	440	+	if viewer == nil \|\| fn == nil {
	441	+	return false
	442	+	}
	443	+	c.mu.Lock()
	444	+	defer c.mu.Unlock()
	445	+	if c.torndown {
	446	+	return false
	447	+	}
	448	+	if current, ok := c.previewWakes[viewer]; ok && current.generation > generation {
	449	+	return false
	450	+	}
	451	+	if c.previewWakes == nil {
	452	+	c.previewWakes = make(map[*attachedClient]previewSubscription)
	453	+	}
	454	+	c.previewWakes[viewer] = previewSubscription{generation: generation, fn: fn}
	455	+	return true
	456	+	}
	457	+
	458	+	// teardownPreviewFor removes viewer's observer only when it still belongs to
	459	+	// generation. A delayed teardown cannot clear a replacement subscription.
	460	+	func (c renderCoordinator) teardownPreviewFor(viewer attachedClient, generation uint64) {
	461	+	c.mu.Lock()
	462	+	if current, ok := c.previewWakes[viewer]; ok && current.generation == generation {
	463	+	delete(c.previewWakes, viewer)
	464	+	}
	465	+	c.mu.Unlock()
	466	+	}
	467	+
	468	+	// primaryDetachedLocked distinguishes an explicitly detached primary from an
	469	+	// unbound coordinator used by headless tests. Preview subscribers remain live
	470	+	// for the former, but it can never regain a primary transport without a lease.
	471	+	func (c *renderCoordinator) primaryDetachedLocked() bool { return c.lease != nil && !c.lease.active }
	472	+
	473	+	func (c renderCoordinator) installLeaseLocked(ac attachedClient, ready bool) *attachmentLease {
	474	+	if c.lease != nil {
	475	+	c.lease.active = false
	476	+	}
	477	+	c.lease = &attachmentLease{attachment: ac, ready: ready, active: ac != nil}
	478	+	return c.lease
	479	+	}
	480	+
	481	+	// attachmentLease returns the exact currently published lease for ac. Welcome
	482	+	// continuations capture this before their transport write and cannot later
	483	+	// bless a same-object park/resume replacement.
	484	+	func (c renderCoordinator) attachmentLease(ac attachedClient) *attachmentLease {
	485	+	c.mu.Lock()
	486	+	defer c.mu.Unlock()
	487	+	if c.torndown \|\| c.lease == nil \|\| !c.lease.active \|\| c.lease.attachment != ac {
	488	+	return nil
	489	+	}
	490	+	return c.lease
	491	+	}
	492	+
	493	+	func (c renderCoordinator) leaseCurrent(lease attachmentLease, ready bool) bool {
	494	+	c.mu.Lock()
	495	+	defer c.mu.Unlock()
	496	+	return c.leaseCurrentLocked(lease, ready)
	497	+	}
	498	+
	499	+	// leaseCurrentLocked is a short, non-reentrant effect admission check. It
	500	+	// must not be used to retain c.mu across input routing, transport operations,
	501	+	// or other arbitrary handler work.
	502	+	func (c renderCoordinator) leaseCurrentLocked(lease attachmentLease, ready bool) bool {
	503	+	return !c.torndown && lease != nil && c.lease == lease && lease.active && (!ready \|\| lease.ready)
	504	+	}
	505	+
	506	+	// attach binds an already-handshaken internal attachment. Route/resume use
	507	+	// attachWithReadiness(..., false) until their Welcome frame is accepted.
	508	+	func (c renderCoordinator) attach(ac attachedClient) { c.attachWithReadiness(ac, true) }
	509	+
	510	+	func (c renderCoordinator) attachWithReadiness(ac attachedClient, ready bool) {
	511	+	c.mu.Lock()
	512	+	if !c.torndown && (c.lease == nil \|\| !c.lease.active \|\| c.lease.attachment != ac) {
	513	+	c.installLeaseLocked(ac, ready)
	514	+	}
	515	+	c.mu.Unlock()
	516	+	}
	517	+
	518	+	// markAttachmentReady completes only the captured attachment incarnation.
	519	+	// A stale Welcome must never bless a replacement that reused its client object.
	520	+	func (c renderCoordinator) markAttachmentReady(lease attachmentLease) bool {
	521	+	c.mu.Lock()
	522	+	defer c.mu.Unlock()
	523	+	if !c.leaseCurrentLocked(lease, false) {
	524	+	return false
	525	+	}
	526	+	lease.ready = true
	527	+	return true
	528	+	}
	529	+
	530	+	// wakeCurrent validates a dispatched wake. paint repeats this check after
	531	+	// sendMu is acquired, so a reused attachedClient cannot revive an old wake.
	532	+	func (c *renderCoordinator) wakeCurrent(w renderWake) bool {
	533	+	c.mu.Lock()
	534	+	defer c.mu.Unlock()
	535	+	return c.leaseCurrentLocked(w.lease, true)
	536	+	}
	537	+
	538	+	type renderLifecycleCleanup struct {
	539	+	tokens []*timerToken
	540	+	observer ports.RuntimeObserver
	541	+	queueCorrelation ports.RuntimeCorrelation
	542	+	queueMarked bool
	543	+	ackBlocked *ackBlockedSpan
	544	+	}
	545	+
	546	+	func (cleanup renderLifecycleCleanup) finish() {
	547	+	for _, token := range cleanup.tokens {
	548	+	stopDetachedTimer(token)
	549	+	}
	550	+	if cleanup.queueMarked && cleanup.observer != nil {
	551	+	cleanup.observer.ObserveRuntime(ports.NewRuntimeMarkWithCorrelation("daemon", cleanup.queueCorrelation, ports.RuntimeQueueDequeued, 0, false))
	552	+	}
	553	+	if cleanup.ackBlocked != nil {
	554	+	cleanup.ackBlocked.finish(false)
	555	+	}
	556	+	}
	557	+
	558	+	// beginDetach invalidates pending work. It never waits for detached workers.
	559	+	func (c renderCoordinator) beginDetach(ac attachedClient) renderLifecycleCleanup {
	560	+	c.mu.Lock()
	561	+	var cleanup renderLifecycleCleanup
	562	+	if c.lease != nil && c.lease.active && c.lease.attachment == ac {
	563	+	c.lease.active = false
	564	+	c.pending = false
	565	+	c.pendingReset = false
	566	+	c.pendingUrgent = false
	567	+	c.ackDeferred = false
	568	+	c.pendingPreview = false
	569	+	c.coalesced = 0
	570	+	_, timer := c.normalLane.replaceLocked()
	571	+	c.armed = false
	572	+	_, resizeTimer := c.resizeLane.replaceLocked()
	573	+	_, retryTimer := c.retryLane.replaceLocked()
	574	+	cleanup = renderLifecycleCleanup{
	575	+	tokens: []*timerToken{timer, resizeTimer, retryTimer},
	576	+	observer: c.opts.observer,
	577	+	queueCorrelation: c.queueCorrelation,
	578	+	queueMarked: c.queueMarked,
	579	+	ackBlocked: c.ackBlocked,
	580	+	}
	581	+	c.queueMarked, c.ackBlocked = false, nil
	582	+	}
	583	+	c.mu.Unlock()
	584	+	return cleanup
	585	+	}
	586	+
	587	+	// noteDetach is for callers that hold no outer lifecycle locks.
	588	+	func (c renderCoordinator) noteDetach(ac attachedClient) {
	589	+	c.beginDetach(ac).finish()
	590	+	}
	591	+
	592	+	// beginReplace hands the coordinator from old to replacement without waiting
	593	+	// for detached workers.
	594	+	func (c renderCoordinator) beginReplace(old, replacement attachedClient, ready bool) renderLifecycleCleanup {
	595	+	c.mu.Lock()
	596	+	var cleanup renderLifecycleCleanup
	597	+	// A coordinator may be installed while replacing a legacy attachment
	598	+	// which predated coordinator ownership. In that case nil has no pending
	599	+	// identity to invalidate, and the replacement becomes the first bound
	600	+	// attachment atomically with this lifecycle transition.
	601	+	if c.lease == nil \|\| c.lease.attachment == old {
	602	+	c.installLeaseLocked(replacement, ready)
	603	+	c.pending = false
	604	+	c.pendingReset = false
	605	+	c.pendingUrgent = false
	606	+	c.ackDeferred = false
	607	+	c.pendingPreview = false
	608	+	c.coalesced = 0
	609	+	_, timer := c.normalLane.replaceLocked()
	610	+	c.armed = false
	611	+	_, resizeTimer := c.resizeLane.replaceLocked()
	612	+	_, retryTimer := c.retryLane.replaceLocked()
	613	+	cleanup = renderLifecycleCleanup{
	614	+	tokens: []*timerToken{timer, resizeTimer, retryTimer},
	615	+	observer: c.opts.observer,
	616	+	queueCorrelation: c.queueCorrelation,
	617	+	queueMarked: c.queueMarked,
	618	+	ackBlocked: c.ackBlocked,
	619	+	}
	620	+	c.queueMarked, c.ackBlocked = false, nil
	621	+	}
	622	+	c.mu.Unlock()
	623	+	return cleanup
	624	+	}
	625	+
	626	+	// noteReplace is for callers that hold no outer lifecycle locks.
	627	+	func (c renderCoordinator) noteReplace(old, replacement attachedClient, readiness ...bool) {
	628	+	ready := true
	629	+	if len(readiness) != 0 {
	630	+	ready = readiness[0]
	631	+	}
	632	+	c.beginReplace(old, replacement, ready).finish()
	633	+	}
	634	+
	635	+	// notePark invalidates pending wakes when the attachment parks for resume.
	636	+	func (c renderCoordinator) notePark(ac attachedClient) { c.noteDetach(ac) }
	637	+
	638	+	// beginSessionTeardown is terminal phase one. It rejects future worker
	639	+	// registration and atomically detaches every complete lane token.
	640	+	func (c *renderCoordinator) beginSessionTeardown() renderLifecycleCleanup {
	641	+	c.mu.Lock()
	642	+	if c.torndown {
	643	+	c.mu.Unlock()
	644	+	return renderLifecycleCleanup{}
	645	+	}
	646	+	c.torndown = true
	647	+	if c.lease != nil {
	648	+	c.lease.active = false
	649	+	}
	650	+	c.previewWake = nil
	651	+	c.previewWakes = nil
	652	+	c.pending = false
	653	+	_, resizeTimer := c.resizeLane.replaceLocked()
	654	+	_, retryTimer := c.retryLane.replaceLocked()
	655	+	c.ackDeferred = false
	656	+	c.pendingPreview = false
	657	+	_, timer := c.normalLane.replaceLocked()
	658	+	c.armed = false
	659	+	workers := c.detachSyncTimersLocked()
	660	+	observer, queueCorrelation, queueMarked := c.opts.observer, c.queueCorrelation, c.queueMarked
	661	+	ackBlocked := c.ackBlocked
	662	+	c.queueMarked, c.ackBlocked = false, nil
	663	+	c.mu.Unlock()
	664	+	return renderLifecycleCleanup{
	665	+	tokens: append([]*timerToken{timer, resizeTimer, retryTimer}, workers...),
	666	+	observer: observer,
	667	+	queueCorrelation: queueCorrelation,
	668	+	queueMarked: queueMarked,
	669	+	ackBlocked: ackBlocked,
	670	+	}
	671	+	}
	672	+
	673	+	// waitForTimerWorkers is terminal phase two. Only the session owner invokes it,
	674	+	// after beginSessionTeardown has made future supervisor registration impossible.
	675	+	func (c *renderCoordinator) waitForTimerWorkers() { c.supervisor.wait() }
	676	+
	677	+	// invalidateRender is the sole producer fan-in. In tests and transitional
	678	+	// headless paths without an attached coordinator it retains the old private
	679	+	// compositor; attached sessions always schedule through their coordinator.
	680	+	func (d Daemon) invalidateRender(sess session, ac *attachedClient, reset bool, producer string) {
	681	+	if sess != nil {
	682	+	if rc := sess.renderCoordinator(); rc != nil {
	683	+	rc.invalidateForAttachment(ac, renderInvalidation{class: invalidateUrgent, reset: reset, producer: producer})
	684	+	return
	685	+	}
	686	+	}
	687	+	if ac != nil {
	688	+	d.paint(sess, ac, reset, nil)
	689	+	}
	690	+	}
	691	+
	692	+	// invalidateRenderNow publishes through the coordinator but immediately
	693	+	// flushes the wake when the client can accept state. Attach uses this path so
	694	+	// the required first full frame never depends on a debounce timer.
	695	+	func (d Daemon) invalidateRenderNow(sess session, ac *attachedClient, reset bool, producer string) {
	696	+	if sess != nil {
	697	+	if rc := sess.renderCoordinator(); rc != nil {
	698	+	rc.invalidateForAttachment(ac, renderInvalidation{class: invalidateUrgent, reset: reset, producer: producer})
	699	+	rc.fireCurrent(false)
	700	+	return
	701	+	}
	702	+	}
	703	+	if ac != nil {
	704	+	d.paint(sess, ac, reset, nil)
	705	+	}
	706	+	}
	707	+
	708	+	func (c *renderCoordinator) fire(gen uint64, watchdog, deadline bool) {
	709	+	c.mu.Lock()
	710	+	if !watchdog && c.normalLane.generation != gen {
	711	+	c.mu.Unlock()
	712	+	return
	713	+	}
	714	+	token := c.normalLane.token
	715	+	c.mu.Unlock()
	716	+	c.fireWithTimerToken(token, gen, watchdog, deadline)
	717	+	}
	718	+
	719	+	// fireFromTimer consumes a normal deadline using its exact immutable token.
	720	+	func (c renderCoordinator) fireFromTimer(token timerToken, watchdog, deadline bool) {
	721	+	c.fireWithTimerToken(token, token.generation, watchdog, deadline)
	722	+	}
	723	+
	724	+	func (c renderCoordinator) fireWithTimerToken(token timerToken, gen uint64, watchdog, deadline bool) {
	725	+	// Readiness probes enter the attachment send path, which may in turn read
	726	+	// coordinator metadata. Never hold c.mu across that external callback.
	727	+	c.mu.Lock()
	728	+	if !c.fireValidLocked(token, gen, watchdog) {
	729	+	c.mu.Unlock()
	730	+	return
	731	+	}
	732	+	// Welcome holds attachment.sendMu while its transport Send is in flight.
	733	+	// Do not probe ACK capacity through that lock before this incarnation is
	734	+	// ready: a deadline that expires during the handshake must complete and
	735	+	// leave work pending, rather than wedging behind Welcome.
	736	+	handshakePending := c.lease != nil && c.lease.active && !c.lease.ready
	737	+	ackReady := c.opts.ackReady
	738	+	c.mu.Unlock()
	739	+	ready := !handshakePending && (ackReady == nil \|\| ackReady())
	740	+
	741	+	// The readiness callback can detach, replace, or publish a sync batch.
	742	+	// syncGateOpenLocked retries predicate evaluation on every registry change
	743	+	// and returns holding c.mu, so pending consumption has no intervening gap.
	744	+	c.mu.Lock()
	745	+	if !c.fireValidLocked(token, gen, watchdog) {
	746	+	c.mu.Unlock()
	747	+	return
	748	+	}
	749	+	if !c.syncGateOpenLocked() {
	750	+	c.mu.Unlock()
	751	+	return
	752	+	}
	753	+	if !c.fireValidLocked(token, gen, watchdog) {
	754	+	c.mu.Unlock()
	755	+	return
	756	+	}
	757	+	// A detached target has no primary transport or ACK window. Its pending
	758	+	// state belongs exclusively to picker previews and must be consumed after
	759	+	// notification rather than accumulating ackDeferred forever. Do not pass
	760	+	// its revoked primary lease to preview consumers.
	761	+	headlessPreviewOnly := c.primaryDetachedLocked()
	762	+	lease := c.lease
	763	+	if headlessPreviewOnly {
	764	+	lease = nil
	765	+	}
	766	+	w := renderWake{reset: c.pendingReset, urgent: c.pendingUrgent, coalesced: c.coalesced, watchdog: watchdog, lease: lease}
	767	+	if !headlessPreviewOnly && (!ready \|\| (c.lease != nil && c.lease.active && !c.lease.ready)) {
	768	+	var ackStart *ackBlockedSpan
	769	+	if !ready && deadline {
	770	+	c.ackDeferred = true
	771	+	if c.ackBlocked == nil && c.opts.observer != nil {
	772	+	// Install the span while coordinator state is owned. Its Start
	773	+	// publication runs unlocked, and its own handoff defers End until
	774	+	// that call has returned.
	775	+	ackStart = newACKBlockedSpan(c.opts.observer)
	776	+	c.ackBlocked = ackStart
	777	+	}
	778	+	}
	779	+	preview, previews := c.takePendingPreviewsLocked()
	780	+	c.mu.Unlock()
	781	+	if ackStart != nil {
	782	+	ackStart.publishStart()
	783	+	}
	784	+	c.notifyPreviews(w, preview, previews)
	785	+	return
	786	+	}
	787	+	preview, previews := c.takePendingPreviewsLocked()
	788	+	c.metrics.wakes.Add(1)
	789	+	c.metrics.coalesced.Add(uint64(w.coalesced))
	790	+	if !w.urgent {
	791	+	if w.coalesced > 1 {
	792	+	c.outputPressure += w.coalesced - 1
	793	+	if c.outputPressure > int((maxOutputRenderDeadline-minOutputRenderDeadline)/time.Millisecond) {
	794	+	c.outputPressure = int((maxOutputRenderDeadline - minOutputRenderDeadline) / time.Millisecond)
	795	+	}
	796	+	} else if c.outputPressure > 0 {
	797	+	c.outputPressure--
	798	+	}
	799	+	}
	800	+	worker := c.detachNormalTimerLocked()
	801	+	c.pending, c.pendingReset, c.pendingUrgent, c.ackDeferred, c.deadlineDue, c.coalesced, c.armed = false, false, false, false, false, 0, false
	802	+	observer, queueCorrelation, queueMarked := c.opts.observer, c.queueCorrelation, c.queueMarked
	803	+	ackBlocked := c.ackBlocked
	804	+	c.queueMarked, c.ackBlocked = false, nil
	805	+	wake := c.opts.wake
	806	+	c.mu.Unlock()
	807	+	stopDetachedTimer(worker)
	808	+	if queueMarked && observer != nil {
	809	+	observer.ObserveRuntime(ports.NewRuntimeMarkWithCorrelation("daemon", queueCorrelation, ports.RuntimeQueueDequeued, 0, true))
	810	+	}
	811	+	if ackBlocked != nil {
	812	+	ackBlocked.finish(true)
	813	+	}
	814	+	if wake != nil && !headlessPreviewOnly {
	815	+	wake(w)
	816	+	}
	817	+	c.notifyPreviews(w, preview, previews)
	818	+	}
	819	+
	820	+	func (c renderCoordinator) fireValidLocked(token timerToken, gen uint64, watchdog bool) bool {
	821	+	if c.torndown \|\| !c.pending \|\| watchdog {
	822	+	return !c.torndown && c.pending
	823	+	}
	824	+	return c.normalLane.generation == gen && (token == nil \|\| c.normalLane.token == token)
	825	+	}
	826	+
	827	+	// syncGateOpen evaluates pane visibility without c.mu, then reacquires and
	828	+	// validates the complete batch identity set before accepting that result.
	829	+	// syncGateOpen evaluates visibility outside c.mu and returns with c.mu held
	830	+	// only when the registry version still matches the snapshot. The caller must
	831	+	// consume pending state before unlocking, eliminating the gate-to-consume gap.
	832	+	func (c *renderCoordinator) syncGateOpenLocked() bool {
	833	+	for {
	834	+	version := c.syncRegistryVersion
	835	+	batches := make([]*syncBatch, 0, len(c.syncBatches))
	836	+	for _, batch := range c.syncBatches {
	837	+	batches = append(batches, batch)
	838	+	}
	839	+	c.mu.Unlock()
	840	+
	841	+	gated := false
	842	+	for _, batch := range batches {
	843	+	if batch.renderable == nil \|\| batch.renderable() {
	844	+	gated = true
	845	+	break
	846	+	}
	847	+	}
	848	+	if hook := c.opts.afterSyncGateEvaluated; hook != nil {
	849	+	hook()
	850	+	}
	851	+
	852	+	c.mu.Lock()
	853	+	if version == c.syncRegistryVersion {
	854	+	return !gated
	855	+	}
	856	+	}
	857	+	}
	858	+
	859	+	// takePendingPreviewsLocked snapshots and consumes the preview delivery for
	860	+	// this target generation. Unlike the target primary, this is never ACK-gated.
	861	+	func (c *renderCoordinator) takePendingPreviewsLocked() (func(renderWake), []func(renderWake)) {
	862	+	if !c.pendingPreview {
	863	+	return nil, nil
	864	+	}
	865	+	c.pendingPreview = false
	866	+	previews := make([]func(renderWake), 0, len(c.previewWakes))
	867	+	for _, subscription := range c.previewWakes {
	868	+	previews = append(previews, subscription.fn)
	869	+	}
	870	+	return c.previewWake, previews
	871	+	}
	872	+
	873	+	func (c *renderCoordinator) notifyPreviews(w renderWake, preview func(renderWake), previews []func(renderWake)) {
	874	+	if preview != nil {
	875	+	preview(w)
	876	+	}
	877	+	for _, fn := range previews {
	878	+	fn(w)
	879	+	}
	880	+	}

A internal/usecase/daemon/render_coordinator_handshake_test.go +70 -0

@@ -0,0 +1,70 @@
	1	+	package daemon
	2	+
	3	+	import (
	4	+	"testing"
	5	+	"time"
	6	+
	7	+	"github.com/stretchr/testify/require"
	8	+	)
	9	+
	10	+	// Fresh producer state may arrive after route has published an attachment but
	11	+	// before the transport has accepted Welcome. It must remain pending until the
	12	+	// exact attachment incarnation completes that handshake.
	13	+	func TestRenderCoordinatorFreshWakeWaitsForWelcome(t *testing.T) {
	14	+	h := newCoordinatorHarness(t)
	15	+	ac := &attachedClient{}
	16	+	h.rc.attachWithReadiness(ac, false)
	17	+	lease := h.rc.attachmentLease(ac)
	18	+	require.NotNil(t, lease)
	19	+
	20	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput, reset: true, producer: "pty"})
	21	+	timer := awaitCoordinatorScheduledTimer(t, h.clk)
	22	+	h.rc.mu.Lock()
	23	+	workerDone := h.rc.normalLane.token.done
	24	+	h.rc.mu.Unlock()
	25	+	require.NotNil(t, workerDone)
	26	+	timer.ch <- time.Time{}
	27	+	requireWorkerExit(t, workerDone)
	28	+	requireNoWake(t, h.wakes)
	29	+
	30	+	require.True(t, h.rc.markAttachmentReady(lease))
	31	+	h.rc.fireCurrent(false)
	32	+	wake := awaitWake(t, h.wakes)
	33	+	require.True(t, wake.reset)
	34	+	require.Equal(t, 1, wake.coalesced)
	35	+	require.Same(t, ac, wake.lease.attachment)
	36	+	requireNoWake(t, h.wakes)
	37	+	}
	38	+
	39	+	// Parking and resuming the same client object is a new incarnation: a failed
	40	+	// Welcome on the replacement transport must not revive readiness from before
	41	+	// the park.
	42	+	func TestRenderCoordinatorParkedResumeRequiresNewWelcome(t *testing.T) {
	43	+	h := newCoordinatorHarness(t)
	44	+	ac := &attachedClient{}
	45	+	h.rc.attachWithReadiness(ac, false)
	46	+	firstLease := h.rc.attachmentLease(ac)
	47	+	require.True(t, h.rc.markAttachmentReady(firstLease))
	48	+	h.rc.notePark(ac)
	49	+	h.rc.attachWithReadiness(ac, false)
	50	+	resumedLease := h.rc.attachmentLease(ac)
	51	+	require.NotNil(t, resumedLease)
	52	+	require.NotSame(t, firstLease, resumedLease)
	53	+	require.False(t, h.rc.markAttachmentReady(firstLease), "stale Welcome must not bless the reused attachment")
	54	+
	55	+	h.rc.invalidate(renderInvalidation{class: invalidateUrgent, reset: true, producer: "session"})
	56	+	timer := awaitCoordinatorScheduledTimer(t, h.clk)
	57	+	h.rc.mu.Lock()
	58	+	workerDone := h.rc.normalLane.token.done
	59	+	h.rc.mu.Unlock()
	60	+	require.NotNil(t, workerDone)
	61	+	timer.ch <- time.Time{}
	62	+	requireWorkerExit(t, workerDone)
	63	+	requireNoWake(t, h.wakes)
	64	+
	65	+	require.True(t, h.rc.markAttachmentReady(resumedLease))
	66	+	h.rc.fireCurrent(false)
	67	+	wake := awaitWake(t, h.wakes)
	68	+	require.True(t, wake.reset)
	69	+	require.Equal(t, 1, wake.coalesced)
	70	+	}

A internal/usecase/daemon/render_coordinator_publication_test.go +28 -0

@@ -0,0 +1,28 @@
	1	+	package daemon
	2	+
	3	+	import (
	4	+	"testing"
	5	+
	6	+	"github.com/stretchr/testify/require"
	7	+	)
	8	+
	9	+	func TestCoordinatorRejectsStaleAttachmentInvalidationAfterReplacement(t *testing.T) {
	10	+	var invalidations []renderInvalidation
	11	+	rc := newRenderCoordinator(renderCoordinatorOptions{
	12	+	onInvalidate: func(inv renderInvalidation) { invalidations = append(invalidations, inv) },
	13	+	})
	14	+	old := &attachedClient{}
	15	+	replacement := &attachedClient{}
	16	+	rc.attach(old)
	17	+	rc.noteReplace(old, replacement)
	18	+
	19	+	// This models a stale resize callback that passed its sendMu generation
	20	+	// check immediately before a replacement was published.
	21	+	rc.invalidateForAttachment(old, renderInvalidation{producer: "stale resize"})
	22	+	require.Empty(t, invalidations)
	23	+
	24	+	// Pane/session producers carry no attachment and remain coordinator-owned.
	25	+	rc.invalidateForAttachment(nil, renderInvalidation{producer: "pane"})
	26	+	require.Len(t, invalidations, 1)
	27	+	require.Equal(t, "pane", invalidations[0].producer)
	28	+	}

A internal/usecase/daemon/render_coordinator_support.go +27 -0

@@ -0,0 +1,27 @@
	1	+	package daemon
	2	+
	3	+	// Small coordinator accessors stay separate from scheduling and lifecycle
	4	+	// transitions, keeping the coordinator state file focused on ownership.
	5	+
	6	+	// burstMetricsSnapshot returns a consistent read-only metric view for
	7	+	// benchmark reporting without exposing coordinator policy outside this package.
	8	+	func (c *renderCoordinator) burstMetricsSnapshot() renderCoordinatorBurstMetricsSnapshot {
	9	+	return renderCoordinatorBurstMetricsSnapshot{
	10	+	invalidations: c.metrics.invalidations.Load(),
	11	+	wakes: c.metrics.wakes.Load(),
	12	+	coalesced: c.metrics.coalesced.Load(),
	13	+	}
	14	+	}
	15	+
	16	+	// renderCoordinator returns the coordinator installed for this session.
	17	+	func (s session) renderCoordinator() renderCoordinator { return s.coordinator.Load() }
	18	+
	19	+	// installRenderCoordinator publishes rc as the session's coordinator.
	20	+	func (s session) installRenderCoordinator(rc renderCoordinator) { s.coordinator.Store(rc) }
	21	+
	22	+	func (c *renderCoordinator) fireCurrent(watchdog bool) {
	23	+	c.mu.Lock()
	24	+	gen := c.normalLane.generation
	25	+	c.mu.Unlock()
	26	+	c.fire(gen, watchdog, watchdog)
	27	+	}

A internal/usecase/daemon/render_coordinator_test.go +1614 -0

@@ -0,0 +1,1614 @@
	1	+	package daemon
	2	+
	3	+	import (
	4	+	"os"
	5	+	"path/filepath"
	6	+	"sync"
	7	+	"sync/atomic"
	8	+	"testing"
	9	+	"time"
	10	+
	11	+	"github.com/stretchr/testify/mock"
	12	+	"github.com/stretchr/testify/require"
	13	+
	14	+	"github.com/bnema/vev/internal/domain"
	15	+	"github.com/bnema/vev/internal/ports"
	16	+	portsmocks "github.com/bnema/vev/internal/ports/mocks"
	17	+	"github.com/bnema/vev/internal/usecase/keys"
	18	+	"github.com/bnema/vev/internal/usecase/layout"
	19	+	)
	20	+
	21	+	// --- coordinator harness ------------------------------------------------------
	22	+
	23	+	// coordinatorMockClock uses generated mocks while retaining deterministic timer
	24	+	// channels and recorded deadlines for coordinator scheduling assertions.
	25	+	type coordinatorMockClock struct {
	26	+	clock *portsmocks.MockClock
	27	+	timers chan *coordinatorMockTimer
	28	+	}
	29	+
	30	+	type coordinatorMockTimer struct {
	31	+	mock *portsmocks.MockTimer
	32	+	ch chan time.Time
	33	+	duration time.Duration
	34	+	}
	35	+
	36	+	func newCoordinatorMockClock(t testing.T, capacity int) coordinatorMockClock {
	37	+	t.Helper()
	38	+	clk := &coordinatorMockClock{
	39	+	clock: portsmocks.NewMockClock(t),
	40	+	timers: make(chan *coordinatorMockTimer, capacity),
	41	+	}
	42	+	clk.clock.EXPECT().Now().Return(time.Time{}).Maybe()
	43	+	clk.clock.EXPECT().NewTimer(mock.MatchedBy(func(d time.Duration) bool {
	44	+	return d == urgentRenderDeadline \|\|
	45	+	(d >= minOutputRenderDeadline && d <= maxOutputRenderDeadline) \|\|
	46	+	d == maxSyncUpdateDuration \|\|
	47	+	d == time.Second \|\|
	48	+	d == 15*time.Minute
	49	+	})).RunAndReturn(func(d time.Duration) ports.Timer {
	50	+	timer := &coordinatorMockTimer{
	51	+	mock: portsmocks.NewMockTimer(t),
	52	+	duration: d,
	53	+	}
	54	+	timer.ch = make(chan time.Time, 1)
	55	+	// A normal deadline reads C both in its worker and in the inert-clock
	56	+	// guard; a watchdog reads it in its worker. The generated expectation
	57	+	// makes every channel access observable without a hand-written Timer.
	58	+	timer.mock.EXPECT().C().Maybe().Return((<-chan time.Time)(timer.ch))
	59	+	timer.mock.EXPECT().Stop().Maybe().Return(true)
	60	+	clk.timers <- timer
	61	+	return timer.mock
	62	+	}).Maybe()
	63	+	return clk
	64	+	}
	65	+
	66	+	// TestRenderCoordinatorTimerCallbacksReenterCoordinator is deliberately a
	67	+	// deadlock regression: every external timer method synchronously reads
	68	+	// coordinator metadata. Before two-phase timer ownership this blocked while
	69	+	// c.mu was held during normal arming, urgent replacement, sync arming, and
	70	+	// detach cancellation.
	71	+	// Detach cancels and stops a selected worker but never synchronously joins it;
	72	+	// the stale worker must be rejected when its external readiness probe returns.
	73	+	func TestRenderCoordinatorDetachDoesNotJoinSelectedDeadlineWorker(t *testing.T) {
	74	+	clock := portsmocks.NewMockClock(t)
	75	+	timer := portsmocks.NewMockTimer(t)
	76	+	timerC := make(chan time.Time, 1)
	77	+	stopCalled := make(chan struct{})
	78	+	clock.EXPECT().NewTimer(minOutputRenderDeadline).Return(timer).Once()
	79	+	timer.EXPECT().C().Return((<-chan time.Time)(timerC)).Once()
	80	+	timer.EXPECT().Stop().Run(func() { close(stopCalled) }).Return(true).Once()
	81	+
	82	+	entered := make(chan struct{})
	83	+	release := make(chan struct{})
	84	+	rc := newRenderCoordinator(renderCoordinatorOptions{
	85	+	clock: clock,
	86	+	ackReady: func() bool {
	87	+	close(entered)
	88	+	<-release
	89	+	return true
	90	+	},
	91	+	})
	92	+	ac := &attachedClient{}
	93	+	rc.attach(ac)
	94	+	rc.invalidate(renderInvalidation{class: invalidateOutput})
	95	+	timerC <- time.Time{}
	96	+	<-entered
	97	+
	98	+	detached := make(chan struct{})
	99	+	go func() {
	100	+	rc.noteDetach(ac)
	101	+	close(detached)
	102	+	}()
	103	+	<-stopCalled
	104	+	awaitTestCompletion(t, detached, "detach synchronously waited for a selected timer worker")
	105	+	close(release)
	106	+	}
	107	+
	108	+	func TestRenderCoordinatorTimerCallbacksReenterCoordinator(t *testing.T) {
	109	+	clock := portsmocks.NewMockClock(t)
	110	+	var rc *renderCoordinator
	111	+	clock.EXPECT().NewTimer(mock.Anything).RunAndReturn(func(time.Duration) ports.Timer {
	112	+	timer := portsmocks.NewMockTimer(t)
	113	+	ch := make(chan time.Time)
	114	+	timer.EXPECT().C().Run(func() {
	115	+	_ = rc.resizeSnapshot()
	116	+	_ = rc.burstMetricsSnapshot()
	117	+	}).Return((<-chan time.Time)(ch)).Once()
	118	+	timer.EXPECT().Stop().Run(func() {
	119	+	_ = rc.resizeSnapshot()
	120	+	_ = rc.burstMetricsSnapshot()
	121	+	}).Return(true).Once()
	122	+	return timer
	123	+	}).Times(3)
	124	+	rc = newRenderCoordinator(renderCoordinatorOptions{clock: clock})
	125	+	ac := &attachedClient{}
	126	+	rc.attach(ac)
	127	+
	128	+	rc.invalidate(renderInvalidation{class: invalidateOutput}) // normal arm
	129	+	rc.invalidate(renderInvalidation{class: invalidateUrgent}) // urgent promotion
	130	+	rc.noteSyncBegin(nil, 1) // sync arm
	131	+	rc.noteDetach(ac) // cancel normal deadline only
	132	+	rc.noteSyncPaneRemoved(nil) // pane destruction cancels sync watchdog
	133	+	}
	134	+
	135	+	// coordinatorHarness wires one coordinator to generated clock/timer mocks and
	136	+	// recording hooks. Every assertion is channel- or counter-based; nothing sleeps.
	137	+	type coordinatorHarness struct {
	138	+	clk *coordinatorMockClock
	139	+	wakes chan renderWake
	140	+	previews chan renderWake
	141	+	ackReady atomic.Bool
	142	+	syncActive atomic.Bool
	143	+	rc *renderCoordinator
	144	+	}
	145	+
	146	+	func newCoordinatorHarness(t testing.T) coordinatorHarness {
	147	+	t.Helper()
	148	+	h := &coordinatorHarness{
	149	+	clk: newCoordinatorMockClock(t, 16),
	150	+	wakes: make(chan renderWake, 16),
	151	+	previews: make(chan renderWake, 16),
	152	+	}
	153	+	h.ackReady.Store(true)
	154	+	h.rc = newRenderCoordinator(renderCoordinatorOptions{
	155	+	clock: h.clk.clock,
	156	+	wake: func(w renderWake) { h.wakes <- w },
	157	+	ackReady: func() bool { return h.ackReady.Load() },
	158	+	syncActive: func() bool { return h.syncActive.Load() },
	159	+	})
	160	+	return h
	161	+	}
	162	+
	163	+	// armedTimers drains every deadline armed so far. Arming happens
	164	+	// synchronously inside the coordinator call, so a non-blocking drain is
	165	+	// deterministic.
	166	+	func (h coordinatorHarness) armedTimers(t testing.T) []*coordinatorMockTimer {
	167	+	t.Helper()
	168	+	var timers []*coordinatorMockTimer
	169	+	for {
	170	+	select {
	171	+	case tm := <-h.clk.timers:
	172	+	timers = append(timers, tm)
	173	+	default:
	174	+	return timers
	175	+	}
	176	+	}
	177	+	}
	178	+
	179	+	// Coordinator calls and fake-clock advancement are synchronous test steps.
	180	+	// These helpers deliberately never wait on wall time: a missing event is a
	181	+	// deterministic contract failure, not a slow behavior to poll for.
	182	+	func awaitWake(t *testing.T, ch chan renderWake) renderWake {
	183	+	t.Helper()
	184	+	return awaitTestValue(t, ch, "coordinator did not publish a wake after fake-clock advancement")
	185	+	}
	186	+
	187	+	func requireNoWake(t *testing.T, ch chan renderWake) {
	188	+	t.Helper()
	189	+	select {
	190	+	case w := <-ch:
	191	+	t.Fatalf("unexpected coordinator wake: %+v", w)
	192	+	default:
	193	+	}
	194	+	}
	195	+
	196	+	func awaitInvalidation(t *testing.T, ch chan renderInvalidation) renderInvalidation {
	197	+	t.Helper()
	198	+	return awaitTestValue(t, ch, "producer did not publish a coordinator invalidation")
	199	+	}
	200	+
	201	+	func requireNoInvalidation(t *testing.T, ch chan renderInvalidation) {
	202	+	t.Helper()
	203	+	select {
	204	+	case inv := <-ch:
	205	+	t.Fatalf("unexpected coordinator invalidation: %+v", inv)
	206	+	default:
	207	+	}
	208	+	}
	209	+
	210	+	func captureResizeCallbackDone(t testing.T, rc renderCoordinator) <-chan struct{} {
	211	+	t.Helper()
	212	+	rc.mu.Lock()
	213	+	defer rc.mu.Unlock()
	214	+	if rc.resizeLane.token == nil {
	215	+	t.Fatal("coordinator did not publish a resize callback completion")
	216	+	}
	217	+	return rc.resizeLane.token.done
	218	+	}
	219	+
	220	+	func awaitCoordinatorScheduledTimer(t testing.T, clk coordinatorMockClock) *coordinatorMockTimer {
	221	+	t.Helper()
	222	+	select {
	223	+	case tm := <-clk.timers:
	224	+	return tm
	225	+	default:
	226	+	t.Fatal("coordinator did not synchronously arm a fake-clock timer")
	227	+	return nil
	228	+	}
	229	+	}
	230	+
	231	+	func awaitLatestCoordinatorTimer(t testing.T, clk coordinatorMockClock) *coordinatorMockTimer {
	232	+	t.Helper()
	233	+	timer := awaitCoordinatorScheduledTimer(t, clk)
	234	+	for {
	235	+	select {
	236	+	case timer = <-clk.timers:
	237	+	default:
	238	+	return timer
	239	+	}
	240	+	}
	241	+	}
	242	+
	243	+	func requireNoCoordinatorOutputFrame(t *testing.T, sends chan ports.Frame) {
	244	+	t.Helper()
	245	+	select {
	246	+	case frame := <-sends:
	247	+	t.Fatalf("unexpected output frame: %+v", frame)
	248	+	default:
	249	+	}
	250	+	}
	251	+
	252	+	// --- wake coalescing and deadlines --------------------------------------------
	253	+
	254	+	func TestRenderCoordinatorCoalescesWakesUnderOneDeadline(t *testing.T) {
	255	+	cases := []struct {
	256	+	name string
	257	+	invalidations []renderInvalidation
	258	+	wantTimers int
	259	+	wantMin time.Duration
	260	+	wantMax time.Duration
	261	+	wantWake renderWake
	262	+	}{
	263	+	{
	264	+	name: "urgent transition wakes within two milliseconds",
	265	+	invalidations: []renderInvalidation{{class: invalidateUrgent, reset: true, producer: "input.go"}},
	266	+	wantTimers: 1,
	267	+	wantMax: urgentRenderDeadline,
	268	+	wantWake: renderWake{reset: true, urgent: true, coalesced: 1},
	269	+	},
	270	+	{
	271	+	name: "output burst coalesces into one adaptive deadline",
	272	+	invalidations: []renderInvalidation{
	273	+	{class: invalidateOutput, producer: "render.go"},
	274	+	{class: invalidateOutput, producer: "render.go"},
	275	+	{class: invalidateOutput, producer: "render.go"},
	276	+	},
	277	+	wantTimers: 1,
	278	+	wantMin: minOutputRenderDeadline,
	279	+	wantMax: maxOutputRenderDeadline,
	280	+	wantWake: renderWake{coalesced: 3},
	281	+	},
	282	+	{
	283	+	name: "full redraw stays sticky across later incremental transitions",
	284	+	invalidations: []renderInvalidation{
	285	+	{class: invalidateOutput},
	286	+	{class: invalidateOutput, reset: true},
	287	+	{class: invalidateOutput},
	288	+	},
	289	+	wantTimers: 1,
	290	+	wantMin: minOutputRenderDeadline,
	291	+	wantMax: maxOutputRenderDeadline,
	292	+	wantWake: renderWake{reset: true, coalesced: 3},
	293	+	},
	294	+	{
	295	+	name: "urgent transition tightens a pending output deadline",
	296	+	invalidations: []renderInvalidation{
	297	+	{class: invalidateOutput},
	298	+	{class: invalidateUrgent},
	299	+	},
	300	+	wantTimers: 2,
	301	+	wantMax: urgentRenderDeadline,
	302	+	wantWake: renderWake{urgent: true, coalesced: 2},
	303	+	},
	304	+	}
	305	+	for _, tc := range cases {
	306	+	t.Run(tc.name, func(t *testing.T) {
	307	+	h := newCoordinatorHarness(t)
	308	+	for _, inv := range tc.invalidations {
	309	+	h.rc.invalidate(inv)
	310	+	}
	311	+	timers := h.armedTimers(t)
	312	+	require.Lenf(t, timers, tc.wantTimers,
	313	+	"cap-1 coalescing must arm exactly %d deadline timer(s) for %d invalidations",
	314	+	tc.wantTimers, len(tc.invalidations))
	315	+	last := timers[len(timers)-1]
	316	+	require.GreaterOrEqual(t, last.duration, tc.wantMin)
	317	+	require.LessOrEqual(t, last.duration, tc.wantMax)
	318	+
	319	+	last.ch <- time.Time{}
	320	+	require.Equal(t, tc.wantWake, awaitWake(t, h.wakes),
	321	+	"one wake must deliver the latest coalesced state")
	322	+	requireNoWake(t, h.wakes)
	323	+	})
	324	+	}
	325	+	}
	326	+
	327	+	func TestRenderCoordinatorAdaptsOutputDeadlineAfterBurstAndDecays(t *testing.T) {
	328	+	h := newCoordinatorHarness(t)
	329	+
	330	+	// A burst establishes pressure without extending its already-armed first
	331	+	// deadline; the following output deadline grows to the 16ms ceiling.
	332	+	for range 9 {
	333	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput, producer: "render.go"})
	334	+	}
	335	+	burst := h.armedTimers(t)
	336	+	require.Len(t, burst, 1)
	337	+	require.Equal(t, minOutputRenderDeadline, burst[0].duration)
	338	+	burst[0].ch <- time.Time{}
	339	+	awaitWake(t, h.wakes)
	340	+
	341	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput, producer: "render.go"})
	342	+	atCap := h.armedTimers(t)
	343	+	require.Len(t, atCap, 1)
	344	+	require.Equal(t, maxOutputRenderDeadline, atCap[0].duration,
	345	+	"burst pressure must be capped at the 16ms output deadline")
	346	+	atCap[0].ch <- time.Time{}
	347	+	awaitWake(t, h.wakes)
	348	+
	349	+	// Quiet singleton batches decay pressure toward the 8ms floor.
	350	+	var previous = maxOutputRenderDeadline
	351	+	for range 8 {
	352	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput, producer: "render.go"})
	353	+	timers := h.armedTimers(t)
	354	+	require.Len(t, timers, 1)
	355	+	require.LessOrEqual(t, timers[0].duration, previous)
	356	+	require.GreaterOrEqual(t, timers[0].duration, minOutputRenderDeadline)
	357	+	previous = timers[0].duration
	358	+	timers[0].ch <- time.Time{}
	359	+	awaitWake(t, h.wakes)
	360	+	}
	361	+	require.Equal(t, minOutputRenderDeadline, previous)
	362	+	}
	363	+
	364	+	func TestRenderCoordinatorUrgentDeadlineCannotBeExtended(t *testing.T) {
	365	+	h := newCoordinatorHarness(t)
	366	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput})
	367	+	h.rc.invalidate(renderInvalidation{class: invalidateUrgent})
	368	+	for range 32 {
	369	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput})
	370	+	h.rc.invalidate(renderInvalidation{class: invalidateUrgent})
	371	+	}
	372	+	timers := h.armedTimers(t)
	373	+	require.Len(t, timers, 2, "only the urgent promotion may replace the output timer")
	374	+	require.Equal(t, urgentRenderDeadline, timers[len(timers)-1].duration)
	375	+	}
	376	+
	377	+	// --- ACK gating ----------------------------------------------------------------
	378	+
	379	+	func TestRenderCoordinatorAckReadinessReentersWithoutBlockingResize(t *testing.T) {
	380	+	owner := &attachedClient{}
	381	+	wakes := make(chan renderWake, 1)
	382	+	ackEntered := make(chan struct{})
	383	+	resizeDone := make(chan uint64, 1)
	384	+	fireDone := make(chan struct{})
	385	+
	386	+	var rc *renderCoordinator
	387	+	rc = newRenderCoordinator(renderCoordinatorOptions{
	388	+	ackReady: func() bool {
	389	+	close(ackEntered)
	390	+	// The readiness probe models the output send path reading coordinator
	391	+	// metadata after sendMu is held. It must never run under c.mu.
	392	+	_ = rc.resizeSnapshot()
	393	+	return true
	394	+	},
	395	+	wake: func(w renderWake) { wakes <- w },
	396	+	})
	397	+	rc.attach(owner)
	398	+	rc.invalidate(renderInvalidation{class: invalidateOutput, producer: "render.go"})
	399	+
	400	+	go func() {
	401	+	rc.fire(1, false, true)
	402	+	close(fireDone)
	403	+	}()
	404	+	awaitTestCompletion(t, ackEntered, "ack readiness was not evaluated")
	405	+
	406	+	go func() { resizeDone <- rc.recordResizeRequest(domain.Size{Cols: 120, Rows: 40}, owner) }()
	407	+	epoch := awaitTestValue(t, resizeDone, "ack readiness re-entry blocked resize metadata")
	408	+	require.Equal(t, uint64(1), epoch)
	409	+	wake := awaitWake(t, wakes)
	410	+	require.Same(t, owner, wake.lease.attachment)
	411	+	wake.lease = nil
	412	+	require.Equal(t, renderWake{coalesced: 1}, wake)
	413	+	awaitTestCompletion(t, fireDone, "fire did not return after wake")
	414	+	}
	415	+
	416	+	func TestRenderCoordinatorAckDoesNotBypassAnUnexpiredDeadline(t *testing.T) {
	417	+	h := newCoordinatorHarness(t)
	418	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput, producer: "render.go"})
	419	+	timer := awaitCoordinatorScheduledTimer(t, h.clk)
	420	+	require.GreaterOrEqual(t, timer.duration, minOutputRenderDeadline)
	421	+
	422	+	// An unrelated ACK must not turn a normal output deadline into an
	423	+	// immediate wake or cancel its timer.
	424	+	h.rc.notifyAck()
	425	+	requireNoWake(t, h.wakes)
	426	+	timer.mock.AssertNotCalled(t, "Stop")
	427	+
	428	+	timer.ch <- time.Time{}
	429	+	awaitWake(t, h.wakes)
	430	+	requireNoWake(t, h.wakes)
	431	+	}
	432	+
	433	+	func TestRenderCoordinatorAckFlushesOnlyExpiredAckDeferredWork(t *testing.T) {
	434	+	t.Run("expired deadline flushes exactly once after readiness", func(t *testing.T) {
	435	+	h := newCoordinatorHarness(t)
	436	+	h.ackReady.Store(false)
	437	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput, reset: true, producer: "render.go"})
	438	+	timer := awaitCoordinatorScheduledTimer(t, h.clk)
	439	+	timer.ch <- time.Time{}
	440	+	requireNoWake(t, h.wakes)
	441	+
	442	+	// Repeated ACKs while capacity remains unavailable are no-ops.
	443	+	h.rc.notifyAck()
	444	+	requireNoWake(t, h.wakes)
	445	+	h.ackReady.Store(true)
	446	+	h.rc.notifyAck()
	447	+	w := awaitWake(t, h.wakes)
	448	+	require.True(t, w.reset)
	449	+	requireNoWake(t, h.wakes)
	450	+	h.rc.notifyAck()
	451	+	requireNoWake(t, h.wakes)
	452	+	})
	453	+
	454	+	t.Run("lifecycle clears deferred work and urgent explicit fires stay immediate", func(t *testing.T) {
	455	+	h := newCoordinatorHarness(t)
	456	+	h.ackReady.Store(false)
	457	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput})
	458	+	timer := awaitCoordinatorScheduledTimer(t, h.clk)
	459	+	timer.ch <- time.Time{}
	460	+	requireNoWake(t, h.wakes)
	461	+	h.rc.beginSessionTeardown().finish()
	462	+	h.ackReady.Store(true)
	463	+	h.rc.notifyAck()
	464	+	requireNoWake(t, h.wakes)
	465	+
	466	+	immediate := newCoordinatorHarness(t)
	467	+	immediate.rc.invalidate(renderInvalidation{class: invalidateUrgent})
	468	+	immediate.rc.fireCurrent(false)
	469	+	w := awaitWake(t, immediate.wakes)
	470	+	require.True(t, w.urgent)
	471	+	})
	472	+	}
	473	+
	474	+	func TestRenderCoordinatorAckGateBlocksCompositionUntilAck(t *testing.T) {
	475	+	h := newCoordinatorHarness(t)
	476	+	h.ackReady.Store(false)
	477	+
	478	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput})
	479	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput, reset: true})
	480	+	h.rc.invalidate(renderInvalidation{class: invalidateUrgent})
	481	+	for _, timer := range h.armedTimers(t) {
	482	+	timer.ch <- time.Time{}
	483	+	}
	484	+	requireNoWake(t, h.wakes)
	485	+
	486	+	h.ackReady.Store(true)
	487	+	h.rc.notifyAck()
	488	+	w := awaitWake(t, h.wakes)
	489	+	require.True(t, w.reset, "the deferred full redraw must stay sticky through the ack gate")
	490	+	require.Equal(t, 3, w.coalesced, "the post-ack wake must carry every deferred transition")
	491	+	requireNoWake(t, h.wakes)
	492	+
	493	+	h.rc.notifyAck()
	494	+	requireNoWake(t, h.wakes)
	495	+	}
	496	+
	497	+	// --- synchronized output ---------------------------------------------------------
	498	+
	499	+	func TestRenderCoordinatorConcurrentSynchronizedPaneBatches(t *testing.T) {
	500	+	t.Run("end orders, watchdogs, and stale callbacks are isolated per pane", func(t *testing.T) {
	501	+	h := newCoordinatorHarness(t)
	502	+	firstForced, secondForced := make(chan struct{}, 1), make(chan struct{}, 1)
	503	+	firstPane, secondPane := &pane{}, &pane{}
	504	+
	505	+	// Distinct stable pane owners must remain independent.
	506	+	h.rc.noteSyncBegin(firstPane, 11, func() { firstForced <- struct{}{} })
	507	+	firstWatchdog := h.armedTimers(t)[0]
	508	+	h.rc.noteSyncBegin(secondPane, 22, func() { secondForced <- struct{}{} })
	509	+	secondWatchdog := h.armedTimers(t)[0]
	510	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput, producer: "render.go"})
	511	+
	512	+	// Completing either pane leaves the aggregate gate closed.
	513	+	h.rc.noteSyncEnd(firstPane, 11)
	514	+	requireNoWake(t, h.wakes)
	515	+	// A stale callback for the completed pane must not force it or wake.
	516	+	firstWatchdog.ch <- time.Time{}
	517	+	select {
	518	+	case <-firstForced:
	519	+	t.Fatal("stale watchdog forced a completed pane")
	520	+	default:
	521	+	}
	522	+	requireNoWake(t, h.wakes)
	523	+
	524	+	// The remaining pane's watchdog forces only its own batch, then the
	525	+	// aggregate completion produces exactly one urgent wake.
	526	+	secondWatchdog.ch <- time.Time{}
	527	+	<-secondForced
	528	+	w := awaitWake(t, h.wakes)
	529	+	require.True(t, w.watchdog)
	530	+	require.True(t, w.urgent, "aggregate completion must wake urgently")
	531	+	requireNoWake(t, h.wakes)
	532	+	})
	533	+
	534	+	t.Run("later pane can end first without releasing an earlier pane", func(t *testing.T) {
	535	+	h := newCoordinatorHarness(t)
	536	+	firstForced := make(chan struct{}, 1)
	537	+	firstPane, secondPane := &pane{}, &pane{}
	538	+	h.rc.noteSyncBegin(firstPane, 11, func() { firstForced <- struct{}{} })
	539	+	firstWatchdog := h.armedTimers(t)[0]
	540	+	h.rc.noteSyncBegin(secondPane, 22)
	541	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput, producer: "render.go"})
	542	+
	543	+	h.rc.noteSyncEnd(secondPane, 22)
	544	+	requireNoWake(t, h.wakes)
	545	+	firstWatchdog.ch <- time.Time{}
	546	+	<-firstForced
	547	+	w := awaitWake(t, h.wakes)
	548	+	require.True(t, w.watchdog)
	549	+	require.True(t, w.urgent, "final pane completion must wake urgently")
	550	+	requireNoWake(t, h.wakes)
	551	+	})
	552	+
	553	+	t.Run("pane removal cancels only that pane watchdog", func(t *testing.T) {
	554	+	h := newCoordinatorHarness(t)
	555	+	firstPane, secondPane := &pane{}, &pane{}
	556	+	firstForced, secondForced := make(chan struct{}, 1), make(chan struct{}, 1)
	557	+	h.rc.noteSyncBegin(firstPane, 11, func() { firstForced <- struct{}{} })
	558	+	firstWatchdog := h.armedTimers(t)[0]
	559	+	h.rc.noteSyncBegin(secondPane, 22, func() { secondForced <- struct{}{} })
	560	+	secondWatchdog := h.armedTimers(t)[0]
	561	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput, producer: "render.go"})
	562	+
	563	+	h.rc.noteSyncPaneRemoved(firstPane)
	564	+	firstWatchdog.ch <- time.Time{}
	565	+	select {
	566	+	case <-firstForced:
	567	+	t.Fatal("removed pane watchdog forced output")
	568	+	default:
	569	+	}
	570	+	requireNoWake(t, h.wakes)
	571	+	secondWatchdog.ch <- time.Time{}
	572	+	<-secondForced
	573	+	require.True(t, awaitWake(t, h.wakes).watchdog)
	574	+	})
	575	+	}
	576	+
	577	+	func TestRenderCoordinatorSyncGateFollowsDynamicRenderability(t *testing.T) {
	578	+	t.Run("inactive batch that becomes active gates activation output until end", func(t *testing.T) {
	579	+	h := newCoordinatorHarness(t)
	580	+	var visible atomic.Bool
	581	+	p := &pane{}
	582	+	h.rc.noteSyncBeginWithRenderability(p, 1, visible.Load)
	583	+	visible.Store(true)
	584	+	h.rc.invalidate(renderInvalidation{class: invalidateUrgent, producer: "activation"})
	585	+	for _, timer := range h.armedTimers(t) {
	586	+	timer.ch <- time.Time{}
	587	+	}
	588	+	requireNoWake(t, h.wakes)
	589	+
	590	+	h.rc.noteSyncEnd(p, 1)
	591	+	w := awaitWake(t, h.wakes)
	592	+	require.True(t, w.urgent)
	593	+	requireNoWake(t, h.wakes)
	594	+	})
	595	+
	596	+	t.Run("hidden batch does not stall newly active output or wake on late end", func(t *testing.T) {
	597	+	h := newCoordinatorHarness(t)
	598	+	var visible atomic.Bool
	599	+	visible.Store(true)
	600	+	p := &pane{}
	601	+	h.rc.noteSyncBeginWithRenderability(p, 1, visible.Load)
	602	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput, producer: "batch"})
	603	+	visible.Store(false)
	604	+	h.rc.invalidate(renderInvalidation{class: invalidateUrgent, producer: "new-active-tab"})
	605	+	timers := h.armedTimers(t)
	606	+	require.NotEmpty(t, timers)
	607	+	timers[len(timers)-1].ch <- time.Time{}
	608	+	w := awaitWake(t, h.wakes)
	609	+	require.False(t, w.watchdog)
	610	+	h.rc.noteSyncEnd(p, 1)
	611	+	requireNoWake(t, h.wakes)
	612	+	})
	613	+	}
	614	+
	615	+	func TestRenderCoordinatorSyncGateRetriesWhenBatchPublishesDuringPredicate(t *testing.T) {
	616	+	h := newCoordinatorHarness(t)
	617	+	first, second := &pane{}, &pane{}
	618	+	gateEvaluated := make(chan struct{})
	619	+	releaseGate := make(chan struct{})
	620	+	var hookOnce sync.Once
	621	+	h.rc.opts.afterSyncGateEvaluated = func() {
	622	+	hookOnce.Do(func() {
	623	+	close(gateEvaluated)
	624	+	<-releaseGate
	625	+	})
	626	+	}
	627	+	h.rc.noteSyncBeginWithRenderability(first, 1, func() bool { return false })
	628	+	h.rc.invalidate(renderInvalidation{class: invalidateUrgent, producer: "test"})
	629	+
	630	+	fireDone := make(chan struct{})
	631	+	go func() {
	632	+	h.rc.fireCurrent(false)
	633	+	close(fireDone)
	634	+	}()
	635	+	<-gateEvaluated
	636	+	// This publication is deliberately between an open predicate result and
	637	+	// pending consumption. A stale gate result must be rejected and retried.
	638	+	h.rc.noteSyncBeginWithRenderability(second, 2, func() bool { return true })
	639	+	close(releaseGate)
	640	+	awaitTestCompletion(t, fireDone, "fire did not finish after retrying the registry snapshot")
	641	+
	642	+	requireNoWake(t, h.wakes)
	643	+	h.rc.noteSyncEnd(second, 2)
	644	+	require.Equal(t, renderWake{urgent: true, coalesced: 1}, awaitWake(t, h.wakes))
	645	+	}
	646	+
	647	+	func TestRenderCoordinatorSynchronizedOutput(t *testing.T) {
	648	+	t.Run("completion flushes pending state in one wake", func(t *testing.T) {
	649	+	h := newCoordinatorHarness(t)
	650	+	h.syncActive.Store(true)
	651	+	h.rc.noteSyncBegin(nil, 1)
	652	+	watchdogs := h.armedTimers(t)
	653	+	require.NotEmpty(t, watchdogs, "sync begin must arm the completion watchdog")
	654	+	require.Equal(t, maxSyncUpdateDuration, watchdogs[len(watchdogs)-1].duration)
	655	+
	656	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput, reset: true})
	657	+	timers := h.armedTimers(t)
	658	+	require.Len(t, timers, 1)
	659	+	h.rc.mu.Lock()
	660	+	normalWorkerDone := h.rc.normalLane.token.done
	661	+	h.rc.mu.Unlock()
	662	+	require.NotNil(t, normalWorkerDone)
	663	+	timers[0].ch <- time.Time{}
	664	+	requireWorkerExit(t, normalWorkerDone)
	665	+	requireNoWake(t, h.wakes)
	666	+
	667	+	h.syncActive.Store(false)
	668	+	h.rc.noteSyncEnd(nil, 1)
	669	+	w := awaitWake(t, h.wakes)
	670	+	require.True(t, w.reset)
	671	+	requireNoWake(t, h.wakes)
	672	+	})
	673	+
	674	+	t.Run("watchdog ends a wedged batch before it flushes", func(t *testing.T) {
	675	+	h := newCoordinatorHarness(t)
	676	+	h.syncActive.Store(true)
	677	+	forced := make(chan struct{}, 1)
	678	+	h.rc.noteSyncBegin(nil, 7, func() { forced <- struct{}{} })
	679	+	watchdogs := h.armedTimers(t)
	680	+	require.NotEmpty(t, watchdogs, "sync begin must arm the completion watchdog")
	681	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput})
	682	+
	683	+	watchdogs[0].ch <- time.Time{}
	684	+	<-forced
	685	+	w := awaitWake(t, h.wakes)
	686	+	require.True(t, w.watchdog, "a wedged synchronized batch must be force-flushed by the watchdog")
	687	+	requireNoWake(t, h.wakes)
	688	+	})
	689	+
	690	+	t.Run("watchdog retains gate while force closes the VT batch", func(t *testing.T) {
	691	+	h := newCoordinatorHarness(t)
	692	+	entered := make(chan struct{})
	693	+	release := make(chan struct{})
	694	+	h.rc.noteSyncBegin(nil, 1, func() {
	695	+	close(entered)
	696	+	<-release
	697	+	})
	698	+	watchdogs := h.armedTimers(t)
	699	+	require.NotEmpty(t, watchdogs)
	700	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput})
	701	+	watchdogs[0].ch <- time.Time{}
	702	+	<-entered
	703	+
	704	+	// A concurrent deadline/fire sees the retained batch gate and cannot
	705	+	// publish the partial VT state while force is blocked.
	706	+	h.rc.fireCurrent(false)
	707	+	requireNoWake(t, h.wakes)
	708	+	close(release)
	709	+	w := awaitWake(t, h.wakes)
	710	+	require.True(t, w.watchdog)
	711	+	require.True(t, w.urgent)
	712	+	requireNoWake(t, h.wakes)
	713	+	})
	714	+
	715	+	t.Run("watchdog flushes after the snapshotted generation is replaced", func(t *testing.T) {
	716	+	h := newCoordinatorHarness(t)
	717	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput})
	718	+	h.rc.mu.Lock()
	719	+	staleGeneration := h.rc.normalLane.generation
	720	+	h.rc.mu.Unlock()
	721	+
	722	+	// Simulate a new urgent publication after fireCurrent(true) snapshots its
	723	+	// generation but before fire validates it.
	724	+	h.rc.invalidate(renderInvalidation{class: invalidateUrgent})
	725	+	h.rc.fire(staleGeneration, true, true)
	726	+
	727	+	w := awaitWake(t, h.wakes)
	728	+	require.True(t, w.watchdog)
	729	+	require.True(t, w.urgent)
	730	+	require.Equal(t, 2, w.coalesced)
	731	+	requireNoWake(t, h.wakes)
	732	+	})
	733	+
	734	+	t.Run("stale watchdog generation cannot wake a completed batch", func(t *testing.T) {
	735	+	h := newCoordinatorHarness(t)
	736	+	h.syncActive.Store(true)
	737	+	h.rc.noteSyncBegin(nil, 1)
	738	+	watchdogs := h.armedTimers(t)
	739	+	require.NotEmpty(t, watchdogs, "sync begin must arm the completion watchdog")
	740	+	h.syncActive.Store(false)
	741	+	h.rc.noteSyncEnd(nil, 1)
	742	+
	743	+	watchdogs[0].ch <- time.Time{}
	744	+	requireNoWake(t, h.wakes)
	745	+	})
	746	+	}
	747	+
	748	+	// --- preview subscription --------------------------------------------------------
	749	+
	750	+	func TestRenderCoordinatorPreviewSubscription(t *testing.T) {
	751	+	h := newCoordinatorHarness(t)
	752	+	h.rc.subscribePreview(func(w renderWake) { h.previews <- w })
	753	+
	754	+	h.rc.invalidate(renderInvalidation{class: invalidateUrgent})
	755	+	timers := h.armedTimers(t)
	756	+	require.NotEmpty(t, timers, "invalidate must arm a deadline timer")
	757	+	timers[len(timers)-1].ch <- time.Time{}
	758	+	owner := awaitWake(t, h.wakes)
	759	+	preview := awaitWake(t, h.previews)
	760	+	require.Equal(t, owner, preview, "a subscribed preview observes the same coalesced wake")
	761	+	requireNoWake(t, h.previews)
	762	+
	763	+	h.rc.teardownPreview()
	764	+	h.rc.invalidate(renderInvalidation{class: invalidateUrgent})
	765	+	timers = h.armedTimers(t)
	766	+	require.NotEmpty(t, timers, "invalidate must arm a deadline timer")
	767	+	timers[len(timers)-1].ch <- time.Time{}
	768	+	awaitWake(t, h.wakes)
	769	+	requireNoWake(t, h.previews)
	770	+	}
	771	+
	772	+	// --- lifecycle and stale callbacks -----------------------------------------------
	773	+
	774	+	func TestRenderCoordinatorPreviewWakesDoNotWaitForTargetAck(t *testing.T) {
	775	+	cases := []struct {
	776	+	name string
	777	+	attach bool
	778	+	}{
	779	+	{name: "attached cross-session target blocked on ack", attach: true},
	780	+	{name: "headless cross-session target blocked on ack"},
	781	+	}
	782	+	for _, tc := range cases {
	783	+	t.Run(tc.name, func(t *testing.T) {
	784	+	h := newCoordinatorHarness(t)
	785	+	h.ackReady.Store(false)
	786	+	if tc.attach {
	787	+	h.rc.attach(&attachedClient{})
	788	+	}
	789	+	viewer := &attachedClient{}
	790	+	h.rc.subscribePreviewFor(viewer, 1, func(w renderWake) { h.previews <- w })
	791	+
	792	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput, producer: "target output"})
	793	+	timers := h.armedTimers(t)
	794	+	require.Len(t, timers, 1)
	795	+	timers[0].ch <- time.Time{}
	796	+
	797	+	preview := awaitWake(t, h.previews)
	798	+	if tc.attach {
	799	+	require.NotNil(t, preview.lease.attachment)
	800	+	}
	801	+	preview.lease = nil
	802	+	require.Equal(t, renderWake{coalesced: 1}, preview,
	803	+	"the viewer preview must receive target output without target ACK capacity")
	804	+	requireNoWake(t, h.wakes)
	805	+	requireNoWake(t, h.previews)
	806	+
	807	+	h.ackReady.Store(true)
	808	+	h.rc.notifyAck()
	809	+	wake := awaitWake(t, h.wakes)
	810	+	if tc.attach {
	811	+	require.NotNil(t, wake.lease.attachment)
	812	+	}
	813	+	wake.lease = nil
	814	+	require.Equal(t, renderWake{coalesced: 1}, wake,
	815	+	"the target primary frame remains pending for its own ACK")
	816	+	requireNoWake(t, h.previews)
	817	+	})
	818	+	}
	819	+	}
	820	+
	821	+	func TestRenderCoordinatorDetachedTargetPublishesPreviewOnlyInvalidations(t *testing.T) {
	822	+	h := newCoordinatorHarness(t)
	823	+	target, viewer := &attachedClient{}, &attachedClient{}
	824	+	h.ackReady.Store(false)
	825	+	h.rc.attach(target)
	826	+	h.rc.subscribePreviewFor(viewer, 1, func(w renderWake) { h.previews <- w })
	827	+	h.rc.noteDetach(target)
	828	+
	829	+	// A detached target only accepts PTY/session-owned invalidations. The old
	830	+	// attachment remains stale and must not revive its render path.
	831	+	h.rc.invalidateForAttachment(target, renderInvalidation{class: invalidateOutput, producer: "stale attachment"})
	832	+	require.Empty(t, h.armedTimers(t))
	833	+
	834	+	for _, producer := range []string{"first PTY output", "second PTY output"} {
	835	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput, producer: producer})
	836	+	timers := h.armedTimers(t)
	837	+	require.Len(t, timers, 1)
	838	+	timers[0].ch <- time.Time{}
	839	+	preview := awaitWake(t, h.previews)
	840	+	require.Nil(t, preview.lease, "headless preview wakes must not retain a revoked primary lease")
	841	+	require.Equal(t, renderWake{coalesced: 1}, preview)
	842	+	requireNoWake(t, h.wakes)
	843	+
	844	+	h.rc.mu.Lock()
	845	+	require.False(t, h.rc.pending)
	846	+	require.False(t, h.rc.pendingPreview)
	847	+	require.False(t, h.rc.ackDeferred)
	848	+	require.False(t, h.rc.deadlineDue)
	849	+	require.Zero(t, h.rc.coalesced)
	850	+	require.True(t, h.rc.primaryDetachedLocked())
	851	+	require.NotNil(t, h.rc.lease)
	852	+	require.False(t, h.rc.lease.active)
	853	+	h.rc.mu.Unlock()
	854	+	}
	855	+
	856	+	// A later attach retains the normal first-paint reset path.
	857	+	replacement := &attachedClient{}
	858	+	h.ackReady.Store(true)
	859	+	h.rc.attach(replacement)
	860	+	h.rc.invalidateForAttachment(replacement, renderInvalidation{class: invalidateUrgent, reset: true, producer: "first paint"})
	861	+	timers := h.armedTimers(t)
	862	+	require.Len(t, timers, 1)
	863	+	timers[0].ch <- time.Time{}
	864	+	wake := awaitWake(t, h.wakes)
	865	+	require.True(t, wake.reset)
	866	+	require.Same(t, replacement, wake.lease.attachment)
	867	+	}
	868	+
	869	+	func TestRenderCoordinatorPreviewLifecycleDropsStaleTargetWakes(t *testing.T) {
	870	+	cases := []struct {
	871	+	name string
	872	+	transition func(renderCoordinator, attachedClient, *attachedClient)
	873	+	}{
	874	+	{"target detach", func(rc renderCoordinator, target, _ attachedClient) { rc.noteDetach(target) }},
	875	+	{"target replacement", func(rc renderCoordinator, target, replacement attachedClient) { rc.noteReplace(target, replacement) }},
	876	+	{"target teardown", func(rc renderCoordinator, _ attachedClient, _ *attachedClient) { rc.beginSessionTeardown().finish() }},
	877	+	}
	878	+	for _, tc := range cases {
	879	+	t.Run(tc.name, func(t *testing.T) {
	880	+	h := newCoordinatorHarness(t)
	881	+	target, replacement, viewer := &attachedClient{}, &attachedClient{}, &attachedClient{}
	882	+	h.rc.attach(target)
	883	+	h.rc.subscribePreviewFor(viewer, 1, func(w renderWake) { h.previews <- w })
	884	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput, producer: "target output"})
	885	+	stale := h.armedTimers(t)
	886	+	require.Len(t, stale, 1)
	887	+
	888	+	tc.transition(h.rc, target, replacement)
	889	+	stale[0].ch <- time.Time{}
	890	+	requireNoWake(t, h.wakes)
	891	+	requireNoWake(t, h.previews)
	892	+	})
	893	+	}
	894	+	}
	895	+
	896	+	func TestRenderCoordinatorPreviewSubscriptionsAreIndependent(t *testing.T) {
	897	+	h := newCoordinatorHarness(t)
	898	+	one, two := &attachedClient{}, &attachedClient{}
	899	+	first, second := make(chan renderWake, 1), make(chan renderWake, 1)
	900	+	h.rc.subscribePreviewFor(one, 1, func(w renderWake) { first <- w })
	901	+	h.rc.subscribePreviewFor(two, 1, func(w renderWake) { second <- w })
	902	+	h.rc.attach(&attachedClient{})
	903	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput, producer: "preview"})
	904	+	last := h.armedTimers(t)
	905	+	last[len(last)-1].ch <- time.Time{}
	906	+	awaitWake(t, first)
	907	+	awaitWake(t, second)
	908	+
	909	+	h.rc.teardownPreviewFor(one, 1)
	910	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput, producer: "preview"})
	911	+	last = h.armedTimers(t)
	912	+	last[len(last)-1].ch <- time.Time{}
	913	+	requireNoWake(t, first)
	914	+	awaitWake(t, second)
	915	+	}
	916	+
	917	+	func TestRenderCoordinatorLifecycleDropsStaleWakes(t *testing.T) {
	918	+	cases := []struct {
	919	+	name string
	920	+	teardown func(rc renderCoordinator, owner attachedClient)
	921	+	}{
	922	+	{"detach", func(rc renderCoordinator, owner attachedClient) { rc.noteDetach(owner) }},
	923	+	{"park", func(rc renderCoordinator, owner attachedClient) { rc.notePark(owner) }},
	924	+	{"session teardown", func(rc renderCoordinator, _ attachedClient) { rc.beginSessionTeardown().finish() }},
	925	+	}
	926	+	for _, tc := range cases {
	927	+	t.Run(tc.name, func(t *testing.T) {
	928	+	h := newCoordinatorHarness(t)
	929	+	owner := &attachedClient{}
	930	+	h.rc.attach(owner)
	931	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput})
	932	+	stale := h.armedTimers(t)
	933	+	require.NotEmpty(t, stale, "invalidate must arm a deadline timer")
	934	+
	935	+	tc.teardown(h.rc, owner)
	936	+	for _, timer := range stale {
	937	+	timer.ch <- time.Time{}
	938	+	}
	939	+	requireNoWake(t, h.wakes)
	940	+
	941	+	h.rc.invalidate(renderInvalidation{class: invalidateUrgent})
	942	+	for _, timer := range h.armedTimers(t) {
	943	+	timer.ch <- time.Time{}
	944	+	}
	945	+	requireNoWake(t, h.wakes)
	946	+	})
	947	+	}
	948	+
	949	+	t.Run("replacement attachment wakes independently", func(t *testing.T) {
	950	+	h := newCoordinatorHarness(t)
	951	+	owner := &attachedClient{}
	952	+	replacement := &attachedClient{}
	953	+	h.rc.attach(owner)
	954	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput})
	955	+	stale := h.armedTimers(t)
	956	+	require.NotEmpty(t, stale, "invalidate must arm a deadline timer")
	957	+
	958	+	h.rc.noteReplace(owner, replacement)
	959	+	h.rc.attach(replacement)
	960	+	for _, timer := range stale {
	961	+	timer.ch <- time.Time{}
	962	+	}
	963	+	requireNoWake(t, h.wakes)
	964	+
	965	+	h.rc.invalidate(renderInvalidation{class: invalidateUrgent, reset: true})
	966	+	fresh := h.armedTimers(t)
	967	+	require.NotEmpty(t, fresh, "a replacement attachment must arm its own deadline")
	968	+	fresh[len(fresh)-1].ch <- time.Time{}
	969	+	w := awaitWake(t, h.wakes)
	970	+	require.True(t, w.reset, "the replacement starts from an independent full state")
	971	+	requireNoWake(t, h.wakes)
	972	+	})
	973	+	}
	974	+
	975	+	// --- resize metadata ownership -----------------------------------------------------
	976	+
	977	+	func TestRenderCoordinatorResizeMetadata(t *testing.T) {
	978	+	sz := func(cols, rows int) domain.Size { return domain.Size{Cols: cols, Rows: rows} }
	979	+	cases := []struct {
	980	+	name string
	981	+	run func(t testing.T, rc renderCoordinator, owner, replacement *attachedClient)
	982	+	}{
	983	+	{
	984	+	name: "latest request wins with strictly increasing epochs",
	985	+	run: func(t testing.T, rc renderCoordinator, owner, _ *attachedClient) {
	986	+	require.Equal(t, uint64(1), rc.recordResizeRequest(sz(100, 30), owner))
	987	+	require.Equal(t, uint64(2), rc.recordResizeRequest(sz(110, 32), owner))
	988	+	require.Equal(t, uint64(3), rc.recordResizeRequest(sz(120, 40), owner))
	989	+	snap := rc.resizeSnapshot()
	990	+	require.Equal(t, sz(120, 40), snap.size)
	991	+	require.Same(t, owner, snap.source)
	992	+	require.Equal(t, uint64(3), snap.epoch)
	993	+
	994	+	// Resize state is exclusively coordinator-owned.
	995	+	require.Zero(t, snap.committed)
	996	+	},
	997	+	},
	998	+	{
	999	+	name: "duplicate sizes still advance the epoch deterministically",
	1000	+	run: func(t testing.T, rc renderCoordinator, owner, _ *attachedClient) {
	1001	+	require.Equal(t, uint64(1), rc.recordResizeRequest(sz(120, 40), owner))
	1002	+	require.Equal(t, uint64(2), rc.recordResizeRequest(sz(120, 40), owner))
	1003	+	snap := rc.resizeSnapshot()
	1004	+	require.Equal(t, sz(120, 40), snap.size)
	1005	+	require.Equal(t, uint64(2), snap.epoch)
	1006	+	},
	1007	+	},
	1008	+	{
	1009	+	name: "stale resize callbacks preserve latest metadata through lifecycle transitions",
	1010	+	run: func(t testing.T, rc renderCoordinator, owner, replacement *attachedClient) {
	1011	+	type lifecycleCase struct {
	1012	+	name string
	1013	+	installNew bool
	1014	+	}
	1015	+	for _, tc := range []lifecycleCase{
	1016	+	{name: "detach"},
	1017	+	{name: "park"},
	1018	+	{name: "session teardown"},
	1019	+	{name: "replacement", installNew: true},
	1020	+	} {
	1021	+	t.Run(tc.name, func(t *testing.T) {
	1022	+	h := newCoordinatorHarness(t)
	1023	+	staleOwner := &attachedClient{}
	1024	+	freshOwner := &attachedClient{}
	1025	+	h.rc.attach(staleOwner)
	1026	+	require.Equal(t, uint64(1), h.rc.recordResizeRequest(sz(120, 40), staleOwner))
	1027	+
	1028	+	// This models a stale callback which captured the old
	1029	+	// attachment before lifecycle ownership changed.
	1030	+	staleCallback := func() uint64 {
	1031	+	return h.rc.recordResizeRequest(sz(80, 20), staleOwner)
	1032	+	}
	1033	+
	1034	+	switch tc.name {
	1035	+	case "detach":
	1036	+	h.rc.noteDetach(staleOwner)
	1037	+	case "park":
	1038	+	h.rc.notePark(staleOwner)
	1039	+	case "session teardown":
	1040	+	h.rc.beginSessionTeardown().finish()
	1041	+	case "replacement":
	1042	+	h.rc.noteReplace(staleOwner, freshOwner)
	1043	+	h.rc.attach(freshOwner)
	1044	+	require.Equal(t, uint64(2), h.rc.recordResizeRequest(sz(100, 50), freshOwner))
	1045	+	}
	1046	+
	1047	+	require.Zero(t, staleCallback(), "stale callback must not advance the resize epoch")
	1048	+	snap := h.rc.resizeSnapshot()
	1049	+	if tc.installNew {
	1050	+	require.Equal(t, sz(100, 50), snap.size)
	1051	+	require.Equal(t, uint64(2), snap.epoch)
	1052	+	require.Same(t, freshOwner, snap.source)
	1053	+	return
	1054	+	}
	1055	+	require.Equal(t, sz(120, 40), snap.size)
	1056	+	require.Equal(t, uint64(1), snap.epoch)
	1057	+	require.Same(t, staleOwner, snap.source,
	1058	+	"the stale callback must not replace the recorded attachment identity")
	1059	+	})
	1060	+	}
	1061	+	},
	1062	+	},
	1063	+	}
	1064	+	for _, tc := range cases {
	1065	+	t.Run(tc.name, func(t *testing.T) {
	1066	+	h := newCoordinatorHarness(t)
	1067	+	owner := &attachedClient{}
	1068	+	replacement := &attachedClient{}
	1069	+	h.rc.attach(owner)
	1070	+	tc.run(t, h.rc, owner, replacement)
	1071	+	})
	1072	+	}
	1073	+	}
	1074	+
	1075	+	// --- producer fan-in ------------------------------------------------------------
	1076	+
	1077	+	// producerFiles is the current production direct-paint inventory. Every file
	1078	+	// gets exactly one exercised state transition in TestProducerInvalidations.
	1079	+	var producerFiles = []string{
	1080	+	"attention.go", "client.go", "copymode.go", "floating.go", "input.go",
	1081	+	"palette.go", "pane_actions.go", "picker.go", "prompt.go", "render.go",
	1082	+	"session.go", "session_back.go",
	1083	+	}
	1084	+
	1085	+	func TestProducerInvalidations(t *testing.T) {
	1086	+	cases := []struct {
	1087	+	file string
	1088	+	name string
	1089	+	tabs int
	1090	+	run func(t testing.T, d Daemon, sess session, ac attachedClient)
	1091	+	}{
	1092	+	{
	1093	+	file: "attention.go",
	1094	+	name: "attention repaint",
	1095	+	run: func(t testing.T, d Daemon, sess session, ac attachedClient) {
	1096	+	d.repaintAttachedClients(sess)
	1097	+	},
	1098	+	},
	1099	+	{
	1100	+	file: "client.go",
	1101	+	name: "client theme application",
	1102	+	run: func(t testing.T, d Daemon, sess session, ac attachedClient) {
	1103	+	d.applyTheme(sess, ac, ports.Theme{TrueColor: true})
	1104	+	},
	1105	+	},
	1106	+	{
	1107	+	file: "copymode.go",
	1108	+	name: "copy mode entry",
	1109	+	run: func(t testing.T, d Daemon, sess session, ac attachedClient) {
	1110	+	d.enterCopyMode(sess, ac)
	1111	+	},
	1112	+	},
	1113	+	{
	1114	+	file: "floating.go",
	1115	+	name: "floating toggle to visible",
	1116	+	run: func(t testing.T, d Daemon, sess session, ac attachedClient) {
	1117	+	fp := newPaneWithStableID(layout.PaneID("floating"), "float-producer", newQuietPTY(), domain.Size{Cols: 40, Rows: 10})
	1118	+	tb := sess.activeTab()
	1119	+	tb.mu.Lock()
	1120	+	tb.floating.pane = fp
	1121	+	tb.floating.state = floatingHidden
	1122	+	tb.floating.generation = 1
	1123	+	tb.mu.Unlock()
	1124	+	require.NoError(t, d.toggleFloating(sess, ac))
	1125	+	},
	1126	+	},
	1127	+	{
	1128	+	file: "input.go",
	1129	+	name: "switch tab key action",
	1130	+	tabs: 2,
	1131	+	run: func(t testing.T, d Daemon, sess session, ac attachedClient) {
	1132	+	daemonKeyHandler{d: d, ac: ac}.Action(keys.ActionSwitchTab2)
	1133	+	require.Equal(t, 1, activeTabIndex(sess))
	1134	+	},
	1135	+	},
	1136	+	{
	1137	+	file: "palette.go",
	1138	+	name: "palette entry",
	1139	+	run: func(t testing.T, d Daemon, sess session, ac attachedClient) {
	1140	+	d.enterPalette(sess, ac)
	1141	+	},
	1142	+	},
	1143	+	{
	1144	+	file: "pane_actions.go",
	1145	+	name: "pane split",
	1146	+	run: func(t testing.T, d Daemon, sess session, ac attachedClient) {
	1147	+	require.NoError(t, d.splitPane(sess, ac, layout.Right))
	1148	+	},
	1149	+	},
	1150	+	{
	1151	+	file: "picker.go",
	1152	+	name: "picker entry",
	1153	+	run: func(t testing.T, d Daemon, sess session, ac attachedClient) {
	1154	+	d.enterPicker(sess, ac)
	1155	+	},
	1156	+	},
	1157	+	{
	1158	+	file: "prompt.go",
	1159	+	name: "prompt entry",
	1160	+	run: func(t testing.T, d Daemon, sess session, ac attachedClient) {
	1161	+	d.enterPrompt(sess, ac, "rename", "", func(string) error { return nil })
	1162	+	},
	1163	+	},
	1164	+	{
	1165	+	file: "render.go",
	1166	+	name: "retained resize dispatch",
	1167	+	run: func(t testing.T, d Daemon, sess session, ac attachedClient) {
	1168	+	clk := newCoordinatorMockClock(t, 2)
	1169	+	// The coordinator owns deadline time; install the deterministic
	1170	+	// clock on the already-attached coordinator rather than changing
	1171	+	// Daemon's construction-time clock afterwards.
	1172	+	sess.renderCoordinator().opts.clock = clk.clock
	1173	+	d.resize(sess, ac, domain.Size{Cols: 100, Rows: 26})
	1174	+	timer := awaitCoordinatorScheduledTimer(t, clk)
	1175	+	done := captureResizeCallbackDone(t, sess.renderCoordinator())
	1176	+	timer.ch <- time.Time{}
	1177	+	awaitTestCompletion(t, done, "resize callback did not complete")
	1178	+	},
	1179	+	},
	1180	+	{
	1181	+	file: "session.go",
	1182	+	name: "tab close repaint",
	1183	+	tabs: 2,
	1184	+	run: func(t testing.T, d Daemon, sess session, ac attachedClient) {
	1185	+	sess.mu.Lock()
	1186	+	tb := sess.tabs[1]
	1187	+	sess.mu.Unlock()
	1188	+	d.closeTab(sess, tb, true)
	1189	+	},
	1190	+	},
	1191	+	{
	1192	+	file: "session_back.go",
	1193	+	name: "back session fallback without a target",
	1194	+	run: func(t testing.T, d Daemon, sess session, ac attachedClient) {
	1195	+	d.backSession(sess, ac)
	1196	+	require.Same(t, sess, ac.currentSession())
	1197	+	},
	1198	+	},
	1199	+	}
	1200	+
	1201	+	exercised := make([]string, 0, len(cases))
	1202	+	for _, tc := range cases {
	1203	+	exercised = append(exercised, tc.file)
	1204	+	}
	1205	+	require.ElementsMatch(t, producerFiles, exercised,
	1206	+	"every current direct-paint producer file needs exactly one exercised transition")
	1207	+
	1208	+	for _, tc := range cases {
	1209	+	t.Run(tc.file+"/"+tc.name, func(t *testing.T) {
	1210	+	tabs := tc.tabs
	1211	+	if tabs == 0 {
	1212	+	tabs = 1
	1213	+	}
	1214	+	d, sess, ac, sends, releases := newManualTabSession(t, tabs)
	1215	+	defer releaseAll(releases)
	1216	+	invs := make(chan renderInvalidation, 8)
	1217	+	sess.installRenderCoordinator(newRenderCoordinator(renderCoordinatorOptions{
	1218	+	clock: d.clock,
	1219	+	wake: func(renderWake) {},
	1220	+	onInvalidate: func(inv renderInvalidation) { invs <- inv },
	1221	+	}))
	1222	+
	1223	+	tc.run(t, d, sess, ac)
	1224	+
	1225	+	awaitInvalidation(t, invs)
	1226	+	requireNoInvalidation(t, invs)
	1227	+	requireNoCoordinatorOutputFrame(t, sends)
	1228	+	})
	1229	+	}
	1230	+	}
	1231	+
	1232	+	// TestProducerInvalidationInventory is a supplementary inventory guard only:
	1233	+	// it keeps producerFiles aligned with the daemon source layout so the
	1234	+	// behavioral table above cannot drift silently. It is never acceptance
	1235	+	// evidence for coordinator behavior.
	1236	+	func TestProducerInvalidationInventory(t *testing.T) {
	1237	+	for _, name := range producerFiles {
	1238	+	_, err := os.Stat(filepath.Join(".", name))
	1239	+	require.NoErrorf(t, err, "producer file %s is gone; update TestProducerInvalidations", name)
	1240	+	}
	1241	+	}
	1242	+
	1243	+	// --- coordinator resize dispatch ----------------------------------------------
	1244	+
	1245	+	func TestConcurrentPaintInitializesOverlayUnderSendOwnership(t *testing.T) {
	1246	+	p, release := newBlockingPTY(t)
	1247	+	defer release()
	1248	+	d, sess, ac, _ := newManualSessionWithPTYs(t, p)
	1249	+	// Exercise the original seam: two fallback paints reach lazy initialization
	1250	+	// together before either can compose.
	1251	+	ac.overlays = nil
	1252	+	ac.overlayOnce = sync.Once{}
	1253	+	start := make(chan struct{})
	1254	+	var wg sync.WaitGroup
	1255	+	wg.Add(2)
	1256	+	for range 2 {
	1257	+	go func() {
	1258	+	defer wg.Done()
	1259	+	<-start
	1260	+	d.paint(sess, ac, true, nil)
	1261	+	}()
	1262	+	}
	1263	+	close(start)
	1264	+	wg.Wait()
	1265	+	require.NotNil(t, ac.overlays)
	1266	+	}
	1267	+
	1268	+	func TestStartPaneGoroutinesAccountsForOneReader(t *testing.T) {
	1269	+	p, release := newBlockingPTY(t)
	1270	+	d, sess, _, _ := newManualSessionWithPTYs(t, p)
	1271	+	tb := sess.activeTab()
	1272	+	require.NotNil(t, tb)
	1273	+	d.startPaneGoroutines(sess, tb, tb.focusedPane())
	1274	+	release()
	1275	+	select {
	1276	+	case <-waitGroupDone(&d.sessWg):
	1277	+	case <-time.After(time.Second):
	1278	+	t.Fatal("one reader must balance exactly one WaitGroup count")
	1279	+	}
	1280	+	}
	1281	+
	1282	+	func requireWorkerExit(t *testing.T, done <-chan struct{}) {
	1283	+	t.Helper()
	1284	+	awaitTestCompletion(t, done, "cancelled coordinator timer worker did not exit")
	1285	+	}
	1286	+
	1287	+	func TestCoordinatorDeadlineCannotPaintPublishedReplacementBeforeOwnershipInstall(t *testing.T) {
	1288	+	p, releasePTY := newBlockingPTY(t)
	1289	+	defer releasePTY()
	1290	+	d, sess, owner, ownerSends := newManualSessionWithPTYs(t, p)
	1291	+	replacementTransport, replacementSends := newCapturingTransport(t)
	1292	+	replacement := &attachedClient{tr: replacementTransport, output: newOutputStateStream(), size: owner.size}
	1293	+	replacement.initOverlays()
	1294	+	replacement.setSession(sess)
	1295	+
	1296	+	clock := newCoordinatorMockClock(t, 2)
	1297	+	d.clock = clock.clock
	1298	+	rc := newRenderCoordinator(renderCoordinatorOptions{
	1299	+	clock: clock.clock,
	1300	+	wake: func(w renderWake) {
	1301	+	// This is the production ownership boundary: composition must use
	1302	+	// the coordinator's captured attachment, never sess.client.
	1303	+	d.paint(sess, w.lease.attachment, w.reset, w.lease)
	1304	+	},
	1305	+	ackReady: func() bool { return true },
	1306	+	})
	1307	+	rc.attach(owner)
	1308	+	sess.installRenderCoordinator(rc)
	1309	+	rc.invalidate(renderInvalidation{class: invalidateOutput, reset: true})
	1310	+	timer := awaitCoordinatorScheduledTimer(t, clock)
	1311	+
	1312	+	// Model attachClient's publication window exactly: sess.client is new,
	1313	+	// but coordinator replacement has not yet invalidated the old deadline.
	1314	+	sess.mu.Lock()
	1315	+	sess.client = replacement
	1316	+	sess.mu.Unlock()
	1317	+	timer.ch <- time.Time{}
	1318	+	requireNoCoordinatorOutputFrame(t, replacementSends)
	1319	+	requireNoCoordinatorOutputFrame(t, ownerSends)
	1320	+	}
	1321	+
	1322	+	func TestRenderCoordinatorResizeLaneRejectsStaleToken(t *testing.T) {
	1323	+	rc := newRenderCoordinator(renderCoordinatorOptions{})
	1324	+	newer := &timerToken{generation: 2, timer: &portsmocks.MockTimer{}, cancel: make(chan struct{}), done: make(chan struct{})}
	1325	+	stale := &timerToken{generation: 1}
	1326	+
	1327	+	rc.mu.Lock()
	1328	+	rc.resizeLane.generation = 2
	1329	+	rc.resizeLane.token = newer
	1330	+	require.False(t, rc.resizeLane.clearLocked(stale))
	1331	+	require.Same(t, newer, rc.resizeLane.token)
	1332	+	require.True(t, rc.resizeLane.clearLocked(newer))
	1333	+	require.Nil(t, rc.resizeLane.token)
	1334	+	rc.mu.Unlock()
	1335	+	}
	1336	+
	1337	+	func TestRenderCoordinatorNilClockRetryCompletesWithoutSynchronousRun(t *testing.T) {
	1338	+	rc := newRenderCoordinator(renderCoordinatorOptions{})
	1339	+	owner := &attachedClient{}
	1340	+	rc.attach(owner)
	1341	+	lease := rc.attachmentLease(owner)
	1342	+	epoch := rc.recordResizeRequestForLease(domain.Size{Cols: 120, Rows: 40}, owner, lease)
	1343	+	require.NotZero(t, epoch)
	1344	+	require.True(t, rc.resizeCurrentForLease(epoch, owner, lease, true))
	1345	+
	1346	+	called := false
	1347	+	rc.scheduleResizeRetryForLease(epoch, owner, lease, func() { called = true })
	1348	+	require.False(t, called, "a nil clock is a disabled retry lane, not a synchronous retry")
	1349	+	rc.mu.Lock()
	1350	+	retryToken := rc.retryLane.token
	1351	+	rc.mu.Unlock()
	1352	+	require.Nil(t, retryToken, "a disabled retry lane must not retain stale completion ownership")
	1353	+	}
	1354	+
	1355	+	func TestRenderCoordinatorInertTimerFiresSynchronouslyWithoutWorker(t *testing.T) {
	1356	+	clock := portsmocks.NewMockClock(t)
	1357	+	timer := portsmocks.NewMockTimer(t)
	1358	+	clock.EXPECT().NewTimer(minOutputRenderDeadline).Return(timer).Once()
	1359	+	timer.EXPECT().C().Return((<-chan time.Time)(nil)).Once()
	1360	+	timer.EXPECT().Stop().Return(true).Once()
	1361	+
	1362	+	wakes := make(chan renderWake, 1)
	1363	+	rc := newRenderCoordinator(renderCoordinatorOptions{
	1364	+	clock: clock,
	1365	+	ackReady: func() bool { return true },
	1366	+	wake: func(w renderWake) { wakes <- w },
	1367	+	})
	1368	+	rc.invalidate(renderInvalidation{class: invalidateOutput})
	1369	+
	1370	+	require.Equal(t, renderWake{coalesced: 1}, <-wakes)
	1371	+	require.Equal(t, renderCoordinatorBurstMetricsSnapshot{invalidations: 1, wakes: 1, coalesced: 1}, rc.burstMetricsSnapshot())
	1372	+	rc.mu.Lock()
	1373	+	require.Nil(t, rc.normalLane.token)
	1374	+	rc.mu.Unlock()
	1375	+	}
	1376	+
	1377	+	func TestRenderCoordinatorSyncBatchSurvivesAttachmentLifecycle(t *testing.T) {
	1378	+	t.Run("detach and park retain a gated headless preview until complete", func(t *testing.T) {
	1379	+	for _, transition := range []struct {
	1380	+	name string
	1381	+	run func(renderCoordinator, attachedClient)
	1382	+	}{
	1383	+	{"detach", func(rc renderCoordinator, ac attachedClient) { rc.noteDetach(ac) }},
	1384	+	{"park", func(rc renderCoordinator, ac attachedClient) { rc.notePark(ac) }},
	1385	+	} {
	1386	+	t.Run(transition.name, func(t *testing.T) {
	1387	+	h := newCoordinatorHarness(t)
	1388	+	target, viewer, p := &attachedClient{}, &attachedClient{}, &pane{}
	1389	+	h.rc.attach(target)
	1390	+	h.rc.subscribePreviewFor(viewer, 1, func(w renderWake) { h.previews <- w })
	1391	+	h.rc.noteSyncBegin(p, 1)
	1392	+	watchdog := h.armedTimers(t)[0]
	1393	+	transition.run(h.rc, target)
	1394	+
	1395	+	// PTY-owned output remains visible to picker observers while the
	1396	+	// target attachment is parked. Its deadline must retain the sync gate.
	1397	+	h.rc.invalidate(renderInvalidation{class: invalidateOutput, producer: "complete preview"})
	1398	+	timers := h.armedTimers(t)
	1399	+	require.Len(t, timers, 1)
	1400	+	h.rc.mu.Lock()
	1401	+	normalWorkerDone := h.rc.normalLane.token.done
	1402	+	h.rc.mu.Unlock()
	1403	+	require.NotNil(t, normalWorkerDone)
	1404	+	timers[0].ch <- time.Time{}
	1405	+	requireWorkerExit(t, normalWorkerDone)
	1406	+	requireNoWake(t, h.previews)
	1407	+
	1408	+	h.rc.mu.Lock()
	1409	+	require.Same(t, h.rc.syncBatches[p].lane.token.timer, watchdog.mock)
	1410	+	h.rc.mu.Unlock()
	1411	+	h.rc.noteSyncEnd(p, 1)
	1412	+	preview := awaitWake(t, h.previews)
	1413	+	preview.lease = nil
	1414	+	require.Equal(t, renderWake{urgent: true, coalesced: 1}, preview)
	1415	+	requireNoWake(t, h.previews)
	1416	+	requireNoWake(t, h.wakes)
	1417	+	})
	1418	+	}
	1419	+	})
	1420	+
	1421	+	t.Run("replacement first paint waits for active batch and rejects stale owner", func(t *testing.T) {
	1422	+	h := newCoordinatorHarness(t)
	1423	+	old, replacement, p := &attachedClient{}, &attachedClient{}, &pane{}
	1424	+	h.rc.attach(old)
	1425	+	h.rc.noteSyncBegin(p, 1)
	1426	+	watchdog := h.armedTimers(t)[0]
	1427	+	h.rc.noteReplace(old, replacement, false)
	1428	+	require.False(t, h.rc.markAttachmentReady(h.rc.attachmentLease(old)), "old attachment Welcome is stale")
	1429	+	require.True(t, h.rc.markAttachmentReady(h.rc.attachmentLease(replacement)))
	1430	+
	1431	+	h.rc.invalidateForAttachment(old, renderInvalidation{class: invalidateUrgent, producer: "stale old attachment"})
	1432	+	require.Empty(t, h.armedTimers(t))
	1433	+	h.rc.invalidateForAttachment(replacement, renderInvalidation{class: invalidateUrgent, reset: true, producer: "replacement first paint"})
	1434	+	timers := h.armedTimers(t)
	1435	+	require.Len(t, timers, 1)
	1436	+	h.rc.mu.Lock()
	1437	+	normalWorkerDone := h.rc.normalLane.token.done
	1438	+	h.rc.mu.Unlock()
	1439	+	require.NotNil(t, normalWorkerDone)
	1440	+	timers[0].ch <- time.Time{}
	1441	+	requireWorkerExit(t, normalWorkerDone)
	1442	+	requireNoWake(t, h.wakes)
	1443	+
	1444	+	h.rc.mu.Lock()
	1445	+	require.Same(t, h.rc.syncBatches[p].lane.token.timer, watchdog.mock)
	1446	+	h.rc.mu.Unlock()
	1447	+	h.rc.noteSyncEnd(p, 1)
	1448	+	wake := awaitWake(t, h.wakes)
	1449	+	require.True(t, wake.reset)
	1450	+	require.True(t, wake.urgent)
	1451	+	require.Same(t, replacement, wake.lease.attachment)
	1452	+	requireNoWake(t, h.wakes)
	1453	+	})
	1454	+
	1455	+	t.Run("pane removal and session teardown cancel sync workers", func(t *testing.T) {
	1456	+	for _, lifecycle := range []struct {
	1457	+	name string
	1458	+	end func(renderCoordinator, pane)
	1459	+	}{
	1460	+	{"pane removal", func(rc renderCoordinator, p pane) { rc.noteSyncPaneRemoved(p) }},
	1461	+	{"session teardown", func(rc renderCoordinator, _ pane) { rc.beginSessionTeardown().finish() }},
	1462	+	} {
	1463	+	t.Run(lifecycle.name, func(t *testing.T) {
	1464	+	clk := newCoordinatorMockClock(t, 2)
	1465	+	rc := newRenderCoordinator(renderCoordinatorOptions{clock: clk.clock})
	1466	+	p := &pane{}
	1467	+	rc.noteSyncBegin(p, 1)
	1468	+	syncTimer := <-clk.timers
	1469	+	rc.mu.Lock()
	1470	+	syncDone := rc.syncBatches[p].lane.token.done
	1471	+	rc.mu.Unlock()
	1472	+	lifecycle.end(rc, p)
	1473	+	syncTimer.mock.AssertNumberOfCalls(t, "Stop", 1)
	1474	+	requireWorkerExit(t, syncDone)
	1475	+	if lifecycle.name == "session teardown" {
	1476	+	rc.waitForTimerWorkers()
	1477	+	}
	1478	+	})
	1479	+	}
	1480	+	})
	1481	+	}
	1482	+
	1483	+	func TestRenderCoordinatorResizeEpochDispatch(t *testing.T) {
	1484	+	newResizeFixture := func(t testing.T) (Daemon, session, attachedClient, chan ports.Frame, *coordinatorMockClock, chan renderInvalidation) {
	1485	+	t.Helper()
	1486	+	p, releasePTY := newBlockingPTY(t)
	1487	+	t.Cleanup(releasePTY)
	1488	+	d, sess, ac, sends := newManualSessionWithPTYs(t, p)
	1489	+	clk := newCoordinatorMockClock(t, 4)
	1490	+	d.clock = clk.clock
	1491	+	invs := make(chan renderInvalidation, 4)
	1492	+	rc := newRenderCoordinator(renderCoordinatorOptions{
	1493	+	clock: clk.clock,
	1494	+	wake: func(renderWake) {},
	1495	+	onInvalidate: func(inv renderInvalidation) { invs <- inv },
	1496	+	})
	1497	+	rc.attach(ac)
	1498	+	sess.installRenderCoordinator(rc)
	1499	+	return d, sess, ac, sends, clk, invs
	1500	+	}
	1501	+
	1502	+	t.Run("bounded timer records metadata and dispatches through the coordinator", func(t *testing.T) {
	1503	+	d, sess, ac, sends, clk, invs := newResizeFixture(t)
	1504	+
	1505	+	d.resize(sess, ac, domain.Size{Cols: 120, Rows: 24})
	1506	+	timer := awaitCoordinatorScheduledTimer(t, clk)
	1507	+	require.GreaterOrEqual(t, timer.duration, minOutputRenderDeadline)
	1508	+	require.LessOrEqual(t, timer.duration, maxOutputRenderDeadline)
	1509	+	snap := sess.renderCoordinator().resizeSnapshot()
	1510	+	require.Equal(t, domain.Size{Cols: 120, Rows: 24}, snap.size,
	1511	+	"the coordinator must record the latest requested resize before delegating")
	1512	+	require.Same(t, ac, snap.source)
	1513	+	require.Equal(t, uint64(1), snap.epoch)
	1514	+	requireNoCoordinatorOutputFrame(t, sends)
	1515	+
	1516	+	done := captureResizeCallbackDone(t, sess.renderCoordinator())
	1517	+	timer.ch <- time.Time{}
	1518	+	awaitTestCompletion(t, done, "resize callback did not complete")
	1519	+	inv := awaitInvalidation(t, invs)
	1520	+	require.True(t, inv.reset, "the resize dispatch must request a full-redraw invalidation")
	1521	+	requireNoInvalidation(t, invs)
	1522	+	requireNoCoordinatorOutputFrame(t, sends)
	1523	+	})
	1524	+
	1525	+	t.Run("stale generations stay rejected", func(t *testing.T) {
	1526	+	d, sess, ac, sends, clk, invs := newResizeFixture(t)
	1527	+
	1528	+	d.resize(sess, ac, domain.Size{Cols: 100, Rows: 24})
	1529	+	first := awaitCoordinatorScheduledTimer(t, clk)
	1530	+	d.resize(sess, ac, domain.Size{Cols: 120, Rows: 24})
	1531	+	latest := awaitCoordinatorScheduledTimer(t, clk)
	1532	+	require.Equal(t, uint64(2), sess.renderCoordinator().resizeSnapshot().epoch,
	1533	+	"every resize request must advance the coordinator epoch")
	1534	+
	1535	+	done := captureResizeCallbackDone(t, sess.renderCoordinator())
	1536	+	first.ch <- time.Time{}
	1537	+	requireNoInvalidation(t, invs)
	1538	+	requireNoCoordinatorOutputFrame(t, sends)
	1539	+
	1540	+	latest.ch <- time.Time{}
	1541	+	awaitTestCompletion(t, done, "latest resize callback did not complete")
	1542	+	inv := awaitInvalidation(t, invs)
	1543	+	require.True(t, inv.reset)
	1544	+	requireNoInvalidation(t, invs)
	1545	+	requireNoCoordinatorOutputFrame(t, sends)
	1546	+	require.Equal(t, domain.Size{Cols: 120, Rows: 24}, sess.renderCoordinator().resizeSnapshot().size)
	1547	+	})
	1548	+
	1549	+	t.Run("cancellation still drops the pending dispatch", func(t *testing.T) {
	1550	+	d, sess, ac, sends, clk, invs := newResizeFixture(t)
	1551	+
	1552	+	d.resize(sess, ac, domain.Size{Cols: 90, Rows: 30})
	1553	+	timer := awaitCoordinatorScheduledTimer(t, clk)
	1554	+	done := captureResizeCallbackDone(t, sess.renderCoordinator())
	1555	+	sess.renderCoordinator().noteDetach(ac)
	1556	+
	1557	+	timer.ch <- time.Time{}
	1558	+	awaitTestCompletion(t, done, "cancelled resize callback did not complete")
	1559	+	requireNoInvalidation(t, invs)
	1560	+	requireNoCoordinatorOutputFrame(t, sends)
	1561	+	})
	1562	+
	1563	+	t.Run("request reports immediate completion and async schedule acceptance", func(t *testing.T) {
	1564	+	t.Run("immediate reset invalidation", func(t *testing.T) {
	1565	+	d, sess, ac, _, _, invs := newResizeFixture(t)
	1566	+
	1567	+	require.True(t, d.requestTransactionalResize(sess, ac, domain.Size{Cols: 120, Rows: 24}, true))
	1568	+	inv := awaitInvalidation(t, invs)
	1569	+	require.True(t, inv.reset)
	1570	+	})
	1571	+
	1572	+	t.Run("superseded immediate transaction", func(t *testing.T) {
	1573	+	resizeStarted := make(chan struct{})
	1574	+	releaseResize := make(chan struct{})
	1575	+	pty := &transactionalResizePTY{onResize: func() {
	1576	+	close(resizeStarted)
	1577	+	<-releaseResize
	1578	+	}}
	1579	+	d, sess, ac, _ := newManualSessionWithPTYs(t, pty)
	1580	+	clk := newCoordinatorMockClock(t, 2)
	1581	+	d.clock = clk.clock
	1582	+	invs := make(chan renderInvalidation, 1)
	1583	+	rc := newRenderCoordinator(renderCoordinatorOptions{
	1584	+	clock: clk.clock,
	1585	+	wake: func(renderWake) {},
	1586	+	onInvalidate: func(inv renderInvalidation) { invs <- inv },
	1587	+	})
	1588	+	rc.attach(ac)
	1589	+	sess.installRenderCoordinator(rc)
	1590	+
	1591	+	result := make(chan bool, 1)
	1592	+	go func() {
	1593	+	result <- d.requestTransactionalResize(sess, ac, domain.Size{Cols: 120, Rows: 24}, true)
	1594	+	}()
	1595	+	<-resizeStarted
	1596	+	require.NotZero(t, rc.recordResizeRequest(domain.Size{Cols: 140, Rows: 30}, ac))
	1597	+	close(releaseResize)
	1598	+
	1599	+	require.False(t, <-result)
	1600	+	requireNoInvalidation(t, invs)
	1601	+	})
	1602	+
	1603	+	t.Run("async schedule and teardown rejection", func(t *testing.T) {
	1604	+	d, sess, ac, _, clk, _ := newResizeFixture(t)
	1605	+
	1606	+	require.True(t, d.requestTransactionalResize(sess, ac, domain.Size{Cols: 120, Rows: 24}, false))
	1607	+	awaitCoordinatorScheduledTimer(t, clk)
	1608	+
	1609	+	// A torn-down coordinator cannot accept a stale attachment's schedule.
	1610	+	sess.renderCoordinator().beginSessionTeardown().finish()
	1611	+	require.False(t, d.requestTransactionalResize(sess, ac, domain.Size{Cols: 140, Rows: 30}, false))
	1612	+	})
	1613	+	})
	1614	+	}

A internal/usecase/daemon/render_coordinator_types.go +176 -0

@@ -0,0 +1,176 @@
	1	+	package daemon
	2	+
	3	+	import (
	4	+	"sync"
	5	+	"sync/atomic"
	6	+	"time"
	7	+
	8	+	"github.com/bnema/vev/internal/domain"
	9	+	"github.com/bnema/vev/internal/ports"
	10	+	)
	11	+
	12	+	// Coordinator contracts and support types are kept with their ownership
	13	+	// semantics, separate from the scheduling transitions in render_coordinator.
	14	+
	15	+	// One coordinator serves a session: producers mutate authoritative state and
	16	+	// publish an invalidation; only the coordinator decides when composition runs.
	17	+	// Producers never compose or send.
	18	+
	19	+	// Coordinator deadline bounds. Urgent transitions (input echo, overlay state
	20	+	// changes) must become visible within urgentRenderDeadline; bulk output
	21	+	// coalesces inside the adaptive [minOutputRenderDeadline,
	22	+	// maxOutputRenderDeadline] window.
	23	+	const (
	24	+	urgentRenderDeadline = 2 * time.Millisecond
	25	+	minOutputRenderDeadline = 8 * time.Millisecond
	26	+	maxOutputRenderDeadline = 16 * time.Millisecond
	27	+	)
	28	+
	29	+	// invalidationClass ranks a producer transition's latency demand.
	30	+	type invalidationClass uint8
	31	+
	32	+	const (
	33	+	// invalidateOutput is bulk PTY/screen churn served by the adaptive window.
	34	+	invalidateOutput invalidationClass = iota
	35	+	// invalidateUrgent is an interactive transition served within 2ms.
	36	+	invalidateUrgent
	37	+	)
	38	+
	39	+	// renderInvalidation is one producer-published state transition.
	40	+	type renderInvalidation struct {
	41	+	class invalidationClass
	42	+	// reset requests a sticky FullRedraw: once requested it survives
	43	+	// coalescing and ACK deferral until a wake actually delivers it.
	44	+	reset bool
	45	+	// producer identifies the publishing call-site for provenance.
	46	+	producer string
	47	+	}
	48	+
	49	+	// renderWake is one coalesced composition request.
	50	+	type renderWake struct {
	51	+	reset bool
	52	+	urgent bool
	53	+	coalesced int
	54	+	// lease is snapshotted under coordinator ownership. The wake consumer must
	55	+	// compose only for this exact active attachment incarnation, never by
	56	+	// rereading session.client after an attach publication.
	57	+	lease *attachmentLease
	58	+	// watchdog marks a flush forced by the synchronized-output watchdog.
	59	+	watchdog bool
	60	+	}
	61	+
	62	+	// attachmentLease is the sole lifecycle identity for a primary render
	63	+	// attachment. It is only mutated while renderCoordinator.mu is held; callbacks
	64	+	// carry its pointer and must revalidate it before every side effect.
	65	+	type attachmentLease struct {
	66	+	attachment *attachedClient
	67	+	ready bool
	68	+	active bool
	69	+	// A lease is an immutable callback capability. Its lifecycle bits are
	70	+	// changed only under renderCoordinator.mu; callbacks revalidate it at each
	71	+	// short effect boundary and never retain coordinator ownership while routing
	72	+	// arbitrary handlers.
	73	+	}
	74	+
	75	+	// resizeRequestMetadata is the coordinator-owned latest requested resize
	76	+	// state. Resize dispatch reads it through resizeSnapshot and validates the
	77	+	// coordinator epoch before applying it.
	78	+	type resizeRequestMetadata struct {
	79	+	size domain.Size
	80	+	source *attachedClient
	81	+	// lease binds every delayed resize/retry callback to the exact attachment
	82	+	// incarnation that published this request.
	83	+	lease *attachmentLease
	84	+	epoch uint64 // latest requested epoch
	85	+	committed uint64 // latest published epoch
	86	+	}
	87	+
	88	+	// renderCoordinatorOptions wires one coordinator instance.
	89	+	type renderCoordinatorOptions struct {
	90	+	clock ports.Clock
	91	+	observer ports.RuntimeObserver
	92	+	// wake composes a coalesced render request.
	93	+	wake func(renderWake)
	94	+	// ackReady reports whether the attachment may compose another output
	95	+	// state (the outputStateStream window has capacity).
	96	+	ackReady func() bool
	97	+	// syncRenderable reports whether a pane's synchronized batch currently
	98	+	// contributes to the attached composition. It must not acquire c.mu.
	99	+	syncRenderable func(*pane) bool
	100	+	// syncActive is retained for existing test harness construction; per-pane
	101	+	// lifecycle and syncRenderable are the coordinator gate.
	102	+	syncActive func() bool
	103	+	// onInvalidate observes every published invalidation (test-visible hook).
	104	+	onInvalidate func(renderInvalidation)
	105	+	// afterSyncGateEvaluated is a package-private deterministic test seam. It
	106	+	// runs unlocked after visibility predicates and before registry validation.
	107	+	afterSyncGateEvaluated func()
	108	+	}
	109	+
	110	+	// renderCoordinator fans in producer invalidations for one attached session.
	111	+	// renderCoordinatorBurstMetrics is deliberately observational. It permits
	112	+	// benchmark/pprof callers to measure producer bursts without influencing
	113	+	// scheduling or transport policy.
	114	+	type renderCoordinatorBurstMetrics struct {
	115	+	invalidations atomic.Uint64
	116	+	wakes atomic.Uint64
	117	+	coalesced atomic.Uint64
	118	+	}
	119	+
	120	+	// ackBlockedSpan hands an ACK-blocked endpoint pair between coordinator
	121	+	// transitions without holding c.mu while the observer runs. A consuming ACK
	122	+	// may arrive while Start is blocked in observer I/O; finish records that
	123	+	// closure and publishStart emits it only after Start has returned.
	124	+	type ackBlockedSpan struct {
	125	+	observer ports.RuntimeObserver
	126	+	correlation ports.RuntimeCorrelation
	127	+
	128	+	mu sync.Mutex
	129	+	startPublished bool
	130	+	finished bool
	131	+	endPending bool
	132	+	endValid bool
	133	+	}
	134	+
	135	+	func newACKBlockedSpan(observer ports.RuntimeObserver) *ackBlockedSpan {
	136	+	return &ackBlockedSpan{observer: observer, correlation: ports.NewRuntimeCorrelation()}
	137	+	}
	138	+
	139	+	func (s *ackBlockedSpan) publishStart() {
	140	+	s.observer.ObserveRuntime(ports.NewRuntimeMarkWithCorrelation("daemon", s.correlation, ports.RuntimeACKBlockedStart, 0, true))
	141	+
	142	+	s.mu.Lock()
	143	+	s.startPublished = true
	144	+	endPending, endValid := s.endPending, s.endValid
	145	+	s.endPending = false
	146	+	s.mu.Unlock()
	147	+	if endPending {
	148	+	s.observer.ObserveRuntime(ports.NewRuntimeMarkWithCorrelation("daemon", s.correlation, ports.RuntimeACKBlockedEnd, 0, endValid))
	149	+	}
	150	+	}
	151	+
	152	+	// finish closes a span at most once. It never invokes the observer before the
	153	+	// blocked start has completed publication.
	154	+	func (s *ackBlockedSpan) finish(valid bool) {
	155	+	s.mu.Lock()
	156	+	if s.finished {
	157	+	s.mu.Unlock()
	158	+	return
	159	+	}
	160	+	s.finished = true
	161	+	if !s.startPublished {
	162	+	s.endPending, s.endValid = true, valid
	163	+	s.mu.Unlock()
	164	+	return
	165	+	}
	166	+	s.mu.Unlock()
	167	+	s.observer.ObserveRuntime(ports.NewRuntimeMarkWithCorrelation("daemon", s.correlation, ports.RuntimeACKBlockedEnd, 0, valid))
	168	+	}
	169	+
	170	+	// renderCoordinatorBurstMetricsSnapshot is an immutable, internal view for
	171	+	// benchmark reporting. Scheduling only writes the atomic counters above.
	172	+	type renderCoordinatorBurstMetricsSnapshot struct {
	173	+	invalidations uint64
	174	+	wakes uint64
	175	+	coalesced uint64
	176	+	}

A internal/usecase/daemon/render_damage_receipt_test.go +128 -0

@@ -0,0 +1,128 @@
	1	+	package daemon
	2	+
	3	+	import (
	4	+	"testing"
	5	+
	6	+	"github.com/bnema/vev/internal/domain"
	7	+	"github.com/bnema/vev/internal/ports"
	8	+	"github.com/bnema/vev/internal/usecase/picker"
	9	+	"github.com/bnema/vev/pkg/renderer"
	10	+	"github.com/stretchr/testify/require"
	11	+	)
	12	+
	13	+	func TestPrimaryCaptureAloneRecordsDamageReceipts(t *testing.T) {
	14	+	_, sess, ac, _ := newManualSessionWithPTYs(t, nil)
	15	+	p := sess.tabs[0].focusedPane()
	16	+	p.mu.Lock()
	17	+	p.screen.ClearDamage()
	18	+	p.screen.Write([]byte("preview-safe"))
	19	+	p.mu.Unlock()
	20	+
	21	+	preview := snapshotPickerPreview(sess.tabs[0])
	22	+	require.NotEmpty(t, preview.Rows)
	23	+	require.Empty(t, ac.renderScratch.receipts, "picker preview must not create primary damage receipts")
	24	+	p.mu.Lock()
	25	+	require.NotEmpty(t, p.screen.Damage(), "picker preview must not acknowledge VT damage")
	26	+	p.mu.Unlock()
	27	+
	28	+	ac.sendMu.Lock()
	29	+	state, ok := capturePrimaryRenderState(sess, ac, barState{}, capturedOverlayRenderState{}, pickerPreviewEmpty(), domain.FloatingConfig{}, false, nil)
	30	+	ac.sendMu.Unlock()
	31	+	require.True(t, ok)
	32	+	require.Len(t, state.receipts, 1)
	33	+	require.Same(t, p, state.receipts[0].pane, "the primary transaction owns the pane receipt")
	34	+	p.mu.Lock()
	35	+	require.NotEmpty(t, p.screen.Damage(), "primary capture remains non-destructive until emission commits")
	36	+	p.mu.Unlock()
	37	+	}
	38	+
	39	+	// These are deliberately end-to-end pipeline tests: the pending damage comes
	40	+	// from a real pane VT Write, not a hand-built FullRedraw.
	41	+	func TestRenderDamageReceiptRetainsRealVTDamageAcrossFailedEmission(t *testing.T) {
	42	+	for _, tt := range []struct {
	43	+	name string
	44	+	fail func(composedRenderFrame, attachedClient)
	45	+	restore func(session, attachedClient, ports.Transport)
	46	+	}{
	47	+	{
	48	+	name: "prepare failure",
	49	+	fail: func(c composedRenderFrame, _ attachedClient) {
	50	+	// An invalid frame makes the real output.prepare reject after capture.
	51	+	c.frame = renderer.Frame{Width: 1}
	52	+	},
	53	+	restore: func(_ session, _ attachedClient, _ ports.Transport) {},
	54	+	},
	55	+	{
	56	+	name: "send failure",
	57	+	fail: func(_ composedRenderFrame, ac attachedClient) { ac.replaceTransport(cacheFailTransport{}) },
	58	+	restore: func(sess session, ac attachedClient, healthy ports.Transport) {
	59	+	sess.mu.Lock()
	60	+	sess.client = ac
	61	+	sess.mu.Unlock()
	62	+	ac.setSession(sess)
	63	+	ac.replaceTransport(healthy)
	64	+	},
	65	+	},
	66	+	} {
	67	+	t.Run(tt.name, func(t *testing.T) {
	68	+	d, sess, ac, sends := newManualSessionWithPTYs(t, nil)
	69	+	healthy := ac.transport()
	70	+	p := sess.tabs[0].focusedPane()
	71	+
	72	+	// Prime both the renderer shadow and the committed capture cache.
	73	+	d.paint(sess, ac, true, nil)
	74	+	<-sends
	75	+	p.mu.Lock()
	76	+	p.screen.Write([]byte("\x1b[1;1Hchanged"))
	77	+	p.mu.Unlock()
	78	+
	79	+	state, composed := captureComposeForReceiptTest(t, sess, ac)
	80	+	tt.fail(&composed, ac)
	81	+	require.True(t, d.emitFrame(sess, ac, state, composed))
	82	+	p.mu.Lock()
	83	+	require.NotEmpty(t, p.screen.Damage(), "failed emission must retain real VT damage")
	84	+	p.mu.Unlock()
	85	+
	86	+	tt.restore(sess, ac, healthy)
	87	+	state, composed = captureComposeForReceiptTest(t, sess, ac)
	88	+	require.True(t, d.emitFrame(sess, ac, state, composed))
	89	+	frame := <-sends
	90	+	out, err := ports.UnmarshalOutput(frame.Payload)
	91	+	require.NoError(t, err)
	92	+	require.Contains(t, string(out.Data), "changed", "retry must emit the retained terminal bytes")
	93	+	p.mu.Lock()
	94	+	require.Empty(t, p.screen.Damage(), "only the successful retry commits its receipt")
	95	+	require.Equal(t, 'c', p.screen.Frame.At(0, 0).Rune, "retry preserves the changed VT shadow")
	96	+	p.mu.Unlock()
	97	+	})
	98	+	}
	99	+	}
	100	+
	101	+	func TestRenderDamageReceiptStaleGenerationForcesFullRedraw(t *testing.T) {
	102	+	d, sess, ac, sends := newManualSessionWithPTYs(t, nil)
	103	+	d.paint(sess, ac, true, nil)
	104	+	<-sends
	105	+	p := sess.tabs[0].focusedPane()
	106	+	p.mu.Lock()
	107	+	p.screen.Write([]byte("\x1b[1;1Hfirst"))
	108	+	p.mu.Unlock()
	109	+	state, composed := captureComposeForReceiptTest(t, sess, ac)
	110	+	p.mu.Lock()
	111	+	p.screen.Write([]byte("\x1b[2;1Hsecond"))
	112	+	p.mu.Unlock()
	113	+	require.True(t, d.emitFrame(sess, ac, state, composed))
	114	+	<-sends
	115	+	p.mu.Lock()
	116	+	defer p.mu.Unlock()
	117	+	require.Equal(t, []renderer.Damage{renderer.FullRedraw()}, p.screen.Damage(), "a stale receipt conservatively preserves intervening VT output")
	118	+	}
	119	+
	120	+	func captureComposeForReceiptTest(t testing.T, sess session, ac attachedClient) (capturedRenderState, composedRenderFrame) {
	121	+	t.Helper()
	122	+	ac.sendMu.Lock() // emitFrame releases the transaction lock.
	123	+	state, ok := capturePrimaryRenderState(sess, ac, barState{}, capturedOverlayRenderState{}, pickerPreviewEmpty(), domain.FloatingConfig{}, false, nil)
	124	+	require.True(t, ok)
	125	+	return state, composeFrame(*state, ac.pipelineCache, ac.pipelineScratch)
	126	+	}
	127	+
	128	+	func pickerPreviewEmpty() picker.Preview { return picker.Preview{} }

A internal/usecase/daemon/render_legacy_test.go +279 -0

@@ -0,0 +1,279 @@
	1	+	package daemon
	2	+
	3	+	import (
	4	+	"github.com/bnema/vev/internal/domain"
	5	+	"github.com/bnema/vev/internal/usecase/layout"
	6	+	themeui "github.com/bnema/vev/internal/usecase/theme"
	7	+	"github.com/bnema/vev/internal/usecase/ui"
	8	+	"github.com/bnema/vev/pkg/renderer"
	9	+	)
	10	+
	11	+	// Legacy composition adapters are test-only. Production rendering is owned by
	12	+	// capturePrimaryRenderState -> composeFrame -> emitFrame.
	13	+	// copyTargetRectLocked maps a captured copy source into the already-composed
	14	+	// client frame. The caller holds tb.mu, preserving the layout/floating
	15	+	// snapshot while the rectangle is chosen.
	16	+	func copyTargetRectLocked(layoutSnap tabLayoutSnapshot, contentArea domain.Rect, p, floating *pane, hasFloating bool, floatingFrameGeometry floatingGeometry) domain.Rect {
	17	+	if p == nil {
	18	+	return domain.Rect{}
	19	+	}
	20	+	if hasFloating && p == floating {
	21	+	return floatingFrameGeometry.Inner
	22	+	}
	23	+	for _, placement := range layoutSnap.placements {
	24	+	if placement.ID == p.id && !placement.Collapsed {
	25	+	r := placement.Content
	26	+	r.X += contentArea.X
	27	+	r.Y += contentArea.Y
	28	+	return r
	29	+	}
	30	+	}
	31	+	p.mu.Lock()
	32	+	width, height := p.screen.Frame.Width, p.screen.Frame.Height
	33	+	p.mu.Unlock()
	34	+	return domain.Rect{X: contentArea.X, Y: contentArea.Y, Width: min(width, contentArea.Width), Height: min(height, contentArea.Height)}
	35	+	}
	36	+
	37	+	type composedFrameCache struct {
	38	+	frame renderer.Frame
	39	+	valid bool
	40	+	layoutSnap tabLayoutSnapshot
	41	+	titleGenerations map[layout.PaneID]uint64
	42	+	floating *pane
	43	+	floatingFrame renderer.Frame
	44	+	floatingGeneration uint64
	45	+	floatingGeometry floatingGeometry
	46	+	floatingTitleGeneration uint64
	47	+	floatingCaptured capturedPaneRenderState
	48	+	}
	49	+
	50	+	func composeClientFrame(sess session, tb tab, full bool, rightStatus string, caches ...*barCache) (renderer.Frame, []renderer.Damage) {
	51	+	return composeClientFrameWithState(barState{status: sess.statusSegments(true), copyFeedback: rightStatus}, tb, full, caches...)
	52	+	}
	53	+
	54	+	func composeClientFrameWithState(bars barState, tb tab, full bool, caches ...barCache) (renderer.Frame, []renderer.Damage) {
	55	+	return composeClientFrameWithLayout(bars, tb, full, solveTabLayoutLocked(tb), caches...)
	56	+	}
	57	+
	58	+	func composeClientFrameWithLayout(bars barState, tb tab, full bool, layoutSnap tabLayoutSnapshot, caches ...barCache) (renderer.Frame, []renderer.Damage) {
	59	+	var cache *barCache
	60	+	if len(caches) > 0 {
	61	+	cache = caches[0]
	62	+	}
	63	+	return composeClientFrameWithLayoutCached(bars, tb, full, layoutSnap, cache, nil)
	64	+	}
	65	+
	66	+	func composeClientFrameWithLayoutCached(bars barState, tb tab, full bool, layoutSnap tabLayoutSnapshot, cache barCache, composed *composedFrameCache) (renderer.Frame, []renderer.Damage) {
	67	+	return composeClientFrameWithLayoutCachedOptions(bars, tb, full, layoutSnap, cache, composed, false)
	68	+	}
	69	+
	70	+	func composeClientFrameWithLayoutCachedConsumeDamage(bars barState, tb tab, full bool, layoutSnap tabLayoutSnapshot, cache barCache, composed *composedFrameCache) (renderer.Frame, []renderer.Damage) {
	71	+	return composeClientFrameWithLayoutCachedOptions(bars, tb, full, layoutSnap, cache, composed, true)
	72	+	}
	73	+
	74	+	func composeClientFrameWithLayoutCachedOptions(bars barState, tb tab, full bool, layoutSnap tabLayoutSnapshot, cache barCache, composed *composedFrameCache, consumeDamage bool) (renderer.Frame, []renderer.Damage) {
	75	+	p := tb.focusedPane()
	76	+	styles := newThemeStyles(bars.theme)
	77	+	if p == nil {
	78	+	width, screenRows := tb.size.Cols, tb.size.Rows
	79	+	if width <= 0 \|\| screenRows <= 0 {
	80	+	return renderer.NewFrame(0, 0), nil
	81	+	}
	82	+	return renderer.NewFrame(width, screenRows+2), nil
	83	+	}
	84	+	p.mu.Lock()
	85	+	width, screenRows := p.screen.Frame.Width, p.screen.Frame.Height
	86	+	p.mu.Unlock()
	87	+	if tb.size.Valid() {
	88	+	width, screenRows = tb.size.Cols, tb.size.Rows
	89	+	}
	90	+	cacheValid := composed != nil && composed.valid && composed.frame.Width == width && composed.frame.Height == screenRows+2
	91	+	layoutSame := composed == nil \|\| (cacheValid && sameTabLayoutSnapshot(composed.layoutSnap, layoutSnap))
	92	+	var frame renderer.Frame
	93	+	if cacheValid {
	94	+	frame = composed.frame
	95	+	} else {
	96	+	frame = renderer.NewFrame(width, screenRows+2)
	97	+	if composed != nil {
	98	+	full = true
	99	+	}
	100	+	}
	101	+	if !layoutSame {
	102	+	full = true
	103	+	}
	104	+	contentArea := domain.Rect{Y: 1, Width: width, Height: screenRows}
	105	+	if cacheValid && !layoutSame {
	106	+	clearFrameRect(frame, contentArea)
	107	+	}
	108	+	topBar := frame.Row(0)
	109	+	drawTopBarSnapshot(topBar, bars.status, bars.attentionFrame, bars.topRight, styles)
	110	+	var titleGenerations map[layout.PaneID]uint64
	111	+	if composed != nil {
	112	+	if composed.titleGenerations == nil {
	113	+	composed.titleGenerations = make(map[layout.PaneID]uint64)
	114	+	}
	115	+	titleGenerations = composed.titleGenerations
	116	+	}
	117	+	contentDamage := composeTabFrameIntoWithLayoutOptions(tb, frame, contentArea, bars.theme, layoutSnap, cacheValid && layoutSame, consumeDamage, titleGenerations)
	118	+	bottomBar := frame.Row(screenRows + 1)
	119	+	drawStatusBarState(bottomBar, bars, styles)
	120	+	if composed != nil {
	121	+	composed.frame = frame
	122	+	composed.valid = true
	123	+	composed.layoutSnap = layoutSnap
	124	+	}
	125	+	if full {
	126	+	if cache != nil {
	127	+	cache.capture(topBar, bottomBar)
	128	+	}
	129	+	return frame, []renderer.Damage{renderer.FullRedraw()}
	130	+	}
	131	+	damage := translateDamage(contentDamage, 0, 1)
	132	+	if cache == nil \|\| !sameCells(cache.top, topBar) {
	133	+	damage = append(damage, renderer.Damage{Kind: renderer.DamageText, X: 0, Y: 0, Width: width, Height: 1})
	134	+	}
	135	+	if cache == nil \|\| !sameCells(cache.bottom, bottomBar) {
	136	+	damage = append(damage, renderer.Damage{Kind: renderer.DamageText, X: 0, Y: screenRows + 1, Width: width, Height: 1})
	137	+	}
	138	+	if cache != nil {
	139	+	cache.capture(topBar, bottomBar)
	140	+	}
	141	+	return frame, damage
	142	+	}
	143	+
	144	+	func composeTabFrame(tb *tab, area domain.Rect, theme themeui.Theme) (renderer.Frame, []renderer.Damage) {
	145	+	return composeTabFrameWithLayout(tb, area, theme, tabLayoutSnapshot{})
	146	+	}
	147	+
	148	+	func composeTabFrameWithLayout(tb *tab, area domain.Rect, theme themeui.Theme, layoutSnap tabLayoutSnapshot) (renderer.Frame, []renderer.Damage) {
	149	+	frame := renderer.NewFrame(area.Width, area.Height)
	150	+	damage := composeTabFrameIntoWithLayout(tb, frame, area, theme, layoutSnap, false)
	151	+	return frame, damage
	152	+	}
	153	+
	154	+	func composeTabFrameIntoWithLayout(tb *tab, frame renderer.Frame, area domain.Rect, theme themeui.Theme, layoutSnap tabLayoutSnapshot, cacheValid bool) []renderer.Damage {
	155	+	return composeTabFrameIntoWithLayoutOptions(tb, frame, area, theme, layoutSnap, cacheValid, false, nil)
	156	+	}
	157	+
	158	+	func composeTabFrameIntoWithLayoutOptions(tb *tab, frame renderer.Frame, area domain.Rect, theme themeui.Theme, layoutSnap tabLayoutSnapshot, cacheValid bool, consumeDamage bool, titleGenerations map[layout.PaneID]uint64) []renderer.Damage {
	159	+	contentArea := domain.Rect{Width: area.Width, Height: area.Height}
	160	+	root := tb.tree.Root
	161	+	placements, ok := layoutSnap.placements, layoutSnap.ok && layoutSnap.root == root && layoutSnap.area == contentArea
	162	+	if !ok {
	163	+	placements, ok = layout.Solve(root, contentArea)
	164	+	}
	165	+	var fallback *pane
	166	+	if !ok {
	167	+	fallback = tb.focusedPane()
	168	+	if fallback == nil {
	169	+	return nil
	170	+	}
	171	+	placements = []layout.Placement{{ID: fallback.id, Content: contentArea}}
	172	+	}
	173	+	// A valid cache has the same layout, so its title IDs remain valid. Reset
	174	+	// the existing cache only when rebuilding after layout or frame churn.
	175	+	if titleGenerations != nil && !cacheValid {
	176	+	clear(titleGenerations)
	177	+	}
	178	+	if ok && !cacheValid {
	179	+	drawDividers(frame, root, area, themeui.DimStyle(newThemeStyles(theme).border, theme))
	180	+	}
	181	+	var damage []renderer.Damage
	182	+	for _, pl := range placements {
	183	+	p := tb.panes[pl.ID]
	184	+	if p == nil && fallback != nil && pl.ID == fallback.id {
	185	+	p = fallback
	186	+	}
	187	+	if p == nil {
	188	+	continue
	189	+	}
	190	+	focused := tb.tree.Focus == pl.ID
	191	+	pl = offsetPlacement(pl, area.X, area.Y)
	192	+	if pl.TitleBar.Height > 0 {
	193	+	generation := drawPaneTitleBar(frame, pl, p, focused, theme)
	194	+	if cacheValid && (titleGenerations == nil \|\| titleGenerations[pl.ID] != generation) {
	195	+	titleDamage := pl.TitleBar
	196	+	titleDamage.X -= area.X
	197	+	titleDamage.Y -= area.Y
	198	+	damage = append(damage, renderer.Damage{Kind: renderer.DamageText, X: titleDamage.X, Y: titleDamage.Y, Width: titleDamage.Width, Height: titleDamage.Height})
	199	+	}
	200	+	if titleGenerations != nil {
	201	+	titleGenerations[pl.ID] = generation
	202	+	}
	203	+	}
	204	+	if pl.Collapsed \|\| pl.Content.Width <= 0 \|\| pl.Content.Height <= 0 {
	205	+	if consumeDamage {
	206	+	p.mu.Lock()
	207	+	captured := capturePaneRenderStateLocked(p, domain.Rect{})
	208	+	p.screen.AcknowledgeDamage(captured.damageGeneration)
	209	+	p.mu.Unlock()
	210	+	}
	211	+	continue
	212	+	}
	213	+	var paneFrame renderer.Frame
	214	+	var paneDamage []renderer.Damage
	215	+	if consumeDamage {
	216	+	p.mu.Lock()
	217	+	captured := capturePaneRenderStateLocked(p, pl.Content)
	218	+	p.screen.AcknowledgeDamage(captured.damageGeneration)
	219	+	p.mu.Unlock()
	220	+	paneFrame, paneDamage = captured.frame, captured.damage
	221	+	} else {
	222	+	p.mu.Lock()
	223	+	paneDamage = p.screen.Damage()
	224	+	if !cacheValid \|\| len(paneDamage) > 0 {
	225	+	blitPaneFrame(frame, pl.Content, p.screen.Frame, !focused, theme)
	226	+	}
	227	+	for _, d := range paneDamage {
	228	+	localContent := pl.Content
	229	+	localContent.X -= area.X
	230	+	localContent.Y -= area.Y
	231	+	damage = append(damage, translatePaneDamage(d, localContent, contentArea)...)
	232	+	}
	233	+	p.mu.Unlock()
	234	+	continue
	235	+	}
	236	+	if !cacheValid \|\| len(paneDamage) > 0 {
	237	+	blitPaneFrame(frame, pl.Content, paneFrame, !focused, theme)
	238	+	}
	239	+	for _, d := range paneDamage {
	240	+	localContent := pl.Content
	241	+	localContent.X -= area.X
	242	+	localContent.Y -= area.Y
	243	+	damage = append(damage, translatePaneDamage(d, localContent, contentArea)...)
	244	+	}
	245	+	}
	246	+	return damage
	247	+	}
	248	+
	249	+	func sameTabLayoutSnapshot(a, b tabLayoutSnapshot) bool {
	250	+	return a.ok == b.ok && a.root == b.root && a.fingerprint == b.fingerprint && a.area == b.area && a.focus == b.focus
	251	+	}
	252	+
	253	+	func clearFrameRect(frame renderer.Frame, r domain.Rect) {
	254	+	blank := renderer.BlankCell()
	255	+	for y := r.Y; y < r.Y+r.Height && y < frame.Height; y++ {
	256	+	for x := r.X; x < r.X+r.Width && x < frame.Width; x++ {
	257	+	frame.Set(x, y, blank)
	258	+	}
	259	+	}
	260	+	}
	261	+
	262	+	func drawPaneTitleBar(frame renderer.Frame, pl layout.Placement, p *pane, focused bool, theme themeui.Theme) uint64 {
	263	+	styles := newThemeStyles(theme)
	264	+	style := styles.border
	265	+	if focused {
	266	+	style = styles.statusBar
	267	+	} else {
	268	+	style = themeui.DimStyle(style, theme)
	269	+	}
	270	+	for x := pl.TitleBar.X; x < pl.TitleBar.X+pl.TitleBar.Width && x < frame.Width; x++ {
	271	+	frame.Set(x, pl.TitleBar.Y, renderer.Cell{Rune: ' ', Style: style})
	272	+	}
	273	+	p.mu.Lock()
	274	+	title := p.displayTitleLocked()
	275	+	generation := p.title.generation
	276	+	p.mu.Unlock()
	277	+	ui.DrawText(frame, pl.TitleBar.X, pl.TitleBar.Y, pl.TitleBar.X+pl.TitleBar.Width, title, style)
	278	+	return generation
	279	+	}

A internal/usecase/daemon/render_pipeline.go +376 -0

@@ -0,0 +1,376 @@
	1	+	package daemon
	2	+
	3	+	import (
	4	+	"errors"
	5	+	"strconv"
	6	+
	7	+	"github.com/bnema/vev/internal/domain"
	8	+	"github.com/bnema/vev/internal/ports"
	9	+	"github.com/bnema/vev/internal/usecase/layout"
	10	+	"github.com/bnema/vev/internal/usecase/palette"
	11	+	"github.com/bnema/vev/internal/usecase/picker"
	12	+	themeui "github.com/bnema/vev/internal/usecase/theme"
	13	+	"github.com/bnema/vev/internal/usecase/ui"
	14	+	"github.com/bnema/vev/pkg/renderer"
	15	+	)
	16	+
	17	+	// composeCacheInput is an attachment-local value snapshot. composeFrame never
	18	+	// retains or accesses the attachment that owns it.
	19	+	// renderStageHooks are optional observability hooks. They are invoked only at
	20	+	// completed production boundaries: capture and compose after their result is
	21	+	// available, and emit after prepare plus transport success (including a
	22	+	// successful no-byte emission). Failed attempts count for capture/compose but
	23	+	// never for emit.
	24	+	type renderStageHooks struct {
	25	+	capture func()
	26	+	compose func()
	27	+	emit func()
	28	+	}
	29	+
	30	+	type composeCacheInput struct {
	31	+	valid bool
	32	+	frame renderer.Frame
	33	+	layoutFingerprint string
	34	+	titleGenerations map[layout.PaneID]uint64
	35	+	damage []renderer.Damage
	36	+	floatingVisible bool
	37	+	floatingGeneration uint64
	38	+	floatingGeometry floatingGeometry
	39	+	floatingTitleGeneration uint64
	40	+	bars barCache
	41	+	theme themeui.Theme
	42	+	}
	43	+
	44	+	type composedRenderFrame struct {
	45	+	frame renderer.Frame
	46	+	damage []renderer.Damage
	47	+	cursor cursorOut
	48	+	cache composeCacheInput
	49	+	reset bool
	50	+	}
	51	+
	52	+	// composeFrame is pure with respect to daemon ownership: it consumes only the
	53	+	// capture, the last committed cache, and an independent attachment-owned
	54	+	// scratch cache. It returns a replacement cache without mutating committed.
	55	+	func composeFrame(state capturedRenderState, in composeCacheInput, scratchIn ...composeCacheInput) composedRenderFrame {
	56	+	scratch := composeCacheInput{}
	57	+	if len(scratchIn) > 0 {
	58	+	scratch = scratchIn[0]
	59	+	}
	60	+	width, rows := state.layout.area.Width, state.layout.area.Height
	61	+	if width <= 0 \|\| rows < 0 {
	62	+	return composedRenderFrame{frame: renderer.NewFrame(0, 0), cursor: cursorOut{hidden: true}, reset: state.reset}
	63	+	}
	64	+	// Compose into the alternate attachment-owned buffer. The committed frame is
	65	+	// copied before incremental drawing so prepare/send failures cannot change it.
	66	+	canReuseFrame := scratch.frame.Width == width && scratch.frame.Height == rows+2
	67	+	frame := scratch.frame
	68	+	if !canReuseFrame {
	69	+	frame = renderer.NewFrame(width, rows+2)
	70	+	}
	71	+	if in.frame.Width == width && in.frame.Height == rows+2 {
	72	+	for y := 0; y < frame.Height; y++ {
	73	+	copy(frame.Row(y), in.frame.Row(y))
	74	+	}
	75	+	}
	76	+	styles := newThemeStyles(state.theme)
	77	+	drawTopBarSnapshot(frame.Row(0), state.bars.status, state.bars.attentionFrame, state.bars.topRight, styles)
	78	+	drawStatusBarState(frame.Row(rows+1), state.bars, styles)
	79	+	content := domain.Rect{Y: 1, Width: width, Height: rows}
	80	+	if state.layout.valid && state.layout.root != nil {
	81	+	drawDividers(frame, state.layout.root, content, themeui.DimStyle(styles.border, state.theme))
	82	+	}
	83	+
	84	+	full := state.reset \|\| !in.valid \|\| in.frame.Width != width \|\| in.frame.Height != rows+2 \|\| in.layoutFingerprint != state.layout.fingerprint \|\| in.theme != state.theme \|\| in.floatingVisible != state.floating.visible
	85	+	titles := scratch.titleGenerations
	86	+	if titles == nil {
	87	+	titles = make(map[layout.PaneID]uint64, len(state.panes))
	88	+	}
	89	+	damage := scratch.damage[:0]
	90	+	for _, pane := range state.panes {
	91	+	pl := offsetPlacement(pane.placement, 0, 1)
	92	+	if pl.TitleBar.Height > 0 {
	93	+	drawCapturedPaneTitleBar(frame, pl, pane.title, pane.focused, state.theme)
	94	+	if !full && in.titleGenerations[pane.id] != pane.titleGeneration {
	95	+	damage = append(damage, renderer.Damage{Kind: renderer.DamageText, X: pl.TitleBar.X, Y: pl.TitleBar.Y, Width: pl.TitleBar.Width, Height: pl.TitleBar.Height})
	96	+	}
	97	+	titles[pane.id] = pane.titleGeneration
	98	+	}
	99	+	if !pl.Collapsed && pl.Content.Width > 0 && pl.Content.Height > 0 && (full \|\| len(pane.damage) > 0) {
	100	+	blitPaneFrame(frame, pl.Content, pane.frame, !pane.focused, state.theme)
	101	+	}
	102	+	for _, d := range pane.damage {
	103	+	if d.Kind != renderer.DamageFullRedraw {
	104	+	d.Y++
	105	+	}
	106	+	damage = append(damage, d)
	107	+	}
	108	+	}
	109	+	// Keep the committed cache unadorned: floating composition clones this
	110	+	// base, so closing or moving a popup cannot retain dimmed/bordered cells.
	111	+	baseFrame := frame
	112	+	if state.floating.visible {
	113	+	var floatingDamage []renderer.Damage
	114	+	frame, floatingDamage = composeCapturedFloatingFrame(baseFrame, damage, state.floating, content, state.layout, state.theme, in, full \|\| state.overlays.active())
	115	+	damage = floatingDamage
	116	+	}
	117	+	if !full {
	118	+	if !sameCells(in.bars.top, frame.Row(0)) {
	119	+	damage = append(damage, renderer.Damage{Kind: renderer.DamageText, X: 0, Y: 0, Width: width, Height: 1})
	120	+	}
	121	+	if !sameCells(in.bars.bottom, frame.Row(rows+1)) {
	122	+	damage = append(damage, renderer.Damage{Kind: renderer.DamageText, X: 0, Y: rows + 1, Width: width, Height: 1})
	123	+	}
	124	+	}
	125	+	if state.overlays.active() {
	126	+	// Without a floating frame, overlay composition would otherwise mutate
	127	+	// the base cache in place. Floating composition already owns a clone.
	128	+	if !state.floating.visible {
	129	+	baseFrame = frame.Clone()
	130	+	}
	131	+	frame, damage = composeCapturedOverlays(state, frame, damage, content)
	132	+	}
	133	+	if full \|\| state.overlays.active() {
	134	+	damage = []renderer.Damage{renderer.FullRedraw()}
	135	+	}
	136	+	cursorInputs := state.cursor
	137	+	cursorInputs.hiddenByOverlay = cursorInputs.hiddenByOverlay \|\| state.overlays.active()
	138	+	cursor := desiredCapturedCursor(cursorInputs)
	139	+	outCache := composeCacheInput{valid: !state.overlays.active(), frame: baseFrame, layoutFingerprint: state.layout.fingerprint, theme: state.theme, titleGenerations: titles, damage: damage, floatingVisible: state.floating.visible, floatingGeneration: state.floating.generation, floatingGeometry: state.floating.geometry.translate(content.X, content.Y), floatingTitleGeneration: state.floating.titleGeneration, bars: scratch.bars}
	140	+	outCache.bars.capture(baseFrame.Row(0), baseFrame.Row(rows+1))
	141	+	return composedRenderFrame{frame: frame, damage: damage, cursor: cursor, cache: outCache, reset: state.reset \|\| state.overlays.active()}
	142	+	}
	143	+
	144	+	func drawCapturedPaneTitleBar(frame renderer.Frame, pl layout.Placement, title string, focused bool, theme themeui.Theme) {
	145	+	styles := newThemeStyles(theme)
	146	+	style := styles.border
	147	+	if focused {
	148	+	style = styles.statusBar
	149	+	} else {
	150	+	style = themeui.DimStyle(style, theme)
	151	+	}
	152	+	for x := pl.TitleBar.X; x < pl.TitleBar.X+pl.TitleBar.Width && x < frame.Width; x++ {
	153	+	frame.Set(x, pl.TitleBar.Y, renderer.Cell{Rune: ' ', Style: style})
	154	+	}
	155	+	ui.DrawText(frame, pl.TitleBar.X, pl.TitleBar.Y, pl.TitleBar.X+pl.TitleBar.Width, title, style)
	156	+	}
	157	+
	158	+	func desiredCapturedCursor(c capturedCursorInputs) cursorOut {
	159	+	if !c.renderable \|\| c.hiddenByOverlay \|\| !c.visible {
	160	+	return cursorOut{hidden: true}
	161	+	}
	162	+	style := c.style
	163	+	if !c.hasStyle {
	164	+	style = 1
	165	+	}
	166	+	return cursorOut{row: c.content.Y + 1 + c.row, col: c.content.X + c.col, style: style, hasStyle: true}
	167	+	}
	168	+
	169	+	func captureOverlayLayers(state capturedRenderState, snap overlayRenderSnapshot, paletteCfg domain.PaletteConfig) {
	170	+	if state == nil \|\| snap == nil {
	171	+	return
	172	+	}
	173	+	o := &state.overlays
	174	+	o.copyActive, o.copySearchActive, o.pickerActive, o.paletteActive, o.promptActive = snap.copyActive, snap.copySearchModel != nil, snap.pickerActive, snap.paletteActive, snap.promptActive
	175	+	o.copyMode, o.copySnapshot = snap.copyMode, snap.copySnapshot
	176	+	if snap.copyPane != nil {
	177	+	o.copyPaneID = snap.copyPane.id
	178	+	}
	179	+	styles := newThemeStyles(state.theme)
	180	+	size := domain.Size{Cols: state.layout.area.Width, Rows: state.layout.area.Height + 2}
	181	+	if snap.copySearchModel != nil {
	182	+	o.copySearch.modal = copySearchModal
	183	+	o.copySearch.inner = snap.copySearchModel.Render(rectSize(copySearchModal.Inner(size)), styles.selection)
	184	+	}
	185	+	if snap.pickerActive && snap.pickerModel != nil {
	186	+	o.picker.modal = pickerModal
	187	+	renderStyles := picker.RenderStyles{Selection: styles.selection, SelectionName: styles.pickerSelectionName, SelectionMuted: styles.pickerSelectionMuted, Name: styles.pickerName, Detail: styles.paletteDesc, Base: renderer.DefaultStyle()}
	188	+	o.picker.inner = snap.pickerModel.Render(rectSize(pickerModal.Inner(size)), state.preview, renderStyles)
	189	+	}
	190	+	if snap.paletteActive && snap.paletteModel != nil {
	191	+	o.palette.modal = paletteModalFor(size, paletteCfg)
	192	+	guidance := ""
	193	+	if snap.paletteHints != nil {
	194	+	guidance = snap.paletteHints.Feedback
	195	+	}
	196	+	o.paletteGuidance = guidance
	197	+	o.palette.inner = snap.paletteModel.Render(rectSize(o.palette.modal.Inner(size)), palette.RenderOptions{Styles: palette.RenderStyles{Selection: styles.selection, Description: styles.paletteDesc}, Guidance: guidance})
	198	+	}
	199	+	if snap.promptActive && snap.promptModel != nil {
	200	+	o.prompt.modal = promptModalFor(snap.promptModel.Title())
	201	+	o.prompt.inner = snap.promptModel.Render(rectSize(o.prompt.modal.Inner(size)), styles.accent)
	202	+	}
	203	+	state.cursor.hiddenByOverlay = o.active()
	204	+	}
	205	+
	206	+	func composeCapturedOverlays(state capturedRenderState, frame renderer.Frame, damage []renderer.Damage, content domain.Rect) (renderer.Frame, []renderer.Damage) {
	207	+	o := state.overlays
	208	+	if o.copyActive {
	209	+	target := domain.Rect{}
	210	+	if state.floating.visible && (o.copyPaneID == "" \|\| o.copyPaneID == state.floating.pane.id) {
	211	+	target = state.floating.geometry.translate(content.X, content.Y).Inner
	212	+	}
	213	+	if target.Width == 0 {
	214	+	id := o.copyPaneID
	215	+	if id == "" {
	216	+	id = state.layout.focus
	217	+	}
	218	+	for _, p := range state.panes {
	219	+	if p.id == id && !p.placement.Collapsed {
	220	+	target = p.placement.Content
	221	+	target.Y += content.Y
	222	+	break
	223	+	}
	224	+	}
	225	+	}
	226	+	frame, damage = composeCopyClientFrame(o.copyMode, o.copySnapshot, target, frame, state.bars)
	227	+	}
	228	+	legacy := tabLayoutSnapshot{placements: state.layout.placements, area: state.layout.area, focus: state.layout.focus, ok: state.layout.valid}
	229	+	if o.paletteActive && !state.floating.visible {
	230	+	(overlayBackdrop{DimPaneContents: true}).apply(frame, content, legacy, state.theme)
	231	+	}
	232	+	for _, modal := range []capturedModal{o.copySearch, o.picker, o.palette, o.prompt} {
	233	+	if modal.inner.Width == 0 && modal.inner.Height == 0 {
	234	+	continue
	235	+	}
	236	+	inner := modal.modal.Composite(frame, newThemeStyles(state.theme).border)
	237	+	for y := range min(inner.Height, modal.inner.Height) {
	238	+	copy(frame.Row(inner.Y + y)[inner.X:inner.X+min(inner.Width, modal.inner.Width)], modal.inner.Row(y)[:min(inner.Width, modal.inner.Width)])
	239	+	}
	240	+	damage = []renderer.Damage{renderer.FullRedraw()}
	241	+	}
	242	+	return frame, damage
	243	+	}
	244	+
	245	+	func (ac *attachedClient) encodeCursorTail(desired cursorOut, force bool) []byte {
	246	+	changed := force \|\| !ac.lastCursor.valid \|\| ac.lastCursor.hidden != desired.hidden \|\| ac.lastCursor.row != desired.row \|\| ac.lastCursor.col != desired.col \|\| ac.lastCursor.style != desired.style \|\| ac.lastCursor.hasStyle != desired.hasStyle
	247	+	if !changed {
	248	+	return nil
	249	+	}
	250	+	prev := ac.lastCursor
	251	+	ac.lastCursor = desired
	252	+	ac.lastCursor.valid = true
	253	+	if desired.hidden {
	254	+	return []byte("\x1b[?25l")
	255	+	}
	256	+	var b []byte
	257	+	b = append(b, "\x1b["...)
	258	+	b = strconv.AppendInt(b, int64(desired.row+1), 10)
	259	+	b = append(b, ';')
	260	+	b = strconv.AppendInt(b, int64(desired.col+1), 10)
	261	+	b = append(b, 'H')
	262	+	if !prev.valid \|\| prev.hidden \|\| prev.style != desired.style \|\| prev.hasStyle != desired.hasStyle {
	263	+	b = append(b, "\x1b["...)
	264	+	b = strconv.AppendInt(b, int64(desired.style), 10)
	265	+	b = append(b, " q"...)
	266	+	}
	267	+	b = append(b, "\x1b[?25h"...)
	268	+	return b
	269	+	}
	270	+
	271	+	// commitDamageReceipts acknowledges only snapshots that reached the client.
	272	+	// sendMu serializes attachment transactions; panes are locked individually so
	273	+	// no pane lock is held while another is acquired. A stale generation is
	274	+	// intentionally acknowledged too: Screen conservatively leaves FullRedraw for
	275	+	// the next transaction.
	276	+	func commitDamageReceipts(receipts []damageReceipt) {
	277	+	for _, receipt := range receipts {
	278	+	if receipt.pane == nil {
	279	+	continue
	280	+	}
	281	+	receipt.pane.mu.Lock()
	282	+	receipt.pane.screen.AcknowledgeDamage(receipt.generation)
	283	+	receipt.pane.mu.Unlock()
	284	+	}
	285	+	}
	286	+
	287	+	// emitFrame is the sole side-effecting half of the pipeline. The caller holds
	288	+	// sendMu for the complete capture/compose/emit transaction.
	289	+	func (d Daemon) emitFrame(sess session, ac attachedClient, state capturedRenderState, composed composedRenderFrame, batches ...*runtimeMarkBatch) bool {
	290	+	var ownedMarks runtimeMarkBatch
	291	+	var marks *runtimeMarkBatch
	292	+	if len(batches) != 0 {
	293	+	marks = batches[0]
	294	+	} else {
	295	+	ownedMarks = d.newRuntimeMarkBatch()
	296	+	marks = &ownedMarks
	297	+	// Direct test callers retain the same guarantee as paint: observer I/O
	298	+	// follows emitFrame's release of attachment ownership.
	299	+	defer marks.flush()
	300	+	}
	301	+	if sess == nil \|\| ac == nil \|\| state == nil {
	302	+	if ac != nil {
	303	+	ac.sendMu.Unlock()
	304	+	}
	305	+	return false
	306	+	}
	307	+	sess.mu.Lock()
	308	+	owned := sess.client == ac
	309	+	sess.mu.Unlock()
	310	+	if !owned \|\| state.attachment != ac \|\| ac.currentSession() != sess {
	311	+	ac.sendMu.Unlock()
	312	+	return false
	313	+	}
	314	+	if state.lease != nil {
	315	+	rc := sess.renderCoordinator()
	316	+	if rc == nil \|\| state.lease.attachment != ac \|\| !rc.leaseCurrent(state.lease, true) {
	317	+	ac.sendMu.Unlock()
	318	+	return false
	319	+	}
	320	+	}
	321	+	ac.sess.mu.Lock()
	322	+	if ac.sess.v != sess {
	323	+	ac.sess.mu.Unlock()
	324	+	ac.sendMu.Unlock()
	325	+	return false
	326	+	}
	327	+	endDiff := marks.span(ports.RuntimeDiffStart, ports.RuntimeDiffEnd, 0)
	328	+	prepared, err := ac.output.prepare(composed.frame, composed.damage, composed.reset)
	329	+	endDiff(0, err == nil)
	330	+	if err != nil {
	331	+	ac.sess.mu.Unlock()
	332	+	ac.sendMu.Unlock()
	333	+	d.log.Error("render draw failed", "err", err, "session", sess.name)
	334	+	return true
	335	+	}
	336	+	data := append([]byte(nil), prepared.data...)
	337	+	data = append(data, ac.encodeCursorTail(composed.cursor, len(data) > 0)...)
	338	+	var sendTr ports.Transport
	339	+	var sendErr error
	340	+	if len(data) > 0 {
	341	+	sendTr = ac.transport()
	342	+	endEmit := marks.span(ports.RuntimeEmitStart, ports.RuntimeEmitEnd, uint64(len(data)))
	343	+	if sendTr == nil {
	344	+	sendErr = errors.New("client transport is nil")
	345	+	} else {
	346	+	send := sendTr.Send
	347	+	if async, ok := sendTr.(ports.AsyncTransport); ok {
	348	+	send = async.SendAsync
	349	+	}
	350	+	sendErr = prepared.send(data, ac.echoAck.Load(), send)
	351	+	}
	352	+	endEmit(uint64(len(data)), sendErr == nil)
	353	+	}
	354	+	if sendErr == nil {
	355	+	// Publish only after output preparation and transport emission both
	356	+	// succeed. Receipt acknowledgement follows after releasing the attachment
	357	+	// session guard, while sendMu still owns this transaction.
	358	+	ac.pipelineScratch = ac.pipelineCache
	359	+	ac.pipelineCache = composed.cache
	360	+	}
	361	+	ac.sess.mu.Unlock()
	362	+	if sendErr == nil {
	363	+	// A successful no-byte emission also commits: its renderer shadow still
	364	+	// represents the captured frame. Lock panes only under sendMu and with no
	365	+	// session guard held.
	366	+	commitDamageReceipts(state.receipts)
	367	+	if ac.renderStages.emit != nil {
	368	+	ac.renderStages.emit()
	369	+	}
	370	+	}
	371	+	ac.sendMu.Unlock()
	372	+	if sendErr != nil {
	373	+	d.detachOnSendError(sess, ac, sendTr)
	374	+	}
	375	+	return true
	376	+	}

A internal/usecase/daemon/render_pipeline_cache_test.go +116 -0

@@ -0,0 +1,116 @@
	1	+	package daemon
	2	+
	3	+	import (
	4	+	"errors"
	5	+	"io"
	6	+	"testing"
	7	+
	8	+	"github.com/bnema/vev/internal/domain"
	9	+	"github.com/bnema/vev/internal/ports"
	10	+	"github.com/bnema/vev/internal/usecase/layout"
	11	+	"github.com/bnema/vev/pkg/renderer"
	12	+	"github.com/stretchr/testify/require"
	13	+	)
	14	+
	15	+	type cacheFailTransport struct{}
	16	+
	17	+	func (cacheFailTransport) Send(ports.Frame) error { return errors.New("send failed") }
	18	+	func (cacheFailTransport) Recv() (ports.Frame, error) { return ports.Frame{}, io.EOF }
	19	+	func (cacheFailTransport) Close() error { return nil }
	20	+
	21	+	func TestPipelineCachePublishesOnlyAfterEmission(t *testing.T) {
	22	+	for _, failure := range []struct {
	23	+	name string
	24	+	apply func(composedRenderFrame, attachedClient)
	25	+	}{
	26	+	{
	27	+	name: "prepare",
	28	+	apply: func(composed composedRenderFrame, _ attachedClient) {
	29	+	// Frame validation fails in output.prepare after composition.
	30	+	composed.frame = renderer.Frame{Width: 1}
	31	+	},
	32	+	},
	33	+	{
	34	+	name: "send",
	35	+	apply: func(_ composedRenderFrame, ac attachedClient) { ac.replaceTransport(cacheFailTransport{}) },
	36	+	},
	37	+	} {
	38	+	t.Run(failure.name, func(t *testing.T) {
	39	+	d, sess, ac, sends := newManualSessionWithPTYs(t)
	40	+	healthy := ac.transport()
	41	+	committed := composeFrame(cacheState("committed", 1), composeCacheInput{})
	42	+	ac.pipelineCache = committed.cache
	43	+	before := cloneComposeCache(ac.pipelineCache)
	44	+	pending := composeFrame(cacheState("next", 2), ac.pipelineCache, ac.pipelineScratch)
	45	+	failure.apply(&pending, ac)
	46	+
	47	+	state := cacheState("next", 2)
	48	+	state.attachment = ac
	49	+	ac.sendMu.Lock()
	50	+	require.True(t, d.emitFrame(sess, ac, &state, pending))
	51	+	require.Equal(t, before, cloneComposeCache(ac.pipelineCache), "failed emission must not publish any composed cache backing storage")
	52	+
	53	+	// A send error detaches the failed link. Re-own this test attachment with
	54	+	// its original healthy transport, then retry the pending state.
	55	+	if failure.name == "send" {
	56	+	sess.mu.Lock()
	57	+	sess.client = ac
	58	+	sess.mu.Unlock()
	59	+	ac.setSession(sess)
	60	+	ac.replaceTransport(healthy)
	61	+	}
	62	+	pending = composeFrame(cacheState("next", 2), ac.pipelineCache, ac.pipelineScratch)
	63	+	state = cacheState("next", 2)
	64	+	state.attachment = ac
	65	+	ac.sendMu.Lock()
	66	+	require.True(t, d.emitFrame(sess, ac, &state, pending))
	67	+	require.Equal(t, pending.cache, ac.pipelineCache)
	68	+	frame := <-sends
	69	+	output, err := ports.UnmarshalOutput(frame.Payload)
	70	+	require.NoError(t, err)
	71	+	require.Contains(t, string(output.Data), "next", "the retry must emit state retained after the failed emission")
	72	+	})
	73	+	}
	74	+	}
	75	+
	76	+	func TestComposeFrameClearsFloatingFrameWhenItCloses(t *testing.T) {
	77	+	visible := cacheState("base", 1)
	78	+	committed := composeFrame(visible, composeCacheInput{})
	79	+
	80	+	hidden := visible
	81	+	hidden.floating = capturedFloatingRenderState{}
	82	+	hidden.panes[0].damage = nil
	83	+	result := composeFrame(hidden, committed.cache, committed.cache)
	84	+
	85	+	require.Equal(t, 's', result.frame.Row(1)[2].Rune)
	86	+	require.Equal(t, []renderer.Damage{renderer.FullRedraw()}, result.damage)
	87	+	}
	88	+
	89	+	func cacheState(title string, generation uint64) capturedRenderState {
	90	+	pane := renderer.NewFrame(6, 1)
	91	+	for x, r := range title[:min(len(title), 6)] {
	92	+	pane.Set(x, 0, renderer.Cell{Rune: r, Style: renderer.DefaultStyle()})
	93	+	}
	94	+	floating := renderer.NewFrame(2, 1)
	95	+	floating.Set(0, 0, renderer.Cell{Rune: 'F', Style: renderer.DefaultStyle()})
	96	+	return capturedRenderState{
	97	+	reset: false,
	98	+	layout: capturedTabLayout{area: domain.Rect{Width: 6, Height: 3}, valid: true, fingerprint: "layout"},
	99	+	panes: []capturedPaneRenderState{{id: "pane", frame: pane, placement: layout.Placement{ID: "pane", Content: domain.Rect{Width: 6, Height: 1}, TitleBar: domain.Rect{Width: 6, Height: 1}}, title: title, titleGeneration: generation, focused: true, damage: []renderer.Damage{renderer.FullRedraw()}}},
	100	+	floating: capturedFloatingRenderState{visible: true, pane: capturedPaneRenderState{frame: floating, damage: []renderer.Damage{renderer.FullRedraw()}}, geometry: floatingGeometry{Bounds: domain.Rect{X: 2, Y: 1, Width: 2, Height: 1}, Inner: domain.Rect{X: 2, Y: 1, Width: 2, Height: 1}}, generation: generation, titleGeneration: generation},
	101	+	bars: barState{topRight: title, bottomRight: title},
	102	+	}
	103	+	}
	104	+
	105	+	func cloneComposeCache(in composeCacheInput) composeCacheInput {
	106	+	out := in
	107	+	out.frame = in.frame.Clone()
	108	+	out.titleGenerations = make(map[layout.PaneID]uint64, len(in.titleGenerations))
	109	+	for id, generation := range in.titleGenerations {
	110	+	out.titleGenerations[id] = generation
	111	+	}
	112	+	out.damage = append([]renderer.Damage(nil), in.damage...)
	113	+	out.bars.top = append([]renderer.Cell(nil), in.bars.top...)
	114	+	out.bars.bottom = append([]renderer.Cell(nil), in.bars.bottom...)
	115	+	return out
	116	+	}

A internal/usecase/daemon/render_pipeline_test.go +281 -0

@@ -0,0 +1,281 @@
	1	+	package daemon
	2	+
	3	+	import (
	4	+	"io"
	5	+	"os"
	6	+	"path/filepath"
	7	+	"strings"
	8	+	"testing"
	9	+
	10	+	"github.com/bnema/vev/internal/domain"
	11	+	"github.com/bnema/vev/internal/ports"
	12	+	"github.com/bnema/vev/internal/testutil/replaytest"
	13	+	"github.com/bnema/vev/internal/usecase/layout"
	14	+	"github.com/bnema/vev/internal/usecase/ui"
	15	+	"github.com/bnema/vev/pkg/renderer"
	16	+	"github.com/bnema/vev/pkg/vt"
	17	+	"github.com/stretchr/testify/require"
	18	+	)
	19	+
	20	+	// scriptedReplayTransport is intentionally a ports.Transport only: replay
	21	+	// assertions must not depend on any adapter's framing implementation.
	22	+	type scriptedReplayTransport struct {
	23	+	t *testing.T
	24	+	frames []ports.Frame
	25	+	next int
	26	+	}
	27	+
	28	+	func (s *scriptedReplayTransport) Send(got ports.Frame) error {
	29	+	s.t.Helper()
	30	+	if s.next >= len(s.frames) {
	31	+	s.t.Errorf("unexpected frame %#v", got)
	32	+	return nil
	33	+	}
	34	+	want := s.frames[s.next]
	35	+	s.next++
	36	+	require.Equal(s.t, want.Type, got.Type)
	37	+	require.Equal(s.t, want.Payload, got.Payload)
	38	+	return nil
	39	+	}
	40	+	func (*scriptedReplayTransport) Recv() (ports.Frame, error) { return ports.Frame{}, io.EOF }
	41	+	func (*scriptedReplayTransport) Close() error { return nil }
	42	+
	43	+	func TestTransportReplayFinalShadowAndTerminalBytes(t *testing.T) {
	44	+	frames := replaytest.Transcript()
	45	+	transport := &scriptedReplayTransport{t: t, frames: frames}
	46	+	terminal := vt.NewScreen(8, 3)
	47	+	for _, frame := range frames {
	48	+	require.NoError(t, transport.Send(frame))
	49	+	output, err := ports.UnmarshalOutput(frame.Payload)
	50	+	require.NoError(t, err)
	51	+	require.Equal(t, frame.Payload, ports.MarshalOutput(output), "output payload must remain byte exact")
	52	+	terminal.Write(output.Data)
	53	+	}
	54	+	require.Equal(t, len(frames), transport.next)
	55	+	require.Equal(t, []string{"one ", "TWO ", " "}, frameRows(terminal.Frame), "terminal replay is the final renderer shadow")
	56	+	}
	57	+
	58	+	func TestCapturePaneRenderStateOwnsVisibleFrameWithoutConsumingDamage(t *testing.T) {
	59	+	p := newPane("p", nil, domain.Size{Cols: 8, Rows: 2})
	60	+	p.screen.ClearDamage()
	61	+	p.screen.Write([]byte("old"))
	62	+
	63	+	p.mu.Lock()
	64	+	captured := capturePaneRenderStateLocked(p, domain.Rect{Width: 3, Height: 1})
	65	+	p.mu.Unlock()
	66	+
	67	+	require.Equal(t, 3, captured.frame.Width)
	68	+	require.Equal(t, 1, captured.frame.Height)
	69	+	require.NotEmpty(t, p.screen.Damage(), "capture is only a transactional snapshot")
	70	+	p.mu.Lock()
	71	+	p.screen.Write([]byte("new"))
	72	+	p.mu.Unlock()
	73	+	require.Equal(t, 'o', captured.frame.At(0, 0).Rune, "capture must not alias the mutable VT frame")
	74	+	}
	75	+
	76	+	func TestCapturePaneRenderStateIsAlwaysNonDestructive(t *testing.T) {
	77	+	p := newPane("p", nil, domain.Size{Cols: 8, Rows: 2})
	78	+	p.screen.ClearDamage()
	79	+	p.screen.Write([]byte("preview"))
	80	+
	81	+	p.mu.Lock()
	82	+	_ = capturePaneRenderStateLocked(p, domain.Rect{Width: 8, Height: 2})
	83	+	p.mu.Unlock()
	84	+
	85	+	require.NotEmpty(t, p.screen.Damage())
	86	+	}
	87	+
	88	+	func TestCapturePaneRenderStateMalformedDamageFallsBackToFullRedraw(t *testing.T) {
	89	+	p := newPane("p", nil, domain.Size{Cols: 8, Rows: 2})
	90	+	p.screen.ClearDamage()
	91	+	p.screen.Write([]byte("x"))
	92	+	p.screen.Damage()[0] = renderer.Damage{Kind: renderer.DamageText, X: -1, Y: 0, Width: 4, Height: 1}
	93	+
	94	+	p.mu.Lock()
	95	+	captured := capturePaneRenderStateLocked(p, domain.Rect{Width: 8, Height: 2})
	96	+	p.mu.Unlock()
	97	+
	98	+	require.Equal(t, []renderer.Damage{renderer.FullRedraw()}, captured.damage)
	99	+	}
	100	+
	101	+	func TestCapturePaneRenderStateCollapsedRetainsDamage(t *testing.T) {
	102	+	p := newPane("p", nil, domain.Size{Cols: 8, Rows: 2})
	103	+	p.screen.ClearDamage()
	104	+	p.screen.Write([]byte("hidden"))
	105	+
	106	+	p.mu.Lock()
	107	+	captured := capturePaneRenderStateLocked(p, domain.Rect{})
	108	+	p.mu.Unlock()
	109	+
	110	+	require.Zero(t, captured.frame.Width)
	111	+	require.Zero(t, captured.frame.Height)
	112	+	require.NotEmpty(t, p.screen.Damage(), "capture keeps hidden-pane damage pending until emission commits")
	113	+	}
	114	+
	115	+	func TestPaintACKBlockedDoesNotDestructivelyCapture(t *testing.T) {
	116	+	pty, release := newBlockingPTY(t)
	117	+	t.Cleanup(release)
	118	+	d, sess, _, sends := newManualSessionWithPTYs(t, pty)
	119	+	ac := sess.client
	120	+	ac.sendMu.Lock()
	121	+	ac.output.next = ac.output.maxOutstanding
	122	+	ac.output.acked = 0
	123	+	ac.sendMu.Unlock()
	124	+	p := sess.tabs[0].focusedPane()
	125	+	p.mu.Lock()
	126	+	p.screen.ClearDamage()
	127	+	p.screen.Write([]byte("blocked"))
	128	+	p.mu.Unlock()
	129	+
	130	+	d.paint(sess, ac, false, nil)
	131	+
	132	+	p.mu.Lock()
	133	+	require.NotEmpty(t, p.screen.Damage())
	134	+	p.mu.Unlock()
	135	+	select {
	136	+	case frame := <-sends:
	137	+	t.Fatalf("ACK-blocked paint sent %#v", frame)
	138	+	default:
	139	+	}
	140	+	}
	141	+
	142	+	func TestComposeFrameUsesCapturedValuesAfterOwnersMutate(t *testing.T) {
	143	+	state := &capturedRenderState{
	144	+	layout: capturedTabLayout{area: domain.Rect{Width: 4, Height: 1}, valid: true},
	145	+	panes: []capturedPaneRenderState{{
	146	+	id: "p", frame: func() renderer.Frame {
	147	+	f := renderer.NewFrame(4, 1)
	148	+	for i, r := range "old " {
	149	+	f.Set(i, 0, renderer.Cell{Rune: r, Style: renderer.DefaultStyle()})
	150	+	}
	151	+	return f
	152	+	}(), placement: layout.Placement{ID: "p", Content: domain.Rect{Width: 4, Height: 1}}, focused: true,
	153	+	damage: []renderer.Damage{renderer.FullRedraw()},
	154	+	}},
	155	+	bars: barState{},
	156	+	cursor: capturedCursorInputs{visible: true, renderable: true, content: domain.Rect{Width: 4, Height: 1}},
	157	+	}
	158	+	cache := composeCacheInput{}
	159	+	out := composeFrame(*state, cache)
	160	+	state.panes[0].frame.Set(0, 0, renderer.Cell{Rune: 'X', Style: renderer.DefaultStyle()})
	161	+	require.Equal(t, "old ", rowText(out.frame.Row(1)))
	162	+	require.False(t, out.cursor.hidden)
	163	+	}
	164	+
	165	+	func TestProductionRenderUsesOnlyTypedComposeEntryPoint(t *testing.T) {
	166	+	for _, name := range []string{"render.go", "floating_render.go", "picker.go"} {
	167	+	data, err := os.ReadFile(filepath.Join(".", name))
	168	+	require.NoError(t, err)
	169	+	source := string(data)
	170	+	for _, forbidden := range []string{
	171	+	"func composeClientFrame", "func composeTabFrame", "func composeFloatingFrame",
	172	+	} {
	173	+	require.NotContains(t, source, forbidden, "%s retains duplicate production composition path", name)
	174	+	}
	175	+	}
	176	+	data, err := os.ReadFile("render_pipeline.go")
	177	+	require.NoError(t, err)
	178	+	body := string(data)
	179	+	start := strings.Index(body, "func composeFrame(")
	180	+	require.NotEqual(t, -1, start)
	181	+	end := strings.Index(body[start:], "\nfunc ")
	182	+	if end > 0 {
	183	+	body = body[start : start+end]
	184	+	} else {
	185	+	body = body[start:]
	186	+	}
	187	+	for _, forbidden := range []string{".Lock()", ".Unlock()", ".Send(", ".prepare("} {
	188	+	require.NotContains(t, body, forbidden, "composeFrame must remain ownership- and transport-free")
	189	+	}
	190	+	}
	191	+
	192	+	func TestComposeEmitExactReplayTiledFloatingBarsOverlayAndCursor(t *testing.T) {
	193	+	paneFrame := renderer.NewFrame(12, 5)
	194	+	for y, text := range []string{"AAAAAAAAAAAA", "BBBBBBBBBBBB", "CCCCCCCCCCCC", "DDDDDDDDDDDD", "EEEEEEEEEEEE"} {
	195	+	for x, r := range text {
	196	+	paneFrame.Set(x, y, renderer.Cell{Rune: r, Style: renderer.DefaultStyle()})
	197	+	}
	198	+	}
	199	+	floatingFrame := renderer.NewFrame(4, 1)
	200	+	for x, r := range "FLOAT"[:4] {
	201	+	floatingFrame.Set(x, 0, renderer.Cell{Rune: r, Style: renderer.DefaultStyle()})
	202	+	}
	203	+	modalInner := renderer.NewFrame(6, 1)
	204	+	for x, r := range "PROMPT" {
	205	+	modalInner.Set(x, 0, renderer.Cell{Rune: r, Style: renderer.DefaultStyle()})
	206	+	}
	207	+	state := capturedRenderState{
	208	+	reset: true,
	209	+	layout: capturedTabLayout{area: domain.Rect{Width: 12, Height: 5}, valid: true, focus: "p", placements: []layout.Placement{{ID: "p", Content: domain.Rect{Width: 12, Height: 5}}}},
	210	+	panes: []capturedPaneRenderState{{id: "p", frame: paneFrame, placement: layout.Placement{ID: "p", Content: domain.Rect{Width: 12, Height: 5}}, focused: true, damage: []renderer.Damage{renderer.FullRedraw()}}},
	211	+	floating: capturedFloatingRenderState{visible: true, pane: capturedPaneRenderState{id: "f", frame: floatingFrame}, geometry: floatingGeometry{Bounds: domain.Rect{X: 3, Y: 1, Width: 6, Height: 3}, Inner: domain.Rect{X: 4, Y: 2, Width: 4, Height: 1}}, title: "float", generation: 1},
	212	+	bars: barState{status: statusSnapshot{session: "sess", tabs: []statusTab{{name: "tab", active: true}}}, topRight: "R", bottomRight: "B"},
	213	+	overlays: capturedOverlayRenderState{promptActive: true, prompt: capturedModal{modal: ui.Modal{FixedWidth: 8, FixedHeight: 3, Title: "Prompt"}, inner: modalInner}},
	214	+	cursor: capturedCursorInputs{row: 1, col: 2, visible: true, renderable: true, content: domain.Rect{X: 4, Y: 2, Width: 4, Height: 1}},
	215	+	}
	216	+	composed := composeFrame(state, composeCacheInput{})
	217	+	require.Equal(t, []string{" tab R", "AAAAAAAAAAAA", "BB┌Prompt┐BB", "CC│PROMPT│CC", "DD└──────┘DD", "EEEEEEEEEEEE", " sess B"}, frameRows(composed.frame))
	218	+	require.True(t, composed.cursor.hidden, "overlay owns cursor visibility")
	219	+
	220	+	stream := newOutputStateStream()
	221	+	prepared, err := stream.prepare(composed.frame, composed.damage, composed.reset)
	222	+	require.NoError(t, err)
	223	+	terminalBytes := append(append([]byte(nil), prepared.data...), (&attachedClient{}).encodeCursorTail(composed.cursor, true)...)
	224	+	require.Equal(t, "\x1b[1;1H\x1b[0;7m tab \x1b[0m R\x1b[2;1HAAAAAAAAAAAA\x1b[3;1HBB┌Prompt┐BB\x1b[4;1HCC│PROMPT│CC\x1b[5;1HDD└──────┘DD\x1b[6;1HEEEEEEEEEEEE\x1b[7;1H\x1b[0;7m sess \x1b[0m B\x1b[0m\x1b[?25l", string(terminalBytes))
	225	+	client := vt.NewScreen(composed.frame.Width, composed.frame.Height)
	226	+	client.Write(terminalBytes)
	227	+	require.Equal(t, frameRows(composed.frame), frameRows(client.Frame))
	228	+	again, err := stream.renderer.Draw(composed.frame, nil)
	229	+	require.NoError(t, err)
	230	+	require.Empty(t, again, "renderer shadow must exactly equal the composed frame")
	231	+	}
	232	+
	233	+	func frameRows(frame renderer.Frame) []string {
	234	+	rows := make([]string, frame.Height)
	235	+	for y := range rows {
	236	+	rows[y] = rowText(frame.Row(y))
	237	+	}
	238	+	return rows
	239	+	}
	240	+
	241	+	func TestComposeEmitExactReplaySafeAndUnsafeScroll(t *testing.T) {
	242	+	for _, tt := range []struct {
	243	+	name string
	244	+	placement layout.Placement
	245	+	initial []string
	246	+	scrolled []string
	247	+	damage []renderer.Damage
	248	+	wantBytes string
	249	+	}{
	250	+	{name: "safe full width", placement: layout.Placement{ID: "p", Content: domain.Rect{Width: 4, Height: 3}}, initial: []string{"AAAA", "BBBB", "CCCC"}, scrolled: []string{"BBBB", "CCCC", "N "}, damage: []renderer.Damage{{Kind: renderer.DamageScrollUp, X: 0, Y: 0, Width: 4, Height: 3, Count: 1}, {Kind: renderer.DamageText, X: 0, Y: 2, Width: 4, Height: 1}}, wantBytes: "\x1b[0m\x1b[2;4r\x1b[4;1H\n\x1b[r\x1b[4;1HN\x1b[K\x1b[0m"},
	251	+	{name: "unsafe partial width", placement: layout.Placement{ID: "p", Content: domain.Rect{X: 2, Width: 2, Height: 3}}, initial: []string{"AA", "BB", "CC"}, scrolled: []string{"BB", "CC", "N "}, damage: []renderer.Damage{{Kind: renderer.DamageText, X: 2, Y: 0, Width: 2, Height: 3}}, wantBytes: "\x1b[2;3HBB\x1b[3;3HCC\x1b[4;3HN \x1b[0m"},
	252	+	} {
	253	+	t.Run(tt.name, func(t *testing.T) {
	254	+	makeState := func(rows []string, damage []renderer.Damage, reset bool) capturedRenderState {
	255	+	f := renderer.NewFrame(tt.placement.Content.Width, tt.placement.Content.Height)
	256	+	for y, text := range rows {
	257	+	for x, r := range text {
	258	+	f.Set(x, y, renderer.Cell{Rune: r, Style: renderer.DefaultStyle()})
	259	+	}
	260	+	}
	261	+	return capturedRenderState{reset: reset, layout: capturedTabLayout{area: domain.Rect{Width: 4, Height: 3}, valid: true, focus: "p", placements: []layout.Placement{tt.placement}}, panes: []capturedPaneRenderState{{id: "p", frame: f, placement: tt.placement, focused: true, damage: damage}}, cursor: capturedCursorInputs{visible: false}}
	262	+	}
	263	+	stream := newOutputStateStream()
	264	+	first := composeFrame(makeState(tt.initial, []renderer.Damage{renderer.FullRedraw()}, true), composeCacheInput{})
	265	+	firstBytes, err := stream.render(first.frame, first.damage, first.reset)
	266	+	require.NoError(t, err)
	267	+	client := vt.NewScreen(4, 5)
	268	+	client.Write(firstBytes)
	269	+
	270	+	second := composeFrame(makeState(tt.scrolled, tt.damage, false), first.cache)
	271	+	secondBytes, err := stream.render(second.frame, second.damage, second.reset)
	272	+	require.NoError(t, err)
	273	+	require.Equal(t, tt.wantBytes, string(secondBytes))
	274	+	client.Write(secondBytes)
	275	+	require.Equal(t, frameRows(second.frame), frameRows(client.Frame))
	276	+	again, err := stream.renderer.Draw(second.frame, nil)
	277	+	require.NoError(t, err)
	278	+	require.Empty(t, again, "renderer shadow must equal replayed terminal state")
	279	+	})
	280	+	}
	281	+	}

A internal/usecase/daemon/render_resize_deadline.go +147 -0

@@ -0,0 +1,147 @@
	1	+	package daemon
	2	+
	3	+	import "github.com/bnema/vev/internal/domain"
	4	+
	5	+	// Resize deadlines and retries retain coordinator-owned attachment, epoch, and
	6	+	// token validation at every callback effect boundary.
	7	+	func (c renderCoordinator) recordResizeRequest(size domain.Size, source attachedClient) uint64 {
	8	+	return c.recordResizeRequestForLease(size, source, c.attachmentLease(source))
	9	+	}
	10	+
	11	+	func (c renderCoordinator) recordResizeRequestForLease(size domain.Size, source attachedClient, lease *attachmentLease) uint64 {
	12	+	c.mu.Lock()
	13	+	if !c.leaseCurrentLocked(lease, false) \|\| lease.attachment != source {
	14	+	c.mu.Unlock()
	15	+	return 0
	16	+	}
	17	+	_, retry := c.retryLane.replaceLocked()
	18	+	c.resize.epoch++
	19	+	c.resize.size, c.resize.source, c.resize.lease = size, source, lease
	20	+	epoch := c.resize.epoch
	21	+	c.mu.Unlock()
	22	+	stopDetachedTimer(retry)
	23	+	return epoch
	24	+	}
	25	+
	26	+	func (c renderCoordinator) scheduleResizeForLease(size domain.Size, source attachedClient, lease *attachmentLease, run func(uint64)) uint64 {
	27	+	epoch := c.recordResizeRequestForLease(size, source, lease)
	28	+	if epoch == 0 {
	29	+	return 0
	30	+	}
	31	+	c.mu.Lock()
	32	+	gen, old := c.resizeLane.replaceLocked()
	33	+	clock := c.opts.clock
	34	+	c.mu.Unlock()
	35	+	stopDetachedTimer(old)
	36	+	if clock == nil {
	37	+	if c.resizeCurrentForLease(epoch, source, lease, false) {
	38	+	run(epoch)
	39	+	}
	40	+	return epoch
	41	+	}
	42	+	timer := clock.NewTimer(minOutputRenderDeadline)
	43	+	timerC := timer.C()
	44	+	if timerC == nil {
	45	+	timer.Stop()
	46	+	if c.resizeCurrentForLease(epoch, source, lease, false) {
	47	+	run(epoch)
	48	+	}
	49	+	return epoch
	50	+	}
	51	+	c.mu.Lock()
	52	+	valid := !c.torndown && c.resizeLane.generation == gen && c.leaseCurrentLocked(lease, false) && c.resize.lease == lease && c.resize.epoch == epoch && c.resize.source == source
	53	+	var token *timerToken
	54	+	if valid {
	55	+	token = c.resizeLane.publishLocked(gen, timer)
	56	+	valid = token != nil
	57	+	if valid {
	58	+	c.supervisor.startLocked(token, timerC, func() {
	59	+	c.mu.Lock()
	60	+	valid := !c.torndown && c.resizeLane.token == token && c.resizeLane.generation == token.generation && c.leaseCurrentLocked(lease, false) && c.resize.lease == lease && c.resize.epoch == epoch && c.resize.source == source
	61	+	if valid {
	62	+	c.resizeLane.clearLocked(token)
	63	+	}
	64	+	c.mu.Unlock()
	65	+	if valid {
	66	+	run(epoch)
	67	+	}
	68	+	})
	69	+	}
	70	+	}
	71	+	c.mu.Unlock()
	72	+	if !valid {
	73	+	timer.Stop()
	74	+	}
	75	+	return epoch
	76	+	}
	77	+
	78	+	func (c renderCoordinator) resizeCurrentForLease(epoch uint64, source attachedClient, lease *attachmentLease, commit bool) bool {
	79	+	c.mu.Lock()
	80	+	defer c.mu.Unlock()
	81	+	if !c.leaseCurrentLocked(lease, false) \|\| lease.attachment != source \|\| c.resize.lease != lease \|\| c.resize.source != source \|\| c.resize.epoch != epoch {
	82	+	return false
	83	+	}
	84	+	if commit {
	85	+	if c.resize.committed >= epoch {
	86	+	return false
	87	+	}
	88	+	c.resize.committed = epoch
	89	+	}
	90	+	return true
	91	+	}
	92	+
	93	+	func (c renderCoordinator) scheduleResizeRetryForLease(epoch uint64, source attachedClient, lease *attachmentLease, run func()) {
	94	+	c.mu.Lock()
	95	+	if !c.leaseCurrentLocked(lease, false) \|\| lease.attachment != source \|\| c.resize.lease != lease \|\| c.resize.source != source \|\| c.resize.epoch != epoch \|\| c.resize.committed != epoch {
	96	+	c.mu.Unlock()
	97	+	return
	98	+	}
	99	+	gen, old := c.retryLane.replaceLocked()
	100	+	clock := c.opts.clock
	101	+	c.mu.Unlock()
	102	+	stopDetachedTimer(old)
	103	+	if clock == nil {
	104	+	return
	105	+	}
	106	+	timer := clock.NewTimer(minOutputRenderDeadline)
	107	+	timerC := timer.C()
	108	+	if timerC == nil {
	109	+	timer.Stop()
	110	+	return
	111	+	}
	112	+	c.mu.Lock()
	113	+	valid := !c.torndown && c.retryLane.generation == gen && c.leaseCurrentLocked(lease, false) && c.resize.lease == lease && c.resize.epoch == epoch && c.resize.committed == epoch && c.resize.source == source
	114	+	if valid {
	115	+	token := c.retryLane.publishLocked(gen, timer)
	116	+	valid = token != nil
	117	+	if valid {
	118	+	c.supervisor.startLocked(token, timerC, func() {
	119	+	c.mu.Lock()
	120	+	valid := !c.torndown && c.retryLane.token == token && c.retryLane.generation == token.generation && c.leaseCurrentLocked(lease, false) && c.resize.lease == lease && c.resize.epoch == epoch && c.resize.committed == epoch && c.resize.source == source
	121	+	if valid {
	122	+	c.retryLane.clearLocked(token)
	123	+	}
	124	+	c.mu.Unlock()
	125	+	if valid {
	126	+	run()
	127	+	}
	128	+	})
	129	+	}
	130	+	}
	131	+	c.mu.Unlock()
	132	+	if !valid {
	133	+	timer.Stop()
	134	+	}
	135	+	}
	136	+
	137	+	func (c renderCoordinator) retryCurrentForLease(epoch uint64, source attachedClient, lease *attachmentLease) bool {
	138	+	c.mu.Lock()
	139	+	defer c.mu.Unlock()
	140	+	return c.leaseCurrentLocked(lease, false) && lease.attachment == source && c.resize.lease == lease && c.resize.source == source && c.resize.epoch == epoch && c.resize.committed == epoch
	141	+	}
	142	+
	143	+	func (c *renderCoordinator) resizeSnapshot() resizeRequestMetadata {
	144	+	c.mu.Lock()
	145	+	defer c.mu.Unlock()
	146	+	return c.resize
	147	+	}

M internal/usecase/daemon/render_test.go +508 -419

 	require.True(t, paletteVisible, "palette must remain composed above copy mode")
 }
+func TestFirstPaintRetainedFloatingPaneEmitsOneReset(t *testing.T) {
+	for _, tc := range []struct {
+		name       string
+		clientSize domain.Size
+	}{
+		{name: "matching outer size", clientSize: domain.Size{Cols: 80, Rows: 25}},
+		{name: "different outer size", clientSize: domain.Size{Cols: 100, Rows: 40}},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			pty, releasePTY := newBlockingPTY(t)
+			defer releasePTY()
+			d, sess, ac, sends := newManualSessionWithPTYs(t, pty)
+			ac.sendMu.Lock()
+			ac.size = tc.clientSize
+			ac.sendMu.Unlock()
+			// A retained visible popup still needs activation warmup. When the
+			// outer terminal also changes, its completed transaction must cover
+			// that popup rather than emitting a second reset-producing frame.
+			floating := newPane(layout.PaneID("floating"), nil, domain.Size{Cols: 20, Rows: 8})
+			installTestFloating(sess.activeTab(), floating, true)
+			d.firstPaint(sess, ac, tc.clientSize)
+			frame := awaitFrame(t, sends, ports.MsgOutput)
+			output, err := ports.UnmarshalOutput(frame.Payload)
+			require.NoError(t, err)
+			require.Zero(t, output.BaseStateNum, "first paint must be a mandatory reset")
+			select {
+			case extra := <-sends:
+				t.Fatalf("first paint emitted duplicate frame after floating activation resize: %#v", extra)
+			default:
+			}
+		})
+	}
+}
 func TestPaletteBackdropKeepsSimultaneousPickerCrisp(t *testing.T) {
 	p, release := newBlockingPTY(t)
 	defer release()
 	client := vt.NewScreen(80, 25)
 	pane := sess.tabs[0].focusedPane()
 	pane.screen.Write([]byte("X"))
-	d.paint(sess, ac, true)
+	d.paint(sess, ac, true, nil)
 	mustApplyOutput(t, client, awaitFrame(t, sends, ports.MsgOutput))
 	undimmed := client.Frame.At(0, 1)
 	topBar := client.Frame.At(0, 0)
 // stubClock returns timers whose channel never fires, so a scheduler under it
 // blocks in its debounce loop until the session context is cancelled. Used by
+// channelPTY drives ptyReader through generated PTY mocks. A processed token is
+// published only when ptyReader asks for its next read, which proves the
+// preceding chunk completed its parser and coordinator path without sleeping.
+type channelPTYStep struct {
+	data []byte
+	err  error
+}
+func newChannelPTY(t *testing.T) (*portsmocks.MockPTY, chan<- channelPTYStep, <-chan struct{}) {
+	t.Helper()
+	pty := portsmocks.NewMockPTY(t)
+	steps := make(chan channelPTYStep)
+	processed := make(chan struct{}, 8)
+	read := false
+	pty.EXPECT().Read(mock.Anything).RunAndReturn(func(buf []byte) (int, error) {
+		if read {
+			processed <- struct{}{}
+		}
+		read = true
+		step := <-steps
+		return copy(buf, step.data), step.err
+	}).Maybe()
+	pty.EXPECT().Close().Return(nil).Maybe()
+	return pty, steps, processed
+}
+func awaitPTYReadProcessed(t *testing.T, processed <-chan struct{}) {
+	t.Helper()
+	<-processed
+}
+func awaitOutputFrameWithoutSleep(t *testing.T, sends <-chan ports.Frame) ports.Frame {
+	t.Helper()
+	frame := <-sends
+	require.Equal(t, ports.MsgOutput, frame.Type)
+	return frame
+}
+func drainCoordinatorTimers(clock *coordinatorMockClock) []*coordinatorMockTimer {
+	var timers []*coordinatorMockTimer
+	for {
+		select {
+		case timer := <-clock.timers:
+			timers = append(timers, timer)
+		default:
+			return timers
+		}
+	}
+}
+func fireCoordinatorTimer(t *testing.T, rc *renderCoordinator, timers []*coordinatorMockTimer, duration time.Duration) {
+	t.Helper()
+	for _, timer := range timers {
+		if timer.duration != duration {
+			continue
+		}
+		var done <-chan struct{}
+		rc.mu.Lock()
+		if rc.normalLane.token.timer == timer.mock {
+			done = rc.normalLane.token.done
+		}
+		rc.mu.Unlock()
+		timer.ch <- time.Time{}
+		if done != nil {
+			<-done
+		}
+		return
+	}
+	t.Fatalf("coordinator did not arm %s timer", duration)
+}
+func TestPTYReaderSyncVisibilityTransitions(t *testing.T) {
+	t.Run("inactive synchronized pane activates only after complete urgent frame", func(t *testing.T) {
+		inactivePTY, inactiveSteps, inactiveProcessed := newChannelPTY(t)
+		activePTY, _, _ := newChannelPTY(t)
+		d, sess, ac, sends := newManualSessionWithPTYs(t, activePTY, inactivePTY)
+		clock := newCoordinatorMockClock(t, 8)
+		d.clock = clock.clock
+		rc := d.attachCoordinator(sess, nil, ac, true)
+		d.sessWg.Add(1)
+		go d.ptyReader(sess, sess.tabs[1], sess.tabs[1].focusedPane())
+		inactiveSteps <- channelPTYStep{data: []byte("\x1b[?2026hpartial")}
+		awaitPTYReadProcessed(t, inactiveProcessed)
+		sess.mu.Lock()
+		sess.active = 1
+		sess.mu.Unlock()
+		d.invalidateRender(sess, ac, true, "sync activation")
+		fireCoordinatorTimer(t, rc, drainCoordinatorTimers(clock), urgentRenderDeadline)
+		requireNoCoordinatorOutputFrame(t, sends)
+		inactiveSteps <- channelPTYStep{data: []byte(" complete\x1b[?2026l")}
+		awaitPTYReadProcessed(t, inactiveProcessed)
+		fireCoordinatorTimer(t, rc, drainCoordinatorTimers(clock), urgentRenderDeadline)
+		frame := awaitOutputFrameWithoutSleep(t, sends)
+		output, err := ports.UnmarshalOutput(frame.Payload)
+		require.NoError(t, err)
+		require.Contains(t, string(output.Data), "partial complete")
+		requireNoCoordinatorOutputFrame(t, sends)
+		inactiveSteps <- channelPTYStep{err: io.EOF}
+		d.sessWg.Wait()
+	})
+	t.Run("hidden synchronized pane cannot stall newly active output or flush late", func(t *testing.T) {
+		oldPTY, oldSteps, oldProcessed := newChannelPTY(t)
+		newPTY, newSteps, newProcessed := newChannelPTY(t)
+		parkedPTY, _, _ := newChannelPTY(t)
+		// Keep one unopened tab so reader EOF cleanup cannot tear down the
+		// session and introduce unrelated detach-notification timers.
+		d, sess, ac, sends := newManualSessionWithPTYs(t, oldPTY, newPTY, parkedPTY)
+		clock := newCoordinatorMockClock(t, 8)
+		d.clock = clock.clock
+		rc := d.attachCoordinator(sess, nil, ac, true)
+		d.sessWg.Add(2)
+		go d.ptyReader(sess, sess.tabs[0], sess.tabs[0].focusedPane())
+		go d.ptyReader(sess, sess.tabs[1], sess.tabs[1].focusedPane())
+		oldSteps <- channelPTYStep{data: []byte("\x1b[?2026hold partial")}
+		awaitPTYReadProcessed(t, oldProcessed)
+		sess.mu.Lock()
+		sess.active = 1
+		sess.mu.Unlock()
+		newSteps <- channelPTYStep{data: []byte("newly active")}
+		awaitPTYReadProcessed(t, newProcessed)
+		fireCoordinatorTimer(t, rc, drainCoordinatorTimers(clock), minOutputRenderDeadline)
+		frame := awaitOutputFrameWithoutSleep(t, sends)
+		output, err := ports.UnmarshalOutput(frame.Payload)
+		require.NoError(t, err)
+		require.Contains(t, string(output.Data), "newly active")
+		require.NotContains(t, string(output.Data), "old partial")
+		oldSteps <- channelPTYStep{data: []byte("\x1b[?2026l")}
+		awaitPTYReadProcessed(t, oldProcessed)
+		requireNoCoordinatorOutputFrame(t, sends)
+		oldSteps <- channelPTYStep{err: io.EOF}
+		newSteps <- channelPTYStep{err: io.EOF}
+		d.sessWg.Wait()
+	})
+}
+func TestPTYReaderRestoresPrimaryScreenImmediatelyAfterDEC1049Exit(t *testing.T) {
+	for _, tc := range []struct {
+		name          string
+		window        uint64
+		exitBeforeACK bool
+		exitSuffix    string
+		overlay       bool
+	}{
+		{name: "datagram output window (1): exit waits for ACK", window: 1},
+		{name: "datagram output window (1): exit and prompt arrive before ACK", window: 1, exitBeforeACK: true, exitSuffix: "PROMPT> "},
+		{name: "local default output window (8): exit waits for ACK", window: maxUnackedOutputStates},
+		{name: "local default output window (8): exit and prompt arrive before ACK", window: maxUnackedOutputStates, exitBeforeACK: true, exitSuffix: "PROMPT> "},
+		{name: "local default output window (8): palette overlay with bars", window: maxUnackedOutputStates, overlay: true},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			pty, steps, processed := newChannelPTY(t)
+			d, sess, ac, sends := newManualSessionWithPTYs(t, pty)
+			// Datagram clients use a one-frame window; local clients retain the
+			// default eight-frame window. Both ACK orderings stay deterministic.
+			ac.output.maxOutstanding = tc.window
+			clock := newCoordinatorMockClock(t, 8)
+			d.clock = clock.clock
+			rc := d.attachCoordinator(sess, nil, ac, true)
+			client := vt.NewScreen(80, 25)
+			pane := sess.tabs[0].focusedPane()
+			d.sessWg.Add(1)
+			go d.ptyReader(sess, sess.tabs[0], pane)
+			t.Cleanup(func() {
+				steps <- channelPTYStep{err: io.EOF}
+				d.sessWg.Wait()
+				d.waitNotifies()
+			})
+			replay := func(data []byte, deadline time.Duration) ports.Output {
+				t.Helper()
+				steps <- channelPTYStep{data: data}
+				awaitPTYReadProcessed(t, processed)
+				fireCoordinatorTimer(t, rc, drainCoordinatorTimers(clock), deadline)
+				return mustApplyOutput(t, client, awaitOutputFrameWithoutSleep(t, sends))
+			}
+			acknowledge := func(out ports.Output) {
+				ac.ackOutputState(out.NewStateNum)
+				rc.notifyAck()
+			}
+			if tc.overlay {
+				d.enterPalette(sess, ac)
+				fireCoordinatorTimer(t, rc, drainCoordinatorTimers(clock), urgentRenderDeadline)
+				acknowledge(mustApplyOutput(t, client, awaitOutputFrameWithoutSleep(t, sends)))
+			}
+			acknowledge(replay([]byte("PRIMARY"), minOutputRenderDeadline))
+			primary := client.Frame.Clone()
+			// The reported PTY stream is deliberately split into enter, body, and
+			// exit reads. MATRIX distinguishes alternate cells in the replayed
+			// terminal model from all primary, chrome, and overlay cells.
+			acknowledge(replay([]byte("\x1b[?1049h"), minOutputRenderDeadline))
+			alternate := replay([]byte(strings.Repeat("MATRIX", 300)), minOutputRenderDeadline)
+			pane.mu.Lock()
+			require.Contains(t, strings.Join(frameRows(pane.screen.Frame), "\n"), "MATRIX", "fixture must make alternate cells visible to the production renderer")
+			pane.mu.Unlock()
+			exitData := []byte("\x1b[?1049l" + tc.exitSuffix)
+			var exit ports.Output
+			if tc.exitBeforeACK {
+				steps <- channelPTYStep{data: exitData}
+				awaitPTYReadProcessed(t, processed)
+				fireCoordinatorTimer(t, rc, drainCoordinatorTimers(clock), minOutputRenderDeadline)
+				if tc.window == 1 {
+					requireNoCoordinatorOutputFrame(t, sends)
+					acknowledge(alternate)
+					exit = mustApplyOutput(t, client, awaitOutputFrameWithoutSleep(t, sends))
+				} else {
+					exit = mustApplyOutput(t, client, awaitOutputFrameWithoutSleep(t, sends))
+					require.Equal(t, alternate.NewStateNum, exit.BaseStateNum,
+						"local exit must be rendered before the alternate-frame ACK")
+					acknowledge(alternate)
+				}
+			} else {
+				acknowledge(alternate)
+				steps <- channelPTYStep{data: exitData}
+				awaitPTYReadProcessed(t, processed)
+				fireCoordinatorTimer(t, rc, drainCoordinatorTimers(clock), minOutputRenderDeadline)
+				exit = mustApplyOutput(t, client, awaitOutputFrameWithoutSleep(t, sends))
+			}
+			require.NotEmpty(t, exit.Data, "the first post-exit output must actively remove alternate cells")
+			rendered := strings.Join(frameRows(client.Frame), "\n")
+			require.NotContains(t, rendered, "MATRIX", "the first post-exit frame must not retain alternate cells")
+			if tc.exitSuffix == "" {
+				require.Equal(t, frameRows(primary), frameRows(client.Frame), "the first post-exit output must restore the primary screen")
+			} else {
+				require.Contains(t, rendered, tc.exitSuffix, "bytes immediately after DEC 1049 exit must appear in the first post-exit frame")
+			}
+		})
+	}
+}
+func TestPTYReaderRepublishesSynchronizedCompletionAfterAttachmentLifecycle(t *testing.T) {
+	t.Run("detached target completes cross-session preview without later PTY output", func(t *testing.T) {
+		pty, steps, processed := newChannelPTY(t)
+		d, target, targetClient, _ := newManualSessionWithPTYs(t, pty)
+		clock := newCoordinatorMockClock(t, 8)
+		d.clock = clock.clock
+		rc := d.attachCoordinator(target, nil, targetClient, true)
+		// A viewer in a different session is previewing the headless target.
+		viewer := &attachedClient{}
+		viewer.initOverlays()
+		viewer.overlays.pickerMu.Lock()
+		viewer.overlays.pickerPreview = target.tabs[0]
+		viewer.overlays.pickerMu.Unlock()
+		d.sessions["viewer"] = &session{id: "viewer", client: viewer}
+		previews := make(chan renderWake, 2)
+		rc.subscribePreviewFor(viewer, 1, func(w renderWake) { previews <- w })
+		d.sessWg.Add(1)
+		go d.ptyReader(target, target.tabs[0], target.tabs[0].focusedPane())
+		steps <- channelPTYStep{data: []byte("\x1b[?2026hpending")}
+		awaitPTYReadProcessed(t, processed)
+		drainCoordinatorTimers(clock) // detach must clear the pending output work.
+		target.mu.Lock()
+		target.client = nil
+		target.mu.Unlock()
+		rc.noteDetach(targetClient)
+		steps <- channelPTYStep{data: []byte(" complete\x1b[?2026l")}
+		awaitPTYReadProcessed(t, processed)
+		fireCoordinatorTimer(t, rc, drainCoordinatorTimers(clock), urgentRenderDeadline)
+		wake := <-previews
+		require.True(t, wake.urgent)
+		require.Equal(t, 1, wake.coalesced, "completion starts a fresh headless preview batch after detach")
+		select {
+		case duplicate := <-previews:
+			t.Fatalf("sync completion must publish exactly one preview wake: %#v", duplicate)
+		default:
+		}
+		steps <- channelPTYStep{err: io.EOF}
+		d.sessWg.Wait()
+	})
+	t.Run("replacement receives one complete reset frame without later PTY output", func(t *testing.T) {
+		pty, steps, processed := newChannelPTY(t)
+		d, target, oldClient, _ := newManualSessionWithPTYs(t, pty)
+		clock := newCoordinatorMockClock(t, 8)
+		d.clock = clock.clock
+		rc := d.attachCoordinator(target, nil, oldClient, true)
+		wakes := make(chan renderWake, 2)
+		rc.opts.wake = func(w renderWake) { wakes <- w }
+		d.sessWg.Add(1)
+		go d.ptyReader(target, target.tabs[0], target.tabs[0].focusedPane())
+		steps <- channelPTYStep{data: []byte("\x1b[?2026hpending")}
+		awaitPTYReadProcessed(t, processed)
+		drainCoordinatorTimers(clock)
+		replacement := &attachedClient{output: newOutputStateStream()}
+		rc.noteReplace(oldClient, replacement, true)
+		target.mu.Lock()
+		target.client = replacement
+		target.mu.Unlock()
+		// The replacement's initial full paint remains gated by the batch. Its
+		// reset must survive completion and coalesce with the completion wake.
+		rc.invalidateForAttachment(replacement, renderInvalidation{class: invalidateUrgent, reset: true, producer: "replacement first paint"})
+		replacementTimers := drainCoordinatorTimers(clock)
+		steps <- channelPTYStep{data: []byte(" complete\x1b[?2026l")}
+		awaitPTYReadProcessed(t, processed)
+		requireNoWake(t, wakes)
+		fireCoordinatorTimer(t, rc, replacementTimers, urgentRenderDeadline)
+		wake := <-wakes
+		require.True(t, wake.urgent)
+		require.True(t, wake.reset, "the replacement's cleared batch must repaint a complete frame")
+		require.Equal(t, 2, wake.coalesced, "completion coalesces only with the replacement's fresh reset batch")
+		require.Same(t, replacement, wake.lease.attachment)
+		select {
+		case duplicate := <-wakes:
+			t.Fatalf("sync completion must publish exactly one replacement wake: %#v", duplicate)
+		default:
+		}
+		steps <- channelPTYStep{err: io.EOF}
+		d.sessWg.Wait()
+	})
+	t.Run("headless pane without a preview stays quiet on completion", func(t *testing.T) {
+		pty, steps, processed := newChannelPTY(t)
+		d, target, client, _ := newManualSessionWithPTYs(t, pty)
+		clock := newCoordinatorMockClock(t, 8)
+		d.clock = clock.clock
+		rc := d.attachCoordinator(target, nil, client, true)
+		wakes := make(chan renderWake, 1)
+		rc.opts.wake = func(w renderWake) { wakes <- w }
+		target.mu.Lock()
+		target.client = nil
+		target.mu.Unlock()
+		rc.noteDetach(client)
+		d.sessWg.Add(1)
+		go d.ptyReader(target, target.tabs[0], target.tabs[0].focusedPane())
+		steps <- channelPTYStep{data: []byte("\x1b[?2026hpending")}
+		awaitPTYReadProcessed(t, processed)
+		drainCoordinatorTimers(clock)
+		steps <- channelPTYStep{data: []byte(" complete\x1b[?2026l")}
+		awaitPTYReadProcessed(t, processed)
+		require.Empty(t, drainCoordinatorTimers(clock))
+		requireNoWake(t, wakes)
+		steps <- channelPTYStep{err: io.EOF}
+		d.sessWg.Wait()
+	})
+}
+// S2 keeps damage pending until the owning render capture consumes it. In
+// particular, a pane becoming invisible must not let a PTY reader erase data
+// that a later attachment or picker preview needs to render.
+func TestPaneRenderableActiveAttachmentDoesNotScanPickerPreviews(t *testing.T) {
+	p, release := newBlockingPTY(t)
+	defer release()
+	d, sess, _, _ := newManualSessionWithPTYs(t, p)
+	tb := sess.activeTab()
+	pane := tb.focusedPane()
+	// Holding daemon ownership makes a picker scan block. An active attached
+	// tab must decide renderability from its session state without that scan.
+	d.mu.Lock()
+	defer d.mu.Unlock()
+	done := make(chan bool, 1)
+	go func() { done <- d.paneRenderable(sess, tb, pane) }()
+	require.True(t, awaitTestValue(t, done, "active attached renderability scanned picker previews"))
+}
+func TestNonRenderablePaneDamageRemainsPendingForCapture(t *testing.T) {
+	newFixture := func(t *testing.T) (*Daemon, *session, *tab, *pane, chan ports.Frame) {
+		t.Helper()
+		p, release := newBlockingPTY(t)
+		t.Cleanup(release)
+		d, sess, _, sends := newManualSessionWithPTYs(t, p)
+		tb := sess.tabs[0]
+		return d, sess, tb, tb.focusedPane(), sends
+	}
+	t.Run("headless, inactive, collapsed, and hidden panes retain damage without output", func(t *testing.T) {
+		for _, tt := range []struct {
+			name  string
+			setup func(*Daemon, *session, *tab, *pane)
+		}{
+			{name: "headless", setup: func(_ *Daemon, sess *session, _ *tab, _ *pane) { sess.client = nil }},
+			{name: "inactive tab", setup: func(_ *Daemon, sess *session, tb *tab, _ *pane) {
+				other := newTab(nil, domain.Size{Cols: 80, Rows: 23})
+				sess.tabs = append([]*tab{other}, sess.tabs...)
+				sess.active = 0
+				require.NotSame(t, other, tb)
+			}},
+			{name: "collapsed", setup: func(_ *Daemon, _ *session, tb *tab, p *pane) {
+				// The expanded leaf is deliberately absent from panes: p is the
+				// inactive stack leaf whose content is not composited.
+				tb.tree = &layout.Tree{Root: &layout.Node{Kind: layout.Stack, Children: []*layout.Node{layout.NewLeaf("visible"), layout.NewLeaf(p.id)}, Expanded: "visible"}, Focus: "visible"}
+			}},
+			{name: "hidden floating", setup: func(_ *Daemon, _ *session, tb *tab, p *pane) {
+				tb.floating = floatingSlot{state: floatingHidden, pane: p, generation: 1}
+				delete(tb.panes, p.id)
+			}},
+		} {
+			t.Run(tt.name, func(t *testing.T) {
+				d, sess, tb, p, sends := newFixture(t)
+				tt.setup(d, sess, tb, p)
+				p.screen.Write([]byte("damage"))
+				_ = d.paneRenderable(sess, tb, p)
+				require.NotEmpty(t, p.screen.Damage(), "only render capture may consume VT damage")
+				select {
+				case frame := <-sends:
+					t.Fatalf("non-renderable output must not compose or send: %#v", frame)
+				default:
+				}
+			})
+		}
+	})
+	t.Run("retains active and picker-preview pane damage", func(t *testing.T) {
+		d, sess, tb, p, _ := newFixture(t)
+		p.screen.Write([]byte("active"))
+		_ = d.paneRenderable(sess, tb, p)
+		require.NotEmpty(t, p.screen.Damage(), "active pane damage belongs to coordinator composition")
+		p.screen.ClearDamage()
+		sess.client = nil
+		viewer := &attachedClient{}
+		viewer.initOverlays()
+		viewer.overlays.pickerMu.Lock()
+		viewer.overlays.pickerPreview = tb
+		viewer.overlays.pickerMu.Unlock()
+		d.sessions["viewer"] = &session{id: "viewer", client: viewer}
+		p.screen.Write([]byte("preview"))
+		_ = d.paneRenderable(sess, tb, p)
+		require.NotEmpty(t, p.screen.Damage(), "picker preview damage must remain for coordinator composition")
+	})
+}
 func TestPTYWriteErrorIsLogged(t *testing.T) {
 	var logs bytes.Buffer
 	d := New(nil, stubClock{}, slog.New(slog.NewTextHandler(&logs, nil)))
 	}
 }
-func TestAltXClosesNonFinalTabScheduler(t *testing.T) {
-	p1, releasePTY1 := newBlockingPTY(t)
-	p2, releasePTY2 := newBlockingPTY(t)
-	clk := &signalClock{called: make(chan struct{})}
-	d, sess, ac, _ := newManualSessionWithPTYs(t, p1, p2)
-	d.clock = clk
-	defer releasePTY1()
-	defer releasePTY2()
-	d.sessWg.Add(1)
-	go d.scheduler(sess, sess.tabs[1], sess.tabs[1].focusedPane())
-	sess.tabs[1].focusedPane().dirty <- struct{}{}
-	<-clk.called
-	d.handleInput(sess, ac, []byte("\x1b2"))
-	d.handleInput(sess, ac, []byte("\x1b "))
-	d.handleInput(sess, ac, []byte("CLT\r"))
-	select {
-	case <-waitGroupDone(&d.sessWg):
-	case <-time.After(2 * time.Second):
-		t.Fatal("scheduler for removed tab did not exit")
-	}
-	require.Equal(t, 1, sessionCount(d))
-}
-func TestPTYEOFClosesNonFinalTabScheduler(t *testing.T) {
-	p1, releasePTY1 := newBlockingPTY(t)
-	p2, releasePTY2 := newBlockingPTY(t)
-	clk := &signalClock{called: make(chan struct{})}
-	d, sess, _, _ := newManualSessionWithPTYs(t, p1, p2)
-	d.clock = clk
-	defer releasePTY1()
-	defer releasePTY2()
-	d.sessWg.Add(2)
-	go d.scheduler(sess, sess.tabs[1], sess.tabs[1].focusedPane())
-	go d.ptyReader(sess, sess.tabs[1], sess.tabs[1].focusedPane())
-	sess.tabs[1].focusedPane().dirty <- struct{}{}
-	<-clk.called
-	releasePTY2()
-	select {
-	case <-waitGroupDone(&d.sessWg):
-	case <-time.After(2 * time.Second):
-		t.Fatal("scheduler for EOF-removed tab did not exit")
-	}
-	require.Equal(t, 1, sessionCount(d))
-}
 func TestAltXClosesFinalTabAndDetaches(t *testing.T) {
 	d, sess, ac, sends, releases := newManualTabSession(t, 1)
 	defer releases[0]()
 	d.sessWg.Wait()
 }
-func TestSchedulerDebounceCoalesces(t *testing.T) {
-	mc := portsmocks.NewMockClock(t)
-	mc.EXPECT().Now().Return(time.Now()).Maybe()
-	mt := portsmocks.NewMockTimer(t)
-	timerCh := make(chan time.Time, 1)
-	newTimerCalled := make(chan struct{}, 4)
-	mc.EXPECT().NewTimer(mock.Anything).RunAndReturn(func(time.Duration) ports.Timer {
-		select {
-		case newTimerCalled <- struct{}{}:
-		default:
-		}
-		return mt
-	}).Maybe()
-	mt.EXPECT().C().Return(timerCh).Maybe()
-	mt.EXPECT().Stop().Return(true).Maybe()
-	var outputs atomic.Int32
-	gotOutput := make(chan struct{}, 1)
-	tr := portsmocks.NewMockTransport(t)
-	tr.EXPECT().Close().Return(nil).Maybe()
-	tr.EXPECT().Send(mock.Anything).RunAndReturn(func(f ports.Frame) error {
-		if f.Type == ports.MsgOutput {
-			outputs.Add(1)
-			select {
-			case gotOutput <- struct{}{}:
-			default:
-			}
-		}
-		return nil
-	}).Maybe()
-	d := New(nil, mc, slog.New(slog.NewTextHandler(io.Discard, nil)))
-	sctx, cancel := context.WithCancel(context.Background())
-	win := newTab(newScriptPTY(nil), domain.Size{Cols: 20, Rows: 5})
-	ac := &attachedClient{tr: tr, output: newOutputStateStream()}
-	ac.initOverlays()
-	sess := &session{id: "s", name: "s", tabs: []*tab{win}, ctx: sctx, cancel: cancel, client: ac}
-	ac.setSession(sess)
-	win.mu.Lock()
-	win.focusedPane().screen.Write([]byte("hi"))
-	win.mu.Unlock()
-	d.sessWg.Add(1)
-	go d.scheduler(sess, win, win.focusedPane())
-	// First dirty opens the debounce window.
-	win.focusedPane().dirty <- struct{}{}
-	<-newTimerCalled
-	// A burst of further dirties inside the window must be absorbed.
-	for range 5 {
-		select {
-		case win.focusedPane().dirty <- struct{}{}:
-		default:
-		}
-	}
-	// Fire the timer once: exactly one render.
-	timerCh <- time.Now()
-	<-gotOutput
-	cancel()
-	d.sessWg.Wait()
-	require.Equal(t, int32(1), outputs.Load(), "N dirties in one tab must render exactly once")
-}
-func TestSchedulerAdaptiveDebounceFloodAndIdle(t *testing.T) {
-	delay := nextDebounceDelay(minDebounceInterval, 3)
-	require.Greater(t, delay, minDebounceInterval, "sustained flood should widen debounce")
-	delay = nextDebounceDelay(delay, 3)
-	require.Greater(t, delay, minDebounceInterval+debounceStep, "continued flood should keep adapting")
-}
-func TestSchedulerAdaptiveDebounceResetsAfterQuietPeriod(t *testing.T) {
-	delay := nextDebounceDelay(minDebounceInterval, 2)
-	require.Greater(t, delay, minDebounceInterval)
-	require.Equal(t, minDebounceInterval, nextDebounceDelay(delay, 0), "isolated update after quiet window should restore idle latency")
-}
 // --- resize ordering --------------------------------------------------------
 func TestResizePreservesLiveContentAndEvictsScrollback(t *testing.T) {
 	d.resize(sess, ac, domain.Size{Cols: 4, Rows: 4})
 	require.Equal(t, "2222", frameRowString(win.focusedPane().screen.Frame, 0))
 	require.Equal(t, "3333", frameRowString(win.focusedPane().screen.Frame, 1))
-	require.Equal(t, 2, win.focusedPane().scrollback.Len())
+	require.Equal(t, 2, win.focusedPane().history.Len())
-	require.Equal(t, "0000", cellsString(win.focusedPane().scrollback.View().Row(0)))
+	require.Equal(t, "0000", cellsString(win.focusedPane().history.View().Row(0)))
-	require.Equal(t, "1111", cellsString(win.focusedPane().scrollback.View().Row(1)))
+	require.Equal(t, "1111", cellsString(win.focusedPane().history.View().Row(1)))
 	d.resize(sess, ac, domain.Size{Cols: 6, Rows: 6})
 	require.Equal(t, "2222  ", frameRowString(win.focusedPane().screen.Frame, 0))
 	require.Equal(t, "3333  ", frameRowString(win.focusedPane().screen.Frame, 1))
-	require.Equal(t, 2, win.focusedPane().scrollback.Len())
+	require.Equal(t, 2, win.focusedPane().history.Len())
 }
 func frameRowString(f renderer.Frame, y int) string {
 	}
 }
-func TestRenderClearsDamageOnEmittingPaneWhenDetached(t *testing.T) {
-	tb := newTab(nil, domain.Size{Cols: 10, Rows: 3})
-	focused := tb.focusedPane()
-	emitting := newPane("pane-2", nil, domain.Size{Cols: 10, Rows: 3})
-	tb.panes[emitting.id] = emitting
-	tb.tree.Root = &layout.Node{Kind: layout.Split, Dir: layout.Horizontal, Children: []*layout.Node{layout.NewLeaf(focused.id), layout.NewLeaf(emitting.id)}}
-	tb.tree.Focus = focused.id
-	sess := &session{id: "s", name: "work", tabs: []*tab{tb}, active: 0}
-	d := newTestDaemon(t, nil, stubClock{})
-	focused.screen.Write([]byte("a"))
-	emitting.screen.Write([]byte("b"))
-	d.render(sess, tb, emitting)
-	require.NotEmpty(t, focused.screen.Damage(), "focused pane damage should be untouched by detached render for another pane")
-	require.Empty(t, emitting.screen.Damage(), "emitting pane damage should be drained")
-}
-func TestRenderWithNilTransportDoesNotAdvanceStateCounter(t *testing.T) {
-	tb := newTab(nil, domain.Size{Cols: 10, Rows: 3})
-	sess := &session{id: "s", name: "work", tabs: []*tab{tb}, active: 0}
-	ac := &attachedClient{output: newOutputStateStream(), size: domain.Size{Cols: 10, Rows: 4}}
-	ac.initOverlays()
-	sess.client = ac
-	ac.setSession(sess)
-	d := newTestDaemon(t, nil, stubClock{})
-	tb.focusedPane().screen.Write([]byte("x"))
-	d.render(sess, tb, tb.focusedPane())
-	require.Zero(t, ac.output.next)
-}
 func TestComposeTabFrameTwoPaneSplitBlitsDividersDimsAndTranslatesDamage(t *testing.T) {
 	win := newTab(nil, domain.Size{Cols: 41, Rows: 4})
 	left := win.focusedPane()
 		{
 			name:       "picker",
 			open:       func(d *Daemon, sess *session, ac *attachedClient) { d.enterPicker(sess, ac) },
-			close:      func(d *Daemon, sess *session, ac *attachedClient) { d.closePicker(ac); d.paint(sess, ac, true) },
+			close:      func(d *Daemon, sess *session, ac *attachedClient) { d.closePicker(ac); d.paint(sess, ac, true, nil) },
 			visible:    "Sessions",
 			notVisible: "Sessions",
 		},
 			open:  func(d *Daemon, sess *session, ac *attachedClient) { d.enterCopyMode(sess, ac) },
 			close: func(d *Daemon, _ *session, ac *attachedClient) { d.handleCopyInput(ac, []byte("q")) },
 			prepareScreen: func(sess *session) {
-				sess.tabs[0].focusedPane().scrollback = scopy.NewScrollback(4)
+				installTestHistory(sess.tabs[0].focusedPane(), vt.HistoryConfig{MaxRows: 4})
 			},
 			visible:      "[SCROLL]",
 			notVisible:   "[SCROLL]",
 			p, release := newBlockingPTY(t)
 			d, sess, ac, sends := newManualSessionWithPTYs(t, p)
 			defer release()
-			sess.tabs[0].focusedPane().screen.Write([]byte("live"))
 			if tt.prepareScreen != nil {
 				tt.prepareScreen(sess)
 			}
+			sess.tabs[0].focusedPane().screen.Write([]byte("live"))
 			if tt.wantRestored == "" {
 				tt.wantRestored = "live"
 			}
 	pane := sess.tabs[0].focusedPane()
 	pane.screen.Write([]byte("live"))
-	d.paint(sess, ac, true)
+	d.paint(sess, ac, true, nil)
 	_ = mustOutputData(t, sends)
-	require.True(t, ac.composed.valid)
+	require.True(t, ac.pipelineCache.valid)
-	require.NotContains(t, frameText(ac.composed.frame), "Rename session")
+	require.NotContains(t, frameText(ac.pipelineCache.frame), "Rename session")
 	d.enterPrompt(sess, ac, " Rename session ", "work", func(string) error { return nil })
 	shown := string(mustOutputData(t, sends))
 	require.Contains(t, shown, "Rename session")
-	require.False(t, ac.composed.valid, "overlay paint must not store the modal-mutated frame in the composed cache")
+	require.False(t, ac.pipelineCache.valid, "overlay paint must not store the modal-mutated frame in the composed cache")
-	require.NotContains(t, frameText(ac.composed.frame), "Rename session")
+	require.NotContains(t, frameText(ac.pipelineCache.frame), "Rename session")
 	d.closePrompt(ac)
-	d.paint(sess, ac, false)
+	d.paint(sess, ac, false, nil)
 	restored := string(mustOutputData(t, sends))
 	require.NotContains(t, restored, "Rename session")
-	require.True(t, ac.composed.valid)
+	require.True(t, ac.pipelineCache.valid)
-	cached := frameText(ac.composed.frame)
+	cached := frameText(ac.pipelineCache.frame)
 	require.Contains(t, cached, "live")
 	require.NotContains(t, cached, "Rename session")
 }
 	d.resize(sess, ac, newSize)
 	// This test verifies resize ordering, not the idle fallback. Consume the
 	// pending resize paint deterministically under the non-firing stub clock.
-	d.paint(sess, ac, false)
+	d.paint(sess, ac, false, nil)
 	require.Equal(t, 80, screenWidthAtResize, "pty.Resize must run before screen.Resize")
 	require.Equal(t, 100, win.focusedPane().screen.Frame.Width, "screen resized after pty")
 	win.focusedPane().screen.Write([]byte("x"))
 	win.mu.Unlock()
-	d.paint(sess, ac, true)
+	d.paint(sess, ac, true, nil)
 	require.Equal(t, 1, sessionCount(d), "ephemeral session survives failed client send")
 	sess.mu.Lock()
 	d.waitNotifies()
 }
-func TestSchedulerDefersPendingDirtyTimerDuringSynchronizedUpdate(t *testing.T) {
-	mc := portsmocks.NewMockClock(t)
-	mc.EXPECT().Now().Return(time.Now()).Maybe()
-	mt := portsmocks.NewMockTimer(t)
-	timerCh := make(chan time.Time, 1)
-	mc.EXPECT().NewTimer(mock.Anything).Return(mt).Maybe()
-	mt.EXPECT().C().Return(timerCh).Maybe()
-	mt.EXPECT().Stop().Return(true).Maybe()
-	var outputs atomic.Int32
-	gotOutput := make(chan struct{}, 1)
-	tr := portsmocks.NewMockTransport(t)
-	tr.EXPECT().Close().Return(nil).Maybe()
-	tr.EXPECT().Send(mock.Anything).RunAndReturn(func(f ports.Frame) error {
-		if f.Type == ports.MsgOutput {
-			outputs.Add(1)
-			select {
-			case gotOutput <- struct{}{}:
-			default:
-			}
-		}
-		return nil
-	}).Maybe()
-	d := New(nil, mc, slog.New(slog.NewTextHandler(io.Discard, nil)))
-	sctx, cancel := context.WithCancel(context.Background())
-	win := newTab(nil, domain.Size{Cols: 20, Rows: 5})
-	win.ctx, win.cancel = sctx, cancel
-	p := win.focusedPane()
-	p.ctx, p.cancel = sctx, cancel
-	ac := &attachedClient{tr: tr, output: newOutputStateStream()}
-	ac.initOverlays()
-	sess := &session{id: "sync", name: "sync", tabs: []*tab{win}, ctx: sctx, cancel: cancel, client: ac}
-	ac.setSession(sess)
-	p.mu.Lock()
-	p.screen.Write([]byte("before"))
-	p.mu.Unlock()
-	d.sessWg.Add(1)
-	go d.scheduler(sess, win, p)
-	p.dirty <- struct{}{}
-	p.mu.Lock()
-	p.screen.Write([]byte("\x1b[?2026hafter"))
-	p.mu.Unlock()
-	timerCh <- time.Now()
-	select {
-	case <-gotOutput:
-		t.Fatal("scheduler rendered while synchronized update was active")
-	case <-time.After(20 * time.Millisecond):
-	}
-	require.Equal(t, int32(0), outputs.Load())
-	p.mu.Lock()
-	p.screen.Write([]byte(" done\x1b[?2026l"))
-	p.mu.Unlock()
-	p.flush <- struct{}{}
-	select {
-	case <-gotOutput:
-	case <-time.After(time.Second):
-		t.Fatal("sync end did not flush deferred damage")
-	}
-	cancel()
-	d.sessWg.Wait()
-}
 func TestPTYQueryGetsResponseWrittenBackToPTY(t *testing.T) {
 	d := newTestDaemon(t, nil, stubClock{})
 	p := portsmocks.NewMockPTY(t)
 	}
 }
-func TestPTYReaderSameReadSynchronizedUpdateStartEndFlushesImmediately(t *testing.T) {
-	d := newTestDaemon(t, nil, stubClock{})
-	p := portsmocks.NewMockPTY(t)
-	chunks := [][]byte{[]byte("\x1b[?2026hhello\x1b[?2026l")}
-	p.EXPECT().Read(mock.Anything).RunAndReturn(func(buf []byte) (int, error) {
-		if len(chunks) == 0 {
-			return 0, io.EOF
-		}
-		n := copy(buf, chunks[0])
-		chunks = chunks[1:]
-		return n, nil
-	})
-	p.EXPECT().Close().Return(nil).Maybe()
-	sctx, cancel := context.WithCancel(context.Background())
-	win := newTestTabWithContext(p, sctx, cancel)
-	sess := &session{id: "sync", name: "sync", tabs: []*tab{win}, ctx: sctx, cancel: cancel}
-	d.sessions[sess.id] = sess
-	d.sessWg.Add(1)
-	d.ptyReader(sess, win, win.focusedPane())
-	select {
-	case <-win.focusedPane().dirty:
-		t.Fatal("same-read synchronized update end should request flush, not dirty")
-	default:
-	}
-	select {
-	case <-win.focusedPane().flush:
-	case <-time.After(time.Second):
-		t.Fatal("same-read synchronized update end did not request immediate flush")
-	}
-}
-func TestPTYReaderDefersDirtyDuringSynchronizedUpdateAndFlushesAtEnd(t *testing.T) {
-	d := newTestDaemon(t, nil, stubClock{})
-	p := portsmocks.NewMockPTY(t)
-	chunks := [][]byte{[]byte("\x1b[?2026hhello"), []byte(" world\x1b[?2026l")}
-	p.EXPECT().Read(mock.Anything).RunAndReturn(func(buf []byte) (int, error) {
-		if len(chunks) == 0 {
-			return 0, io.EOF
-		}
-		n := copy(buf, chunks[0])
-		chunks = chunks[1:]
-		return n, nil
-	})
-	p.EXPECT().Close().Return(nil).Maybe()
-	sctx, cancel := context.WithCancel(context.Background())
-	win := newTestTabWithContext(p, sctx, cancel)
-	sess := &session{id: "sync", name: "sync", tabs: []*tab{win}, ctx: sctx, cancel: cancel}
-	d.sessions[sess.id] = sess
-	d.sessWg.Add(1)
-	d.ptyReader(sess, win, win.focusedPane())
-	select {
-	case <-win.focusedPane().dirty:
-		t.Fatal("dirty signaled while synchronized update was active")
-	default:
-	}
-	select {
-	case <-win.focusedPane().flush:
-	case <-time.After(time.Second):
-		t.Fatal("sync end did not request immediate flush")
-	}
-}
-func TestSyncUpdateWatchdogAbandonedSyncRenders(t *testing.T) {
-	clk := &signalClock{timers: make(chan *signalTimer, 1)}
-	d := newTestDaemon(t, nil, clk)
-	p := portsmocks.NewMockPTY(t)
-	release := make(chan struct{})
-	read := 0
-	p.EXPECT().Read(mock.Anything).RunAndReturn(func(buf []byte) (int, error) {
-		if read == 0 {
-			read++
-			return copy(buf, []byte("\x1b[?2026habandoned")), nil
-		}
-		<-release
-		return 0, io.EOF
-	})
-	p.EXPECT().Close().Return(nil).Maybe()
-	tr, sends := newCapturingTransport(t)
-	sctx, cancel := context.WithCancel(context.Background())
-	win := newTestTabWithContext(p, sctx, cancel)
-	ac := &attachedClient{tr: tr, output: newOutputStateStream()}
-	ac.initOverlays()
-	sess := &session{id: "sync", name: "sync", tabs: []*tab{win}, ctx: sctx, cancel: cancel, client: ac}
-	ac.setSession(sess)
-	d.sessions[sess.id] = sess
-	d.sessWg.Add(2)
-	go d.scheduler(sess, win, win.focusedPane())
-	go d.ptyReader(sess, win, win.focusedPane())
-	timer := <-clk.timers
-	timer.ch <- time.Now()
-	awaitFrame(t, sends, ports.MsgOutput)
-	win.mu.Lock()
-	active := win.focusedPane().screen.SyncUpdateActive()
-	got := screenLineText(win.focusedPane().screen, 0)
-	win.mu.Unlock()
-	require.False(t, active)
-	require.Contains(t, got, "abandoned")
-	close(release)
-	cancel()
-	d.sessWg.Wait()
-}
-func TestSyncUpdateWatchdogStaleGenerationNoopAfterEnd(t *testing.T) {
-	clk := &signalClock{timers: make(chan *signalTimer, 1)}
-	d := newTestDaemon(t, nil, clk)
-	p := portsmocks.NewMockPTY(t)
-	release := make(chan struct{})
-	chunks := [][]byte{[]byte("\x1b[?2026hhello"), []byte(" world\x1b[?2026l")}
-	p.EXPECT().Read(mock.Anything).RunAndReturn(func(buf []byte) (int, error) {
-		if len(chunks) > 0 {
-			n := copy(buf, chunks[0])
-			chunks = chunks[1:]
-			return n, nil
-		}
-		<-release
-		return 0, io.EOF
-	})
-	p.EXPECT().Close().Return(nil).Maybe()
-	tr, sends := newCapturingTransport(t)
-	sctx, cancel := context.WithCancel(context.Background())
-	win := newTestTabWithContext(p, sctx, cancel)
-	ac := &attachedClient{tr: tr, output: newOutputStateStream()}
-	ac.initOverlays()
-	sess := &session{id: "sync", name: "sync", tabs: []*tab{win}, ctx: sctx, cancel: cancel, client: ac}
-	ac.setSession(sess)
-	d.sessions[sess.id] = sess
-	d.sessWg.Add(2)
-	go d.scheduler(sess, win, win.focusedPane())
-	go d.ptyReader(sess, win, win.focusedPane())
-	timer := <-clk.timers
-	awaitFrame(t, sends, ports.MsgOutput)
-	timer.ch <- time.Now()
-	select {
-	case f := <-sends:
-		if f.Type == ports.MsgOutput {
-			t.Fatal("stale synchronized update watchdog produced an extra render")
-		}
-	case <-time.After(50 * time.Millisecond):
-	}
-	win.mu.Lock()
-	active := win.focusedPane().screen.SyncUpdateActive()
-	win.mu.Unlock()
-	require.False(t, active)
-	close(release)
-	cancel()
-	d.sessWg.Wait()
-}
 func TestCursorTailVisibleHideAndMoveOnly(t *testing.T) {
 	p, _ := newBlockingPTY(t)
 	d, sess, ac, sends := newManualSessionWithPTYs(t, p)
 	win := sess.tabs[0]
 	win.focusedPane().screen.Write([]byte("A"))
-	d.paint(sess, ac, true)
+	d.paint(sess, ac, true, nil)
 	data := mustOutputData(t, sends)
 	require.Contains(t, string(data), "\x1b[1 q")
 	require.Contains(t, string(data), "\x1b[?25h")
 	win.focusedPane().screen.Write([]byte("\x1b[2;3H"))
-	d.paint(sess, ac, false)
+	d.paint(sess, ac, false, nil)
 	data = mustOutputData(t, sends)
 	require.Contains(t, string(data), "\x1b[3;3H")
 	require.Contains(t, string(data), "\x1b[?25h")
 	win.focusedPane().screen.Write([]byte("\x1b[?25l"))
-	d.paint(sess, ac, false)
+	d.paint(sess, ac, false, nil)
 	data = mustOutputData(t, sends)
 	require.Contains(t, string(data), "\x1b[?25l")
 }
 	floating.popupGeometry = committed
 	installTestFloating(sess.activeTab(), floating, true)
-	d.paint(sess, ac, true)
+	d.paint(sess, ac, true, nil)
 	data := mustOutputData(t, sends)
 	want := committed.translate(contentArea.X, contentArea.Y)
 	require.Contains(t, string(data), cursorCSI(want.Inner.Y+1, 1))
 	require.True(t, ok)
 	rightContent := placementContent(placements, right.id)
-	d.paint(sess, ac, true)
+	d.paint(sess, ac, true, nil)
 	data := mustOutputData(t, sends)
 	want := cursorCSI(rightContent.Y+right.screen.CursorRow()+2, rightContent.X+right.screen.CursorCol()+1)
 	require.Contains(t, string(data), want)
 	twoContent := placementContent(placements, two.id)
 	require.Greater(t, twoContent.Y, 0, "stack title bars should offset content")
-	d.paint(sess, ac, true)
+	d.paint(sess, ac, true, nil)
 	data := mustOutputData(t, sends)
 	want := cursorCSI(twoContent.Y+two.screen.CursorRow()+2, twoContent.X+two.screen.CursorCol()+1)
 	require.Contains(t, string(data), want)
 	}
 	p := newPane("floating", nil, domain.Size{Cols: 18, Rows: 2})
 	p.screen.Write([]byte("ab\r\ncd"))
-	document := scopy.NewSnapshot(p.scrollback, p.screen.Frame)
+	document := scopy.NewSnapshot(p.history, p.screen.Frame)
 	mode := scopy.NewMode(document)
 	target := domain.Rect{X: 2, Y: 3, Width: 18, Height: 2}
 	defer releaseFloating()
 	d, sess, ac, sends := newManualSessionWithPTYs(t, normal)
 	fp := newPane("floating", floatingPTY, domain.Size{Cols: 20, Rows: 3})
-	fp.scrollback.Append(testRow("flt-old"))
+	fp.history.Append(testRow("flt-old"))
 	fp.screen.Write([]byte("flt-live"))
 	installTestFloating(sess.activeTab(), fp, true)

A internal/usecase/daemon/render_timer.go +90 -0

@@ -0,0 +1,90 @@
	1	+	package daemon
	2	+
	3	+	import (
	4	+	"sync"
	5	+	"time"
	6	+
	7	+	"github.com/bnema/vev/internal/ports"
	8	+	)
	9	+
	10	+	// timerToken is the complete, immutable identity of one timer worker. A lane
	11	+	// owns either this token or nothing; cancellation and completion are never
	12	+	// retained separately where a stale callback could reacquire them.
	13	+	type timerToken struct {
	14	+	generation uint64
	15	+	timer ports.Timer
	16	+	cancel chan struct{}
	17	+	done chan struct{}
	18	+	}
	19	+
	20	+	// timerLane is coordinator-owned and must be accessed while renderCoordinator.mu
	21	+	// is held. Detach removes the complete token before any external Timer call.
	22	+	type timerLane struct {
	23	+	generation uint64
	24	+	token *timerToken
	25	+	}
	26	+
	27	+	func (l timerLane) replaceLocked() (uint64, timerToken) {
	28	+	l.generation++
	29	+	return l.generation, l.detachLocked()
	30	+	}
	31	+
	32	+	func (l timerLane) detachLocked() timerToken {
	33	+	token := l.token
	34	+	l.token = nil
	35	+	if token != nil {
	36	+	close(token.cancel)
	37	+	}
	38	+	return token
	39	+	}
	40	+
	41	+	func (l timerLane) publishLocked(generation uint64, timer ports.Timer) timerToken {
	42	+	if l.generation != generation \|\| l.token != nil {
	43	+	return nil
	44	+	}
	45	+	token := &timerToken{
	46	+	generation: generation,
	47	+	timer: timer,
	48	+	cancel: make(chan struct{}),
	49	+	done: make(chan struct{}),
	50	+	}
	51	+	l.token = token
	52	+	return token
	53	+	}
	54	+
	55	+	func (l timerLane) clearLocked(token timerToken) bool {
	56	+	if token == nil \|\| l.generation != token.generation \|\| l.token != token {
	57	+	return false
	58	+	}
	59	+	l.token = nil
	60	+	return true
	61	+	}
	62	+
	63	+	// timerSupervisor owns worker accounting. Workers are registered only while
	64	+	// coordinator ownership is held and terminal teardown rejects registrations
	65	+	// before Wait begins, so Add cannot race the terminal Wait.
	66	+	type timerSupervisor struct{ workers sync.WaitGroup }
	67	+
	68	+	func (s timerSupervisor) startLocked(token timerToken, timerC <-chan time.Time, fire func()) {
	69	+	s.workers.Add(1)
	70	+	go func() {
	71	+	defer s.workers.Done()
	72	+	defer close(token.done)
	73	+	select {
	74	+	case <-timerC:
	75	+	fire()
	76	+	case <-token.cancel:
	77	+	}
	78	+	}()
	79	+	}
	80	+
	81	+	func (s *timerSupervisor) wait() { s.workers.Wait() }
	82	+
	83	+	// stopDetachedTimer is intentionally non-blocking. Ordinary lifecycle and
	84	+	// replacement transitions may cancel/stop stale workers but callback stacks
	85	+	// never wait for a worker to finish.
	86	+	func stopDetachedTimer(token *timerToken) {
	87	+	if token != nil && token.timer != nil {
	88	+	token.timer.Stop()
	89	+	}
	90	+	}

A internal/usecase/daemon/render_timer_test.go +144 -0

@@ -0,0 +1,144 @@
	1	+	package daemon
	2	+
	3	+	import (
	4	+	"testing"
	5	+	"time"
	6	+
	7	+	"github.com/stretchr/testify/require"
	8	+
	9	+	"github.com/bnema/vev/internal/domain"
	10	+	portsmocks "github.com/bnema/vev/internal/ports/mocks"
	11	+	)
	12	+
	13	+	func TestTimerLaneDetachRemovesCompleteToken(t *testing.T) {
	14	+	lane := timerLane{generation: 1}
	15	+	timer := portsmocks.NewMockTimer(t)
	16	+	token := lane.publishLocked(1, timer)
	17	+	require.NotNil(t, token)
	18	+	detached := lane.detachLocked()
	19	+	require.Same(t, token, detached)
	20	+	require.Nil(t, lane.token)
	21	+	select {
	22	+	case <-token.cancel:
	23	+	default:
	24	+	t.Fatal("detach did not cancel token")
	25	+	}
	26	+	require.Nil(t, lane.publishLocked(0, timer), "stale generation reacquired detached lane")
	27	+	}
	28	+
	29	+	func TestRenderDeadlineCallbackCanInvalidateAgainWithoutJoiningItself(t *testing.T) {
	30	+	clock := newCoordinatorMockClock(t, 2)
	31	+	ac := &attachedClient{}
	32	+	wakeReturned := make(chan struct{})
	33	+	var rc *renderCoordinator
	34	+	rc = newRenderCoordinator(renderCoordinatorOptions{
	35	+	clock: clock.clock,
	36	+	wake: func(renderWake) {
	37	+	require.True(t, rc.invalidateForAttachment(ac, renderInvalidation{class: invalidateUrgent, producer: "reentrant-wake-test"}))
	38	+	close(wakeReturned)
	39	+	},
	40	+	})
	41	+	rc.attach(ac)
	42	+	t.Cleanup(func() {
	43	+	rc.beginSessionTeardown().finish()
	44	+	rc.waitForTimerWorkers()
	45	+	})
	46	+	require.True(t, rc.invalidateForAttachment(ac, renderInvalidation{class: invalidateUrgent, producer: "initial-wake-test"}))
	47	+	awaitCoordinatorScheduledTimer(t, clock).ch <- time.Time{}
	48	+	awaitTestCompletion(t, wakeReturned, "timer callback deadlocked while publishing its next invalidation")
	49	+	}
	50	+
	51	+	func TestRenderTimerCallbacksCanReplaceTheirOwnLane(t *testing.T) {
	52	+	cases := []struct {
	53	+	name string
	54	+	run func(testing.T, renderCoordinator, coordinatorMockClock, attachedClient)
	55	+	}{
	56	+	{
	57	+	name: "normal wake invalidates", run: func(t testing.T, rc renderCoordinator, clock coordinatorMockClock, ac attachedClient) {
	58	+	woke := make(chan struct{})
	59	+	rc.opts.wake = func(renderWake) {
	60	+	rc.invalidateForAttachment(ac, renderInvalidation{class: invalidateUrgent, producer: "table-normal"})
	61	+	close(woke)
	62	+	}
	63	+	rc.invalidateForAttachment(ac, renderInvalidation{class: invalidateUrgent, producer: "table-normal"})
	64	+	awaitCoordinatorScheduledTimer(t, clock).ch <- time.Time{}
	65	+	awaitTestCompletion(t, woke, "normal callback did not reenter invalidation")
	66	+	},
	67	+	},
	68	+	{
	69	+	name: "resize callback schedules resize", run: func(t testing.T, rc renderCoordinator, clock coordinatorMockClock, ac attachedClient) {
	70	+	lease := rc.attachmentLease(ac)
	71	+	rc.scheduleResizeForLease(domain.Size{Cols: 80, Rows: 24}, ac, lease, func(uint64) {
	72	+	rc.scheduleResizeForLease(domain.Size{Cols: 81, Rows: 24}, ac, lease, func(uint64) {})
	73	+	})
	74	+	timer := awaitCoordinatorScheduledTimer(t, clock)
	75	+	done := captureResizeCallbackDone(t, rc)
	76	+	timer.ch <- time.Time{}
	77	+	_ = awaitTestValue(t, clock.timers, "resize callback did not schedule replacement")
	78	+	awaitTestCompletion(t, done, "resize callback did not complete")
	79	+	},
	80	+	},
	81	+	{
	82	+	name: "retry callback schedules retry", run: func(t testing.T, rc renderCoordinator, clock coordinatorMockClock, ac attachedClient) {
	83	+	lease := rc.attachmentLease(ac)
	84	+	epoch := rc.recordResizeRequestForLease(domain.Size{Cols: 80, Rows: 24}, ac, lease)
	85	+	require.True(t, rc.resizeCurrentForLease(epoch, ac, lease, true))
	86	+	rc.scheduleResizeRetryForLease(epoch, ac, lease, func() { rc.scheduleResizeRetryForLease(epoch, ac, lease, func() {}) })
	87	+	awaitCoordinatorScheduledTimer(t, clock).ch <- time.Time{}
	88	+	_ = awaitTestValue(t, clock.timers, "retry callback did not schedule replacement")
	89	+	},
	90	+	},
	91	+	{
	92	+	name: "sync force transitions sync generation", run: func(t testing.T, rc renderCoordinator, clock coordinatorMockClock, _ attachedClient) {
	93	+	p := &pane{}
	94	+	rc.noteSyncBegin(p, 1, func() { rc.noteSyncBegin(p, 2) })
	95	+	awaitCoordinatorScheduledTimer(t, clock).ch <- time.Time{}
	96	+	_ = awaitTestValue(t, clock.timers, "sync force did not schedule replacement")
	97	+	},
	98	+	},
	99	+	}
	100	+	for _, tc := range cases {
	101	+	t.Run(tc.name, func(t *testing.T) {
	102	+	clock := newCoordinatorMockClock(t, 4)
	103	+	rc, ac := newRenderCoordinator(renderCoordinatorOptions{clock: clock.clock}), &attachedClient{}
	104	+	rc.attach(ac)
	105	+	tc.run(t, rc, clock, ac)
	106	+	rc.beginSessionTeardown().finish()
	107	+	rc.waitForTimerWorkers()
	108	+	})
	109	+	}
	110	+	}
	111	+
	112	+	func TestRenderTimerScheduleRacesTerminalTeardown(t *testing.T) {
	113	+	for range 32 {
	114	+	clock := newCoordinatorMockClock(t, 2)
	115	+	rc, ac := newRenderCoordinator(renderCoordinatorOptions{clock: clock.clock}), &attachedClient{}
	116	+	rc.attach(ac)
	117	+	start := make(chan struct{})
	118	+	scheduled := make(chan struct{})
	119	+	go func() {
	120	+	<-start
	121	+	rc.invalidateForAttachment(ac, renderInvalidation{class: invalidateUrgent, producer: "race"})
	122	+	close(scheduled)
	123	+	}()
	124	+	close(start)
	125	+	cleanup := rc.beginSessionTeardown()
	126	+	awaitTestCompletion(t, scheduled, "schedule raced terminal teardown")
	127	+	cleanup.finish()
	128	+	rc.waitForTimerWorkers()
	129	+	rc.mu.Lock()
	130	+	require.True(t, rc.torndown)
	131	+	require.Nil(t, rc.normalLane.token)
	132	+	rc.mu.Unlock()
	133	+	}
	134	+	}
	135	+
	136	+	func TestTimerSupervisorCancelsNeverFiringWorker(t *testing.T) {
	137	+	var lane timerLane
	138	+	generation, _ := lane.replaceLocked()
	139	+	token := lane.publishLocked(generation, portsmocks.NewMockTimer(t))
	140	+	var supervisor timerSupervisor
	141	+	supervisor.startLocked(token, nil, func() { t.Fatal("cancelled worker fired") })
	142	+	lane.detachLocked()
	143	+	supervisor.wait()
	144	+	}

M internal/usecase/daemon/resume.go +6 -5

 	if !d.prepareParkAttachment(sess, ac) {
 		return false
 	}
-	// A parked attachment has no valid visible destination. Retire both the
+	// A parked attachment has no session owner. Release pane snapshots before
-	// deferred diff and any resize-idle callback before registering its grace
+	// publishing it in the daemon registry so headless pane/session teardown
-	// timer; resume performs a fresh full paint.
+	// cannot leave closed panes retained through its capture cache. This must
-	ac.cancelResizePaint()
+	// remain outside d.mu: sendMu is ordered before the daemon lock.
-	ac.clearPaintDeferred()
+	ac.clearCaptureFrames()
 	d.mu.Lock()
 	if d.closing {
 		d.mu.Unlock()
 	sess.client = ac
 	sess.terminal = terminalEnv{TrueColor: h.TrueColor}
 	sess.mu.Unlock()
+	d.attachCoordinator(sess, nil, ac, false)
 	d.touchMRU(sess)
 	d.log.Info("client resumed", "session", sess.name)
 	return sess, ac, true, nil

M internal/usecase/daemon/resume_test.go +163 -1


@@ -2,6 +2,7 @@ package daemon
2			2
3		import (	3		import (
4		"errors"	4		"errors"
			5	+	"strings"
5		"sync"	6		"sync"
6		"testing"	7		"testing"
7		"time"	8		"time"

@@ -10,6 +11,8 @@ import (
10			11
11		"github.com/bnema/vev/internal/domain"	12		"github.com/bnema/vev/internal/domain"
12		"github.com/bnema/vev/internal/ports"	13		"github.com/bnema/vev/internal/ports"
			14	+	"github.com/bnema/vev/internal/usecase/layout"
			15	+	"github.com/bnema/vev/pkg/vt"
13		)	16		)
14			17
15		type closeTrackingTransport struct {	18		type closeTrackingTransport struct {

@@ -77,6 +80,110 @@ func TestNamedLinkLossParks(t *testing.T) {
77		require.True(t, parked, "named resume-capable link loss is parked")	80		require.True(t, parked, "named resume-capable link loss is parked")
78		}	81		}
79			82
			83	+	func TestBoundedSendTimeoutCannotTargetResumedTransport(t *testing.T) {
			84	+	pty, release := newBlockingPTY(t)
			85	+	defer release()
			86	+	d := newTestDaemon(t, newFactory(t, pty), stubClock{})
			87	+	oldTransport := &closeTrackingTransport{}
			88	+	sess, ac, err := d.route(helloResumeCapable(ports.IntentNew, "bounded-send-resume", 0), oldTransport)
			89	+	require.NoError(t, err)
			90	+	token := ac.resumeToken
			91	+	d.clientGone(sess, ac, oldTransport, false)
			92	+
			93	+	clock := &signalClock{timers: make(chan *signalTimer, 1)}
			94	+	d.clock = clock
			95	+	ac.sendMu.Lock()
			96	+	expected := ac.transportSnapshot()
			97	+	orphanDone := make(chan struct{})
			98	+	type boundedResult struct {
			99	+	transport ports.Transport
			100	+	err error
			101	+	}
			102	+	result := make(chan boundedResult, 1)
			103	+	go func() {
			104	+	transport, sendErr := d.boundedSendWithTimeout(time.Second, expected.transport, func() error {
			105	+	defer close(orphanDone)
			106	+	return ac.sendExpectedTransport(expected, frameDetached(ports.ReasonDetach))
			107	+	})
			108	+	result <- boundedResult{transport: transport, err: sendErr}
			109	+	}()
			110	+
			111	+	// The worker is parked behind sendMu, so it has not observed a transport.
			112	+	// Expire its deadline before resuming this same attachment with a new link.
			113	+	timer := <-clock.timers
			114	+	timer.ch <- time.Time{}
			115	+	got := <-result
			116	+	require.ErrorIs(t, got.err, errSendTimedOut)
			117	+	require.Same(t, oldTransport, got.transport)
			118	+	require.NoError(t, ac.closeCapturedTransport(got.transport))
			119	+
			120	+	newTransport := &closeTrackingTransport{}
			121	+	d.mu.Lock()
			122	+	resumedSess, resumedAC, ok, err := d.resumeParkedLocked(
			123	+	helloResumeCapable(ports.IntentResume, sess.name, token),
			124	+	newTransport,
			125	+	domain.Size{Cols: 80, Rows: 24},
			126	+	)
			127	+	d.mu.Unlock()
			128	+	require.NoError(t, err)
			129	+	require.True(t, ok)
			130	+	require.Same(t, sess, resumedSess)
			131	+	require.Same(t, ac, resumedAC)
			132	+
			133	+	ac.sendMu.Unlock()
			134	+	<-orphanDone
			135	+	require.Empty(t, newTransport.Sends(), "orphaned send must not write to the resumed transport")
			136	+	require.False(t, newTransport.Closed(), "orphaned send must not close the resumed transport")
			137	+	}
			138	+
			139	+	func TestHandleHelloResumeDefersFreshOutputUntilWelcome(t *testing.T) {
			140	+	pty, releasePTY := newBlockingPTY(t)
			141	+	defer releasePTY()
			142	+	clock := newCoordinatorMockClock(t, 4)
			143	+	d := newTestDaemon(t, newFactorySeq(t, pty), clock.clock)
			144	+	oldTransport := &closeTrackingTransport{}
			145	+	sess, ac, err := d.route(helloResumeCapable(ports.IntentNew, "resume-welcome-gate", 0), oldTransport)
			146	+	require.NoError(t, err)
			147	+	token := ac.resumeToken
			148	+	d.clientGone(sess, ac, oldTransport, false)
			149	+
			150	+	tr := newWelcomeBlockingTransport(t)
			151	+	done := make(chan struct{})
			152	+	resumeHello := helloResumeCapable(ports.IntentResume, sess.name, token)
			153	+	go func() {
			154	+	d.handleHello(tr.tr, ports.Frame{Type: ports.MsgHello, Payload: ports.MarshalHello(resumeHello)})
			155	+	close(done)
			156	+	}()
			157	+
			158	+	<-tr.welcomeEntered
			159	+	welcome := <-tr.sends
			160	+	require.Equal(t, ports.MsgWelcome, welcome.Type)
			161	+	sess.mu.Lock()
			162	+	resumed := sess.client
			163	+	sess.mu.Unlock()
			164	+	require.Same(t, ac, resumed)
			165	+
			166	+	d.invalidateRender(sess, resumed, true, "resume-welcome-gate-test")
			167	+	timer := awaitLatestCoordinatorTimer(t, clock)
			168	+	rc := sess.renderCoordinator()
			169	+	rc.mu.Lock()
			170	+	workerDone := rc.normalLane.token.done
			171	+	rc.mu.Unlock()
			172	+	require.NotNil(t, workerDone)
			173	+	timer.ch <- time.Time{}
			174	+	awaitTestCompletion(t, workerDone, "coordinator deadline worker did not complete")
			175	+	requireNoCoordinatorOutputFrame(t, tr.sends)
			176	+
			177	+	tr.release()
			178	+	output := awaitFrame(t, tr.sends, ports.MsgOutput)
			179	+	first, err := ports.UnmarshalOutput(output.Payload)
			180	+	require.NoError(t, err)
			181	+	require.Zero(t, first.BaseStateNum)
			182	+	tr.finish()
			183	+	<-done
			184	+	requireNoCoordinatorOutputFrame(t, tr.sends)
			185	+	}
			186	+
80		func TestEphemeralLinkLossParksAndResumes(t *testing.T) {	187		func TestEphemeralLinkLossParksAndResumes(t *testing.T) {
81		pty, release := newBlockingPTY(t)	188		pty, release := newBlockingPTY(t)
82		defer release()	189		defer release()

@@ -228,6 +335,7 @@ func TestResumeRebasesFullOutputWindowBeforeFirstPaint(t *testing.T) {
228		resumedSess, resumedAC, err := d.route(helloResumeCapable(ports.IntentResume, "work", token), newTr)	335		resumedSess, resumedAC, err := d.route(helloResumeCapable(ports.IntentResume, "work", token), newTr)
229		require.NoError(t, err)	336		require.NoError(t, err)
230		require.Same(t, ac, resumedAC)	337		require.Same(t, ac, resumedAC)
			338	+	require.True(t, resumedSess.renderCoordinator().markAttachmentReady(resumedSess.renderCoordinator().attachmentLease(resumedAC)))
231		d.firstPaint(resumedSess, resumedAC, resumedAC.size)	339		d.firstPaint(resumedSess, resumedAC, resumedAC.size)
232			340
233		sends := newTr.Sends()	341		sends := newTr.Sends()

@@ -239,7 +347,7 @@ func TestResumeRebasesFullOutputWindowBeforeFirstPaint(t *testing.T) {
239		resumedAC.ackOutputState(first.NewStateNum)	347		resumedAC.ackOutputState(first.NewStateNum)
240			348
241		resumedSess.tabs[0].focusedPane().screen.Write([]byte("A"))	349		resumedSess.tabs[0].focusedPane().screen.Write([]byte("A"))
242	-	d.paint(resumedSess, resumedAC, false)	350	+	d.paint(resumedSess, resumedAC, false, nil)
243		sends = newTr.Sends()	351		sends = newTr.Sends()
244		require.Len(t, sends, 2)	352		require.Len(t, sends, 2)
245		second, err := ports.UnmarshalOutput(sends[1].Payload)	353		second, err := ports.UnmarshalOutput(sends[1].Payload)

@@ -247,6 +355,60 @@ func TestResumeRebasesFullOutputWindowBeforeFirstPaint(t *testing.T) {
247		require.Equal(t, first.NewStateNum, second.BaseStateNum)	355		require.Equal(t, first.NewStateNum, second.BaseStateNum)
248		}	356		}
249			357
			358	+	func TestParkingReleasesPaneCapturesBeforeHeadlessCloseAndResume(t *testing.T) {
			359	+	pty, release := newBlockingPTY(t)
			360	+	defer release()
			361	+	d := newTestDaemon(t, newFactorySeq(t, pty), stubClock{})
			362	+	oldTransport := &closeTrackingTransport{}
			363	+	sess, ac, err := d.route(helloResumeCapable(ports.IntentNew, "work", 0), oldTransport)
			364	+	require.NoError(t, err)
			365	+
			366	+	tb := sess.tabs[0]
			367	+	survivor := tb.panes["pane-1"]
			368	+	closed := newPane("pane-2", nil, domain.Size{Cols: 40, Rows: 23})
			369	+	tb.mu.Lock()
			370	+	tb.tree = &layout.Tree{Root: &layout.Node{Kind: layout.Split, Dir: layout.Horizontal, Children: []*layout.Node{layout.NewLeaf("pane-1"), layout.NewLeaf("pane-2")}}, Focus: "pane-1"}
			371	+	tb.panes[closed.id] = closed
			372	+	d.applyLayoutLocked(tb)
			373	+	tb.mu.Unlock()
			374	+	survivor.screen.Write([]byte("survivor"))
			375	+	closed.screen.Write([]byte("closed"))
			376	+	d.paint(sess, ac, true, nil)
			377	+
			378	+	ac.sendMu.Lock()
			379	+	require.Contains(t, ac.captureFrames, closed, "fixture must render and capture the pane before parking")
			380	+	ac.sendMu.Unlock()
			381	+	token := ac.resumeToken
			382	+	require.True(t, sess.detachIfCurrent(ac))
			383	+	require.True(t, d.parkAttachment(sess, ac))
			384	+
			385	+	// Headless close cannot find the parked attachment through sess.client.
			386	+	// Its capture must already have been released before the attachment parked.
			387	+	require.NoError(t, d.closePane(sess, tb, closed.id, nil, false))
			388	+	ac.sendMu.Lock()
			389	+	require.NotContains(t, ac.captureFrames, closed, "parked attachment must not retain a pane closed while headless")
			390	+	ac.sendMu.Unlock()
			391	+
			392	+	newTransport := &closeTrackingTransport{}
			393	+	resumedSess, resumedAC, err := d.route(helloResumeCapable(ports.IntentResume, sess.name, token), newTransport)
			394	+	require.NoError(t, err)
			395	+	require.Same(t, sess, resumedSess)
			396	+	require.Same(t, ac, resumedAC)
			397	+	require.True(t, resumedSess.renderCoordinator().markAttachmentReady(resumedSess.renderCoordinator().attachmentLease(resumedAC)))
			398	+	d.firstPaint(resumedSess, resumedAC, resumedAC.size)
			399	+
			400	+	sends := newTransport.Sends()
			401	+	require.Len(t, sends, 1)
			402	+	output, err := ports.UnmarshalOutput(sends[0].Payload)
			403	+	require.NoError(t, err)
			404	+	require.Zero(t, output.BaseStateNum, "resume must start with a complete frame")
			405	+	terminal := vt.NewScreen(resumedAC.size.Cols, resumedAC.size.Rows)
			406	+	terminal.Write(output.Data)
			407	+	contents := strings.Join(frameRows(terminal.Frame), "\n")
			408	+	require.Contains(t, contents, "survivor", "resume first paint must contain current headless content")
			409	+	require.NotContains(t, contents, "closed", "resume first paint must not contain closed pane content")
			410	+	}
			411	+
250		func TestExplicitDetachDoesNotPark(t *testing.T) {	412		func TestExplicitDetachDoesNotPark(t *testing.T) {
251		pty, release := newBlockingPTY(t)	413		pty, release := newBlockingPTY(t)
252		defer release()	414		defer release()

A internal/usecase/daemon/resume_wake_incarnation_test.go +68 -0

@@ -0,0 +1,68 @@
	1	+	package daemon
	2	+
	3	+	import (
	4	+	"testing"
	5	+
	6	+	"github.com/stretchr/testify/require"
	7	+
	8	+	"github.com/bnema/vev/internal/domain"
	9	+	"github.com/bnema/vev/internal/ports"
	10	+	)
	11	+
	12	+	// A wake can have passed coordinator scheduling before a link loss parks its
	13	+	// attachment. Resume deliberately reuses that attachment object, so pointer
	14	+	// identity alone must not let the old wake paint the new transport before its
	15	+	// Welcome frame.
	16	+	func TestParkedResumeRejectsDispatchedWakeFromPriorAttachment(t *testing.T) {
	17	+	pty, releasePTY := newBlockingPTY(t)
	18	+	defer releasePTY()
	19	+	d := newTestDaemon(t, newFactorySeq(t, pty), stubClock{})
	20	+
	21	+	oldTransport := &closeTrackingTransport{}
	22	+	sess, ac, err := d.route(helloResumeCapable(ports.IntentNew, "work", 0), oldTransport)
	23	+	require.NoError(t, err)
	24	+	token := ac.resumeToken
	25	+	rc := sess.renderCoordinator()
	26	+	require.NotNil(t, rc)
	27	+	require.True(t, rc.markAttachmentReady(rc.attachmentLease(ac)))
	28	+
	29	+	// Pause after dispatch but before the coordinator's normal wake callback
	30	+	// reaches paint/sendMu. This is the precise stale-wake window.
	31	+	originalWake := rc.opts.wake
	32	+	wakeDispatched := make(chan struct{})
	33	+	releaseWake := make(chan struct{})
	34	+	rc.opts.wake = func(w renderWake) {
	35	+	close(wakeDispatched)
	36	+	<-releaseWake
	37	+	originalWake(w)
	38	+	}
	39	+	wakeDone := make(chan struct{})
	40	+	go func() {
	41	+	d.invalidateRenderNow(sess, ac, true, "resume-wake-test")
	42	+	close(wakeDone)
	43	+	}()
	44	+	<-wakeDispatched
	45	+
	46	+	d.clientGone(sess, ac, oldTransport, false)
	47	+	newTransport := &closeTrackingTransport{}
	48	+	resumedSess, resumedAC, ok, err := d.resumeParked(helloResumeCapable(ports.IntentResume, sess.name, token), newTransport, domain.Size{Cols: 80, Rows: 24})
	49	+	require.NoError(t, err)
	50	+	require.True(t, ok)
	51	+	require.Same(t, sess, resumedSess)
	52	+	require.Same(t, ac, resumedAC)
	53	+
	54	+	close(releaseWake)
	55	+	<-wakeDone
	56	+	require.Empty(t, newTransport.Sends(), "a stale pre-resume wake must not emit Output before Welcome")
	57	+	ac.sendMu.Lock()
	58	+	staleWakeComposed := ac.pipelineCache.valid
	59	+	ac.sendMu.Unlock()
	60	+	require.False(t, staleWakeComposed, "a stale wake must not mutate the resumed attachment shadow")
	61	+
	62	+	// The new attachment incarnation accepts the required paint only after its
	63	+	// replacement transport has completed Welcome.
	64	+	rc.opts.wake = originalWake
	65	+	require.True(t, rc.markAttachmentReady(rc.attachmentLease(ac)))
	66	+	d.firstPaint(sess, ac, ac.size)
	67	+	require.NotEmpty(t, newTransport.Sends())
	68	+	}

M internal/usecase/daemon/scrollback_test.go +15 -6

 		wantRow string
 	}{
 		{
-			name:    "tab owns scrollback wired to vt eviction callback",
+			name:    "primary evictions enter screen owned history",
 			write:   []byte("AAAA\r\nBBBB\r\nCCCC\r\n"),
 			wantLen: 1,
 			wantRow: "AAAA",
 		},
+		{
+			name:    "alternate screen evictions do not enter pane history",
+			write:   []byte("AAAA\r\nBBBB\r\nCCCC\r\n\x1b[?1049h1111\r\n2222\r\n3333\r\n4444\x1b[?1049l"),
+			wantLen: 1,
+			wantRow: "AAAA",
+		},
 	}
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
 			win := newTab(nil, domain.Size{Cols: 4, Rows: 3})
-			if win.focusedPane().scrollback == nil {
+			if win.focusedPane().history == nil {
 				t.Fatal("scrollback is nil")
 			}
-			if win.focusedPane().scrollback.Cap() != defaultScrollbackRows {
+			if win.focusedPane().history != win.focusedPane().screen.History() {
-				t.Fatalf("scrollback cap = %d, want %d", win.focusedPane().scrollback.Cap(), defaultScrollbackRows)
+				t.Fatal("pane history is not owned by its screen")
+			}
+			if win.focusedPane().history.Cap() != defaultScrollbackRows {
+				t.Fatalf("scrollback cap = %d, want %d", win.focusedPane().history.Cap(), defaultScrollbackRows)
 			}
 			win.focusedPane().screen.Write(tt.write)
-			if got := win.focusedPane().scrollback.Len(); got != tt.wantLen {
+			if got := win.focusedPane().history.Len(); got != tt.wantLen {
 				t.Fatalf("scrollback len = %d, want %d", got, tt.wantLen)
 			}
-			row := win.focusedPane().scrollback.View().Row(0)
+			row := win.focusedPane().history.View().Row(0)
 			runes := make([]rune, len(row))
 			for i, c := range row {
 				runes[i] = c.Rune

M internal/usecase/daemon/session.go +68 -40

@@ -1,27 +1,4 @@
1	-	// Package daemon holds vev's server-side session multiplexer use case: the	1	+	// Package daemon holds vev's server-side session multiplexer use case.
2	-	// accept loop, the ephemeral/named session registry, the per-tab PTY reader
3	-	// and VT screen, and the per-client debounced render scheduler.
4	-	//
5	-	// Concurrency model (sessions own one or more PTY-backed tabs):
6	-	//
7	-	// - Serve runs the accept loop. Each accepted connection is handled by its
8	-	// own goroutine (handleConn): it reads the first frame and routes it to a
9	-	// session create/attach, a list, or a kill.
10	-	// - Per session there are exactly two long-lived goroutines: the PTY reader
11	-	// (drains child output into the VT screen and pokes a cap-1 dirty channel)
12	-	// and the render scheduler (debounces dirties and paints the attached
13	-	// client). Both are tied to the session context and unwind when the
14	-	// session is killed (pty.Close unblocks the reader; ctx cancel stops the
15	-	// scheduler).
16	-	// - The daemon exits (Serve returns) when the last session is removed, or
17	-	// when the parent context is cancelled (graceful shutdown notifies any
18	-	// attached clients with ReasonServerShutdown).
19	-	//
20	-	// Locking: a pane's screen/scrollback and per-client renderer shadow are
21	-	// guarded by pane.mu/tab.mu as appropriate; the attached-client pointer by
22	-	// session.mu; the registry by Daemon.mu. When more than one is held the order
23	-	// is always attachedClient.sendMu > Daemon.mu > session.mu > tab.mu > pane.mu.
24	-	// The PTY reader only ever takes pane.mu, so it never blocks on a slow client.
25		package daemon	2		package daemon
26			3
27		import (	4		import (

@@ -67,15 +44,35 @@ type session struct {
67		mruAt atomic.Uint64	44		mruAt atomic.Uint64
68		snapDirty atomic.Bool	45		snapDirty atomic.Bool
69		snapEligible atomic.Bool	46		snapEligible atomic.Bool
			47	+	// snapshotMu serializes the dirty generation with worker completion. It is
			48	+	// intentionally independent from mu: persistence never holds session state
			49	+	// locks while encoding or writing.
			50	+	snapshotMu sync.Mutex
			51	+	snapshotGeneration uint64
			52	+	snapshotPending bool
			53	+	snapshotPendingCaptures uint
			54	+	snapshotChanged chan struct{}
			55	+	// syncGen makes synchronized-output watchdog generations unique across all
			56	+	// panes in this session.
			57	+	syncGen atomic.Uint64
			58	+	// coordinator fans in this session's producer render invalidations.
			59	+	coordinator atomic.Pointer[renderCoordinator]
70		// clipFiles records clipboard-image-transfer temp file paths (see	60		// clipFiles records clipboard-image-transfer temp file paths (see
71		// clipboard.go) written for this session, removed best-effort in	61		// clipboard.go) written for this session, removed best-effort in
72		// killSession.	62		// killSession.
73		clipFiles []string	63		clipFiles []string
			64	+
			65	+	// floatingLaunchMu is ownership synchronization for external floating
			66	+	// launches. It is separate from mu so PTYFactory.Open never holds a
			67	+	// session, tab, or daemon architecture lock.
			68	+	floatingLaunchMu sync.Mutex
			69	+	floatingLaunchStopping bool
			70	+	floatingLaunches map[*floatingLaunch]struct{}
74		}	71		}
75			72
76		// tab is a pane layout container; pane owns PTY/screen/scrollback/render scheduling state.	73		// tab is a pane layout container; pane owns PTY/screen/scrollback/render scheduling state.
77		type tab struct {	74		type tab struct {
78	-	mu sync.Mutex // guards tree, panes, floating, nextPaneID, size, previewClient, and pane map membership	75	+	mu sync.Mutex // guards tree, panes, floating, nextPaneID, size, and pane map membership
79			76
80		stableID string	77		stableID string
81		name string	78		name string

@@ -88,9 +85,6 @@ type tab struct {
88		// floating is independent from the normal layout tree and pane map.	85		// floating is independent from the normal layout tree and pane map.
89		floating floatingSlot	86		floating floatingSlot
90			87
91	-	// previewClient tracks the one client currently previewing this tab in the picker.
92	-	// v1 is last-writer-wins: multiple clients previewing the same tab are not supported.
93	-	previewClient *attachedClient
94		// attention fields are guarded by the owning session.mu.	88		// attention fields are guarded by the owning session.mu.
95		attention bool	89		attention bool
96		attentionAt time.Time	90		attentionAt time.Time

@@ -100,7 +94,7 @@ type tab struct {
100		// attachedClient is a client currently attached to a session's tab. rend is	94		// attachedClient is a client currently attached to a session's tab. rend is
101		// its private renderer shadow (so each client gets output minimised against	95		// its private renderer shadow (so each client gets output minimised against
102		// what it has actually seen). sendMu serialises the two senders — the render	96		// what it has actually seen). sendMu serialises the two senders — the render
103	-	// scheduler and the connection handler — so the transport's single-writer	97	+	// coordinator and the connection handler — so the transport's single-writer
104			98
105		func (d Daemon) touchMRU(sess session) {	99		func (d Daemon) touchMRU(sess session) {
106		if d == nil \|\| sess == nil {	100		if d == nil \|\| sess == nil {

@@ -145,7 +139,7 @@ func (d *Daemon) createSessionLocked(name string, ephemeral bool, cwd string, sz
145		closeTabs(tabs)	139		closeTabs(tabs)
146		return nil, err	140		return nil, err
147		}	141		}
148	-	pty, err := d.ptys.Open(d.shell, d.shellArgs, d.childEnv(name, tabStableID, paneStableID, term), cwd, tbSize)	142	+	pty, err := d.ptys.Open(d.serveCtx, d.shell, d.shellArgs, d.childEnv(name, tabStableID, paneStableID, term), cwd, tbSize)
149		if err != nil {	143		if err != nil {
150		closeTabs(tabs)	144		closeTabs(tabs)
151		d.log.Warn("pty spawn failed", "err", err, "session", name, "kind", "session")	145		d.log.Warn("pty spawn failed", "err", err, "session", name, "kind", "session")

@@ -258,16 +252,19 @@ func (d Daemon) createSessionAndSwitch(from session, ac *attachedClient, name
258		ac.setSession(nil)	252		ac.setSession(nil)
259		from.mu.Unlock()	253		from.mu.Unlock()
260			254
			255	+	// Prepare source invalidation, output rebase, and destination coordinator
			256	+	// identity before publishing the destination attachment.
			257	+	cleanups := d.handoffCoordinator(from, newSess, nil, ac)
			258	+	ac.setSession(newSess)
261		newSess.mu.Lock()	259		newSess.mu.Lock()
262		newSess.client = ac	260		newSess.client = ac
263		newSess.mu.Unlock()	261		newSess.mu.Unlock()
264			262
265		d.touchMRU(newSess)	263		d.touchMRU(newSess)
266	-	ac.setSession(newSess)
267		ac.recordPreviousSession(from)	264		ac.recordPreviousSession(from)
268		d.log.Info("client attached", "session", newSess.name, "resume", ac.resumeCapable)	265		d.log.Info("client attached", "session", newSess.name, "resume", ac.resumeCapable)
269		d.mu.Unlock()	266		d.mu.Unlock()
270	-		267	+	finishRenderLifecycleCleanups(cleanups)
271		d.firstPaint(newSess, ac, sz)	268		d.firstPaint(newSess, ac, sz)
272		return nil	269		return nil
273		}	270		}

@@ -284,7 +281,7 @@ func (d Daemon) createTab(sess session, sz domain.Size) error {
284		if err != nil {	281		if err != nil {
285		return err	282		return err
286		}	283		}
287	-	pty, err := d.ptys.Open(d.shell, d.shellArgs, d.childEnv(name, tabStableID, paneStableID, term), cwd, tbSize)	284	+	pty, err := d.ptys.Open(sess.ctx, d.shell, d.shellArgs, d.childEnv(name, tabStableID, paneStableID, term), cwd, tbSize)
288		if err != nil {	285		if err != nil {
289		d.log.Warn("pty spawn failed", "err", err, "session", name, "kind", "tab")	286		d.log.Warn("pty spawn failed", "err", err, "session", name, "kind", "tab")
290		return fmt.Errorf("daemon: spawning tab for session %q: %w", name, err)	287		return fmt.Errorf("daemon: spawning tab for session %q: %w", name, err)

@@ -418,9 +415,9 @@ func (d Daemon) startPaneGoroutines(sess session, tb tab, p pane) {
418		sess.mu.Unlock()	415		sess.mu.Unlock()
419		d.log.Info("pane created", "session", name, "pane", p.id)	416		d.log.Info("pane created", "session", name, "pane", p.id)
420		}	417		}
421	-	d.sessWg.Add(2)	418	+	// Scheduler ownership was removed; this launch creates exactly one reader.
			419	+	d.sessWg.Add(1)
422		go d.ptyReader(sess, tb, p)	420		go d.ptyReader(sess, tb, p)
423	-	go d.scheduler(sess, tb, p)
424		}	421		}
425			422
426		// attachClient makes ac the session's current client, displacing any prior one	423		// attachClient makes ac the session's current client, displacing any prior one

@@ -637,16 +634,27 @@ func (d Daemon) closeTab(sess session, tb *tab, repaint bool) {
637		markSnapshotDirty(sess)	634		markSnapshotDirty(sess)
638			635
639		d.clearDestroyedTabPreview(tb)	636		d.clearDestroyedTabPreview(tb)
			637	+	tb.mu.Lock()
			638	+	panes := tb.panesSnapshot()
			639	+	tb.mu.Unlock()
			640	+	if ac != nil {
			641	+	ac.pruneCaptureFrames(panes...)
			642	+	}
640		if tb.cancel != nil {	643		if tb.cancel != nil {
641		tb.cancel()	644		tb.cancel()
642		}	645		}
643		d.teardownFloating(tb, ac)	646		d.teardownFloating(tb, ac)
			647	+	if rc := sess.renderCoordinator(); rc != nil {
			648	+	for _, p := range panes {
			649	+	rc.noteSyncPaneRemoved(p)
			650	+	}
			651	+	}
644		tb.closeAllPanes()	652		tb.closeAllPanes()
645		if wasActive {	653		if wasActive {
646		d.activateTab(sess, destination)	654		d.activateTab(sess, destination)
647		}	655		}
648		if repaint && ac != nil {	656		if repaint && ac != nil {
649	-	d.paint(sess, ac, true)	657	+	d.invalidateRender(sess, ac, true, "session.go")
650		}	658		}
651		if ringing {	659		if ringing {
652		d.repaintAllAttachedClients()	660		d.repaintAllAttachedClients()

@@ -664,7 +672,7 @@ func (d Daemon) killSession(sess session, reason uint8, purge bool) error {
664		if !purge && !isEphemeral && d.persistEnabled {	672		if !purge && !isEphemeral && d.persistEnabled {
665		d.refreshSessionCwd(sess)	673		d.refreshSessionCwd(sess)
666		}	674		}
667	-	var snapshotDeleteErr error	675	+	var snapshotDeleteErr, terminalSnapshotErr error
668		if d.snapsEnabled && !isEphemeral {	676		if d.snapsEnabled && !isEphemeral {
669		if purge {	677		if purge {
670		sess.mu.Lock()	678		sess.mu.Lock()

@@ -675,7 +683,16 @@ func (d Daemon) killSession(sess session, reason uint8, purge bool) error {
675		d.log.Warn("deleting session snapshot failed", "err", err, "session", name)	683		d.log.Warn("deleting session snapshot failed", "err", err, "session", name)
676		}	684		}
677		} else {	685		} else {
678	-	d.captureSession(sess)	686	+	markSnapshotDirty(sess)
			687	+	if !d.scheduleFinalSnapshot(sess) {
			688	+	sess.mu.Lock()
			689	+	name := sess.name
			690	+	sess.mu.Unlock()
			691	+	terminalSnapshotErr = fmt.Errorf("retain final snapshot for session %q: snapshot worker unavailable or saturated", name)
			692	+	if reason != ports.ReasonServerShutdown {
			693	+	return terminalSnapshotErr
			694	+	}
			695	+	}
679		}	696		}
680		}	697		}
681		d.mu.Lock()	698		d.mu.Lock()

@@ -731,13 +748,24 @@ func (d Daemon) killSession(sess session, reason uint8, purge bool) error {
731		ac := sess.client	748		ac := sess.client
732		sess.client = nil	749		sess.client = nil
733		sess.mu.Unlock()	750		sess.mu.Unlock()
			751	+	if rc := sess.renderCoordinator(); rc != nil {
			752	+	// Terminal teardown has two phases: this session owner first prevents any
			753	+	// new worker registration and detaches all tokens, then stops and waits
			754	+	// outside coordinator callbacks and session locks.
			755	+	rc.beginSessionTeardown().finish()
			756	+	rc.waitForTimerWorkers()
			757	+	}
734		if ac != nil {	758		if ac != nil {
735	-	ac.cancelResizePaint()
736		d.unregisterPreview(ac)	759		d.unregisterPreview(ac)
737		ac.clearPreviousSession()	760		ac.clearPreviousSession()
738		ac.setSession(nil)	761		ac.setSession(nil)
			762	+	ac.clearCaptureFrames()
739		}	763		}
740			764
			765	+	// Prevent queued launches from entering Open, then cancel the parent
			766	+	// context. The launch worker owns any late PTY result and closes it rather
			767	+	// than publishing it into the torn-down tab.
			768	+	sess.stopFloatingLaunches()
741		sess.cancel()	769		sess.cancel()
742		sess.mu.Lock()	770		sess.mu.Lock()
743		tabs := append([]*tab(nil), sess.tabs...)	771		tabs := append([]*tab(nil), sess.tabs...)

@@ -763,7 +791,7 @@ func (d Daemon) killSession(sess session, reason uint8, purge bool) error {
763		if ac != nil {	791		if ac != nil {
764		d.notifyDetachedAsync(ac, reason)	792		d.notifyDetachedAsync(ac, reason)
765		}	793		}
766	-	return purgeErr	794	+	return errors.Join(purgeErr, terminalSnapshotErr)
767		}	795		}
768			796
769		// allocEphemeralNameLocked returns the lowest free decimal name. Caller holds	797		// allocEphemeralNameLocked returns the lowest free decimal name. Caller holds

M internal/usecase/daemon/session_back.go +1 -1


@@ -11,7 +11,7 @@ func (d Daemon) backSession(current session, ac *attachedClient) {
11		target := ac.previousSession.Get()	11		target := ac.previousSession.Get()
12		if target == nil \|\| target == current \|\| d.sessionByID(target.id) != target {	12		if target == nil \|\| target == current \|\| d.sessionByID(target.id) != target {
13		ac.clearPreviousSessionIf(target)	13		ac.clearPreviousSessionIf(target)
14	-	d.paint(current, ac, true)	14	+	d.invalidateRender(current, ac, true, "session_back.go")
15		return	15		return
16		}	16		}
17		d.switchToTarget(current, ac, picker.Target{Session: target.id, TabIndex: -1})	17		d.switchToTarget(current, ac, picker.Target{Session: target.id, TabIndex: -1})

M internal/usecase/daemon/session_test.go +104 -51

 )
 // --- test doubles -----------------------------------------------------------
+func expectFloatingPrewarmOpen(factory *portsmocks.MockPTYFactory, normalSize domain.Size, floating ports.PTY) {
+	factory.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.MatchedBy(func(got domain.Size) bool {
+		return got != normalSize && got.Valid()
+	})).Return(floating, nil).Maybe()
+}
 // stubClock returns timers whose channel never fires, so a scheduler under it
 // blocks in its debounce loop until the session context is cancelled. Used by
 	f := portsmocks.NewMockPTYFactory(t)
 	var dirs []string
 	normalSize := domain.Size{Cols: sz.Cols, Rows: sz.Rows - 2}
-	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, normalSize).RunAndReturn(
+	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything, normalSize).RunAndReturn(
-		func(_ string, _ []string, _ []string, dir string, _ domain.Size) (ports.PTY, error) {
+		func(_ context.Context, _ string, _ []string, _ []string, dir string, _ domain.Size) (ports.PTY, error) {
 			dirs = append(dirs, dir)
 			if len(dirs) == 1 {
 				return first, nil
 		},
 	).Twice()
 	floating := newQuietPTY()
-	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.MatchedBy(func(got domain.Size) bool {
+	expectFloatingPrewarmOpen(f, normalSize, floating)
-		return got != normalSize && got.Valid()
-	})).Return(floating, nil).Once()
 	d := newTestDaemon(t, f, stubClock{})
 	tr := portsmocks.NewMockTransport(t)
 	releaseFirst()
 	releaseSecond()
 	d.sessWg.Wait()
-	select {
+	// Teardown may cancel a queued prewarm before it enters Open; an opened
-	case <-floating.done:
+	// prewarm is closed by teardownFloating.
-	default:
-		t.Fatal("floating prewarm PTY was not closed")
-	}
 }
 func TestHandshakeEphemeralHappy(t *testing.T) {
 	p1, release1 := newBlockingPTY(t)
 	defer release1()
 	p2 := portsmocks.NewMockPTY(t)
-	opened := make(chan struct{})
+	openCtx := make(chan context.Context, 1)
 	releaseOpen := make(chan struct{})
 	closed := make(chan struct{})
 	p2.EXPECT().Close().RunAndReturn(func() error {
 	f := portsmocks.NewMockPTYFactory(t)
 	normalSize := domain.Size{Cols: 80, Rows: 22}
-	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, normalSize).Return(p1, nil).Once()
+	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything, normalSize).Return(p1, nil).Once()
 	floating := newQuietPTY()
-	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.MatchedBy(func(got domain.Size) bool {
+	expectFloatingPrewarmOpen(f, normalSize, floating)
-		return got != normalSize && got.Valid()
+	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything, normalSize).RunAndReturn(
-	})).Return(floating, nil).Once()
+		func(ctx context.Context, _ string, _ []string, _ []string, _ string, _ domain.Size) (ports.PTY, error) {
-	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, normalSize).RunAndReturn(
+			openCtx <- ctx
-		func(string, []string, []string, string, domain.Size) (ports.PTY, error) {
+			select {
-			close(opened)
+			case <-ctx.Done():
-			<-releaseOpen
+			case <-releaseOpen:
+			}
 			return p2, nil
 		},
 	).Once()
 	errCh := make(chan error, 1)
 	go func() { errCh <- d.createTab(sess, domain.Size{Cols: 80, Rows: 24}) }()
-	<-opened
+	ctx := <-openCtx
 	_ = d.killSession(sess, ports.ReasonSessionKilled, false)
+	cancelled := true
+	select {
+	case <-ctx.Done():
+	case <-time.After(time.Second):
+		cancelled = false
+	}
 	close(releaseOpen)
+	if !cancelled {
+		t.Error("killSession did not cancel the context passed to PTYFactory.Open")
+	}
 	require.Error(t, <-errCh)
 	select {
 	case <-closed:
 	releaseConn()
 	hg.Wait()
 	d.sessWg.Wait()
-	select {
+	// A queued prewarm may be cancelled before Open during teardown.
-	case <-floating.done:
-	default:
-		t.Fatal("floating prewarm PTY was not closed")
-	}
 }
 // --- ephemeral numbering ----------------------------------------------------
 // --- shutdown/create interlock ------------------------------------------------
 // TestHelloRacingShutdownIsRejected covers the interlock between the
-// registry-empty shutdown decision and session creation: once killSession has
+// registry-empty shutdown decision and session creation. In particular, the
-// removed the last session (making shutdown irreversible via doneOnce), a
+// old `.Once()` factory expectation was wrong: firstPaint legitimately queues
-// Hello landing at any later instant must be rejected cleanly — never spawn a
+// a floating prewarm, so an Open that starts before killSession is valid work,
-// child that no teardown pass would reap. The interleaving is deterministic:
+// not a leaked post-shutdown child. The factory markers classify timing at the
-// killSession runs to completion before the racing Hellos are handled, which
+// Open boundary: this test first observes that valid prewarm, then verifies
-// models the worst case (insertion attempt after the shutdown snapshot).
+// that d.done (closed synchronously by killSession) is never followed by an
+// Open start.
 func TestHelloRacingShutdownIsRejected(t *testing.T) {
 	p, releasePTY := newBlockingPTY(t)
+	floating := newQuietPTY()
 	f := portsmocks.NewMockPTYFactory(t)
-	// Exactly one child may ever be spawned; a second Open call (a leaked
+	normalSize := domain.Size{Cols: 80, Rows: 22}
-	// child) fails the test via mock expectations.
+	preShutdownFloatingOpen := make(chan struct{}, 1)
-	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything).Return(p, nil).Once()
+	var opensAfterShutdown atomic.Int32
 	d := newTestDaemon(t, f, stubClock{})
+	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything).RunAndReturn(
+		func(_ context.Context, _ string, _ []string, _ []string, _ string, size domain.Size) (ports.PTY, error) {
+			select {
+			case <-d.done:
+				opensAfterShutdown.Add(1)
+			default:
+			}
+			if size == normalSize {
+				return p, nil
+			}
+			select {
+			case preShutdownFloatingOpen <- struct{}{}:
+			default:
+			}
+			return floating, nil
+		},
+	).Maybe()
 	tr1, sends1, release1 := newConn(t, mustHello(ports.IntentEphemeral, "", domain.Size{Cols: 80, Rows: 24}))
 	var hg sync.WaitGroup
 	hg.Go(func() { d.handleConn(tr1) })
 	awaitFrame(t, sends1, ports.MsgWelcome)
+	// firstPaint starts the background prewarm asynchronously. It is valid for
+	// this Open to occur before shutdown, so do not mistake it for the racing
+	// Hello's forbidden launch.
+	awaitFrame(t, sends1, ports.MsgOutput)
+	select {
+	case <-preShutdownFloatingOpen:
+	case <-time.After(time.Second):
+		t.Fatal("legitimate pre-shutdown floating prewarm did not start")
+	}
 	sess := firstSession(d)
 	require.NotNil(t, sess)
 		require.Equal(t, ports.ErrServerShutdown, em.Code, "%s hello racing shutdown must be rejected", intent.name)
 	}
 	require.Equal(t, 0, sessionCount(d), "no session may be inserted after shutdown began")
+	require.Zero(t, opensAfterShutdown.Load(), "no PTYFactory.Open may start after d.done/shutdown completion")
 	release1()
 	releasePTY()
 	hg.Wait()
 	d.sessWg.Wait()
 	d.waitNotifies()
+	require.Zero(t, opensAfterShutdown.Load(), "racing Hellos must not start a PTY after shutdown")
 }
 // --- wedged-client teardown ---------------------------------------------------
 	var gotEnv []string
 	f := portsmocks.NewMockPTYFactory(t)
-	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything).RunAndReturn(
+	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything).RunAndReturn(
-		func(_ string, _ []string, env []string, _ string, _ domain.Size) (ports.PTY, error) {
+		func(_ context.Context, _ string, _ []string, env []string, _ string, _ domain.Size) (ports.PTY, error) {
 			gotEnv = append([]string(nil), env...)
 			return p, nil
 		},
 	var opens [][]string
 	f := portsmocks.NewMockPTYFactory(t)
 	normalSize := domain.Size{Cols: sz.Cols, Rows: sz.Rows - 2}
-	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, normalSize).RunAndReturn(
+	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything, normalSize).RunAndReturn(
-		func(_ string, _ []string, env []string, _ string, _ domain.Size) (ports.PTY, error) {
+		func(_ context.Context, _ string, _ []string, env []string, _ string, _ domain.Size) (ports.PTY, error) {
 			opens = append(opens, append([]string(nil), env...))
 			if len(opens) == 1 {
 				return p1, nil
 		},
 	).Twice()
 	floating := newQuietPTY()
-	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.MatchedBy(func(got domain.Size) bool {
+	expectFloatingPrewarmOpen(f, normalSize, floating)
-		return got != normalSize && got.Valid()
-	})).Return(floating, nil).Once()
 	d := newTestDaemon(t, f, stubClock{})
 	d.baseEnv = []string{"KEEP=1", "COLORTERM=old"}
 	var opens [][]string
 	f := portsmocks.NewMockPTYFactory(t)
 	normalSize := domain.Size{Cols: sz.Cols, Rows: sz.Rows - 2}
-	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, normalSize).RunAndReturn(
+	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything, normalSize).RunAndReturn(
-		func(_ string, _ []string, env []string, _ string, _ domain.Size) (ports.PTY, error) {
+		func(_ context.Context, _ string, _ []string, env []string, _ string, _ domain.Size) (ports.PTY, error) {
 			opens = append(opens, append([]string(nil), env...))
 			if len(opens) == 1 {
 				return p1, nil
 		},
 	).Twice()
 	floating := newQuietPTY()
-	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.MatchedBy(func(got domain.Size) bool {
+	expectFloatingPrewarmOpen(f, normalSize, floating)
-		return got != normalSize && got.Valid()
-	})).Return(floating, nil).Once()
 	d := newTestDaemon(t, f, stubClock{})
 	tr1 := portsmocks.NewMockTransport(t)
 	var opens [][]string
 	f := portsmocks.NewMockPTYFactory(t)
 	normalSize := domain.Size{Cols: sz.Cols, Rows: sz.Rows - 2}
-	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, normalSize).RunAndReturn(
+	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.Anything, normalSize).RunAndReturn(
-		func(_ string, _ []string, env []string, _ string, _ domain.Size) (ports.PTY, error) {
+		func(_ context.Context, _ string, _ []string, env []string, _ string, _ domain.Size) (ports.PTY, error) {
 			opens = append(opens, append([]string(nil), env...))
 			if len(opens) == 1 {
 				return p1, nil
 		},
 	).Twice()
 	floating := newQuietPTY()
-	f.EXPECT().Open(mock.Anything, mock.Anything, mock.Anything, mock.Anything, mock.MatchedBy(func(got domain.Size) bool {
+	expectFloatingPrewarmOpen(f, normalSize, floating)
-		return got != normalSize && got.Valid()
-	})).Return(floating, nil).Once()
 	d := newTestDaemon(t, f, stubClock{})
 	tr := portsmocks.NewMockTransport(t)
 	tr.EXPECT().Send(mock.Anything).Return(nil).Maybe()
 	tr.EXPECT().Close().Return(nil).Maybe()
 	sess, ac, err := d.route(ports.Hello{Version: ports.ProtocolVersion, Intent: ports.IntentNew, Name: "work", Size: sz, TrueColor: true}, tr)
 	require.NoError(t, err)
+	// The source coordinator is deliberately made pending so the switch must
+	// invalidate it rather than letting its stale callback render for ac.
+	sourceCoordinator := sess.renderCoordinator()
+	require.NotNil(t, sourceCoordinator)
+	sourceCoordinator.invalidate(renderInvalidation{class: invalidateOutput})
+	ac.sendMu.Lock()
+	ac.output.next = 3
+	ac.output.acked = 1
+	ac.sendMu.Unlock()
 	require.NoError(t, d.createSessionAndSwitch(sess, ac, "next"))
 	got := ac.sess.Get()
 	require.NotNil(t, got)
+	sourceCoordinator.mu.Lock()
+	require.NotNil(t, sourceCoordinator.lease)
+	require.False(t, sourceCoordinator.lease.active, "source callbacks must be stale after handoff")
+	sourceCoordinator.mu.Unlock()
+	destinationCoordinator := got.renderCoordinator()
+	require.NotNil(t, destinationCoordinator, "destination coordinator precedes first paint")
+	destinationCoordinator.mu.Lock()
+	require.Same(t, ac, destinationCoordinator.lease.attachment)
+	destinationCoordinator.mu.Unlock()
+	ac.sendMu.Lock()
+	require.Equal(t, uint64(1), ac.output.next-ac.output.acked, "only the destination first paint may follow the rebase")
+	ac.sendMu.Unlock()
 	got.mu.Lock()
 	require.True(t, got.terminal.TrueColor)
 	got.mu.Unlock()

M internal/usecase/daemon/snapshot.go +455 -79


@@ -4,17 +4,58 @@ import (
4		"context"	4		"context"
5		"fmt"	5		"fmt"
6		"math"	6		"math"
			7	+	"sync"
7		"time"	8		"time"
8			9
9		"github.com/bnema/vev/internal/domain"	10		"github.com/bnema/vev/internal/domain"
10	-	scopy "github.com/bnema/vev/internal/usecase/copy"
11		"github.com/bnema/vev/internal/usecase/layout"	11		"github.com/bnema/vev/internal/usecase/layout"
12		snapcodec "github.com/bnema/vev/internal/usecase/snapshot"	12		snapcodec "github.com/bnema/vev/internal/usecase/snapshot"
13		"github.com/bnema/vev/pkg/renderer"	13		"github.com/bnema/vev/pkg/renderer"
14		"github.com/bnema/vev/pkg/vt"	14		"github.com/bnema/vev/pkg/vt"
15		)	15		)
16			16
17	-	const snapshotInterval = 30 * time.Second	17	+	// snapshotQueueCapacity bounds retained immutable captures. A full queue never
			18	+	// stalls session producers; the session remains dirty for the next interval.
			19	+	const snapshotQueueCapacity = 1
			20	+
			21	+	// snapshotFinalQueueCapacity is a global bound for terminal captures retained
			22	+	// while the worker is blocked. On saturation a new terminal capture is rejected
			23	+	// (and left dirty) rather than silently dropping a persistence acknowledgement.
			24	+	const snapshotFinalQueueCapacity = 32
			25	+
			26	+	type snapshotCapture struct {
			27	+	session *session
			28	+	generation uint64
			29	+	name string
			30	+	createdAt uint64
			31	+	active uint16
			32	+	tabs []snapshotCaptureTab
			33	+	finishOnce sync.Once
			34	+	}
			35	+
			36	+	type snapshotCaptureTab struct {
			37	+	stableID string
			38	+	cols uint16
			39	+	rows uint16
			40	+	nextPaneID uint64
			41	+	focus layout.PaneID
			42	+	tree *layout.Tree
			43	+	panes []snapshotCapturePane
			44	+	}
			45	+
			46	+	type snapshotCapturePane struct {
			47	+	id layout.PaneID
			48	+	stableID string
			49	+	cwd string
			50	+	history vt.HistoryView
			51	+	visible renderer.Frame
			52	+	process *snapcodec.Process
			53	+	}
			54	+
			55	+	const (
			56	+	snapshotInterval = 30 * time.Second
			57	+	snapshotFinalFlushTimeout = time.Second
			58	+	)
18			59
19		func (d *Daemon) closeRestoreDone() {	60		func (d *Daemon) closeRestoreDone() {
20		d.restoreOnce.Do(func() { close(d.restoreDone) })	61		d.restoreOnce.Do(func() { close(d.restoreDone) })

@@ -51,6 +92,13 @@ func (d *Daemon) restoreSession(ctx context.Context, snap snapcodec.Session) err
51		if err := ctx.Err(); err != nil {	92		if err := ctx.Err(); err != nil {
52		return err	93		return err
53		}	94		}
			95	+	if len(snap.Tabs) == 0 {
			96	+	if snap.Active != 0 {
			97	+	return fmt.Errorf("snapshot: active tab out of range")
			98	+	}
			99	+	} else if int(snap.Active) >= len(snap.Tabs) {
			100	+	return fmt.Errorf("snapshot: active tab out of range")
			101	+	}
54		if snap.Name == "" {	102		if snap.Name == "" {
55		return fmt.Errorf("snapshot: empty session name")	103		return fmt.Errorf("snapshot: empty session name")
56		}	104		}

@@ -141,7 +189,7 @@ func (d *Daemon) restoreSession(ctx context.Context, snap snapcodec.Session) err
141		return fmt.Errorf("snapshot: generating pane identity: %w", err)	189		return fmt.Errorf("snapshot: generating pane identity: %w", err)
142		}	190		}
143		}	191		}
144	-	pty, err := d.ptys.Open(d.shell, d.shellArgs, d.childEnv(snap.Name, tabStableID, paneStableID, restoreTerm), paneSnap.Cwd, contentSize)	192	+	pty, err := d.ptys.Open(ctx, d.shell, d.shellArgs, d.childEnv(snap.Name, tabStableID, paneStableID, restoreTerm), paneSnap.Cwd, contentSize)
145		if err != nil {	193		if err != nil {
146		closeOpened()	194		closeOpened()
147		return err	195		return err

@@ -149,7 +197,13 @@ func (d *Daemon) restoreSession(ctx context.Context, snap snapcodec.Session) err
149		p := newPaneWithStableID(paneSnap.ID, paneStableID, pty, contentSize)	197		p := newPaneWithStableID(paneSnap.ID, paneStableID, pty, contentSize)
150		p.ctx, p.cancel = context.WithCancel(tb.ctx)	198		p.ctx, p.cancel = context.WithCancel(tb.ctx)
151		p.rect = contentRect	199		p.rect = contentRect
152	-	seedPaneRows(p, paneSnap.Scrollback, paneSnap.Visible)	200	+	if err := restorePaneTerminal(p, paneSnap); err != nil {
			201	+	// p is not in tb.panes yet, so closeOpened cannot reach it.
			202	+	p.cancel()
			203	+	_ = pty.Close()
			204	+	closeOpened()
			205	+	return err
			206	+	}
153		if restoreCommand != "" {	207		if restoreCommand != "" {
154		if _, err := pty.Write([]byte(restoreCommand + "\n")); err != nil {	208		if _, err := pty.Write([]byte(restoreCommand + "\n")); err != nil {
155		d.log.Warn("writing snapshot restore command failed", "err", err, "session", snap.Name, "pane", paneSnap.ID)	209		d.log.Warn("writing snapshot restore command failed", "err", err, "session", snap.Name, "pane", paneSnap.ID)

@@ -164,9 +218,6 @@ func (d *Daemon) restoreSession(ctx context.Context, snap snapcodec.Session) err
164		}	218		}
165		createdAt := int64(snap.CreatedAt)	219		createdAt := int64(snap.CreatedAt)
166		sess := &session{name: snap.Name, ctx: sctx, cancel: cancel, tabs: opened, active: int(snap.Active), terminal: restoreTerm, createdAt: createdAt}	220		sess := &session{name: snap.Name, ctx: sctx, cancel: cancel, tabs: opened, active: int(snap.Active), terminal: restoreTerm, createdAt: createdAt}
167	-	if sess.active < 0 \|\| sess.active >= len(sess.tabs) {
168	-	sess.active = 0
169	-	}
170		sess.snapEligible.Store(true)	221		sess.snapEligible.Store(true)
171		if len(snap.Tabs) > 0 && len(snap.Tabs[0].Panes) > 0 {	222		if len(snap.Tabs) > 0 && len(snap.Tabs[0].Panes) > 0 {
172		sess.cwd = snap.Tabs[0].Panes[0].Cwd	223		sess.cwd = snap.Tabs[0].Panes[0].Cwd

@@ -201,36 +252,51 @@ func (d *Daemon) restoreSession(ctx context.Context, snap snapcodec.Session) err
201		}	252		}
202			253
203		func markSnapshotDirty(sess *session) {	254		func markSnapshotDirty(sess *session) {
204	-	if sess == nil {	255	+	if sess == nil \|\| !sess.snapEligible.Load() {
205		return	256		return
206		}	257		}
207	-	if sess.snapEligible.Load() {	258	+	sess.snapshotMu.Lock()
208	-	sess.snapDirty.Store(true)	259	+	sess.snapshotGeneration++
			260	+	sess.snapDirty.Store(true)
			261	+	sess.signalSnapshotChangedLocked()
			262	+	sess.snapshotMu.Unlock()
			263	+	}
			264	+
			265	+	// snapshotChangeLocked returns a channel closed by the next snapshot state
			266	+	// transition. snapshotMu must be held.
			267	+	func (sess *session) snapshotChangeLocked() chan struct{} {
			268	+	if sess.snapshotChanged == nil {
			269	+	sess.snapshotChanged = make(chan struct{})
209		}	270		}
			271	+	return sess.snapshotChanged
210		}	272		}
211			273
212	-	func seedPaneRows(p *pane, scrollback, visible [][]renderer.Cell) {	274	+	// signalSnapshotChangedLocked wakes waiters after a snapshot state transition.
			275	+	// snapshotMu must be held.
			276	+	func (sess *session) signalSnapshotChangedLocked() {
			277	+	close(sess.snapshotChangeLocked())
			278	+	sess.snapshotChanged = make(chan struct{})
			279	+	}
			280	+
			281	+	func restorePaneTerminal(p *pane, snap snapcodec.Pane) error {
			282	+	if len(snap.Tail) == 0 {
			283	+	return fmt.Errorf("snapshot history: missing tail blob")
			284	+	}
			285	+	if len(snap.Visible) == 0 {
			286	+	return fmt.Errorf("snapshot visible: missing visible blob")
			287	+	}
213		p.mu.Lock()	288		p.mu.Lock()
214		defer p.mu.Unlock()	289		defer p.mu.Unlock()
215	-	p.scrollback = scopy.NewScrollback(defaultScrollbackRows)	290	+	screen, err := vt.NewScreenWithRestoredHistory(p.screen.Frame.Width, p.screen.Frame.Height, vt.HistoryConfig{MaxRows: defaultScrollbackRows}, snap.SealedChunks, snap.Tail)
216	-	for _, row := range scrollback {	291	+	if err != nil {
217	-	p.scrollback.Append(row)	292	+	return fmt.Errorf("snapshot history: %w", err)
218	-	}	293	+	}
219	-	if len(visible) > 0 {	294	+	if err := screen.RestorePrimaryVisible(snap.Visible); err != nil {
220	-	p.screen = vt.NewScreen(p.screen.Frame.Width, p.screen.Frame.Height)	295	+	return fmt.Errorf("snapshot visible: %w", err)
221	-	for y, row := range visible {
222	-	if y >= p.screen.Frame.Height {
223	-	break
224	-	}
225	-	for x, cell := range row {
226	-	if x >= p.screen.Frame.Width {
227	-	break
228	-	}
229	-	p.screen.Frame.Set(x, y, cell)
230	-	}
231	-	}
232		}	296		}
233	-	p.screen.OnLineEvicted = p.scrollback.Append	297	+	p.screen = screen
			298	+	p.history = screen.History()
			299	+	return nil
234		}	300		}
235			301
236		func restorePTYSize(contentRect domain.Rect, tabSize domain.Size) domain.Size {	302		func restorePTYSize(contentRect domain.Rect, tabSize domain.Size) domain.Size {

@@ -251,35 +317,135 @@ func restorePTYSize(contentRect domain.Rect, tabSize domain.Size) domain.Size {
251		return domain.Size{Cols: cols, Rows: rows}	317		return domain.Size{Cols: cols, Rows: rows}
252		}	318		}
253			319
254	-	func (d Daemon) captureSession(sess session) bool {	320	+	// scheduleSnapshot captures a session only once while an immutable capture is
255	-	if d == nil \|\| sess == nil \|\| !d.snapsEnabled \|\| d.snaps == nil {	321	+	// queued or being written. It never waits for the bounded worker queue.
			322	+	func (d Daemon) scheduleSnapshot(sess session) bool {
			323	+	return d.scheduleSnapshotWithFinalFallback(sess, false)
			324	+	}
			325	+
			326	+	// scheduleFinalSnapshot captures the terminal state even when an older capture
			327	+	// is pending. A removed session cannot be retried by snapshotSaver, so the
			328	+	// worker drains its retained terminal state before it stops.
			329	+	func (d Daemon) scheduleFinalSnapshot(sess session) bool {
			330	+	return d.scheduleSnapshotWithFinalFallback(sess, true)
			331	+	}
			332	+
			333	+	func (d Daemon) scheduleSnapshotWithFinalFallback(sess session, final bool) bool {
			334	+	if d == nil \|\| sess == nil \|\| !d.snapsEnabled \|\| d.snaps == nil \|\| !sess.snapEligible.Load() {
			335	+	return true
			336	+	}
			337	+	sess.snapshotMu.Lock()
			338	+	if !sess.snapDirty.Load() \|\| (!final && sess.snapshotPending) {
			339	+	sess.snapshotMu.Unlock()
256		return true	340		return true
257		}	341		}
			342	+	generation := sess.snapshotGeneration
			343	+	sess.snapshotPendingCaptures++
			344	+	sess.snapshotPending = true
			345	+	sess.signalSnapshotChangedLocked()
			346	+	sess.snapshotMu.Unlock()
258			347
			348	+	capture, ok := d.captureSnapshotState(sess, generation)
			349	+	if !ok {
			350	+	d.finishSnapshotCapture(capture, false)
			351	+	return false
			352	+	}
			353	+	if d.enqueueSnapshotCapture(capture) \|\| (final && d.enqueueFinalSnapshotCapture(capture)) {
			354	+	return true
			355	+	}
			356	+	// Coalesce under saturation or shutdown: the latest state stays dirty and
			357	+	// will be captured on a later tick once an active worker has room.
			358	+	d.finishSnapshotCapture(capture, false)
			359	+	return false
			360	+	}
			361	+
			362	+	// enqueueSnapshotCapture serializes worker shutdown with the non-blocking
			363	+	// queue send. The queue is deliberately never closed: producers can race
			364	+	// shutdown without risking a send-on-closed panic.
			365	+	func (d Daemon) enqueueSnapshotCapture(capture snapshotCapture) bool {
			366	+	d.snapshotWorkerMu.Lock()
			367	+	defer d.snapshotWorkerMu.Unlock()
			368	+	if d.snapshotWorkerClosing \|\| d.snapshotWorkerCancel == nil \|\| d.snapshotWorkerCtx == nil \|\| d.snapshotWorkerCtx.Err() != nil {
			369	+	return false
			370	+	}
			371	+	select {
			372	+	case d.snapshotJobs <- capture:
			373	+	return true
			374	+	default:
			375	+	return false
			376	+	}
			377	+	}
			378	+
			379	+	// enqueueFinalSnapshotCapture hands terminal captures to the existing worker
			380	+	// without making session teardown wait behind its normal bounded queue. A
			381	+	// blocked worker retains at most one terminal capture per session, replacing
			382	+	// stale state with the newest immutable capture.
			383	+	func (d Daemon) enqueueFinalSnapshotCapture(capture snapshotCapture) bool {
			384	+	if capture == nil \|\| capture.session == nil {
			385	+	return false
			386	+	}
			387	+	d.snapshotWorkerMu.Lock()
			388	+	if d.snapshotWorkerClosing \|\| d.snapshotWorkerCancel == nil \|\| d.snapshotWorkerCtx == nil \|\| d.snapshotWorkerCtx.Err() != nil {
			389	+	d.snapshotWorkerMu.Unlock()
			390	+	return false
			391	+	}
			392	+	if d.snapshotFinalJobs == nil {
			393	+	d.snapshotFinalJobs = make(map[session]snapshotCapture)
			394	+	}
			395	+	replaced, exists := d.snapshotFinalJobs[capture.session]
			396	+	if !exists && len(d.snapshotFinalJobs) >= snapshotFinalQueueCapacity {
			397	+	d.snapshotWorkerMu.Unlock()
			398	+	d.log.Warn("terminal snapshot retention saturated; capture rejected", "session", capture.name, "capacity", snapshotFinalQueueCapacity)
			399	+	return false
			400	+	}
			401	+	// Finish the stale capture before publishing its replacement. Otherwise a
			402	+	// fast worker could persist the replacement and then have this stale failure
			403	+	// mark the session dirty again.
			404	+	d.finishSnapshotCapture(replaced, false)
			405	+	if !exists {
			406	+	d.snapshotFinalOrder = append(d.snapshotFinalOrder, capture.session)
			407	+	}
			408	+	d.snapshotFinalJobs[capture.session] = capture
			409	+	select {
			410	+	case d.snapshotWorkerFinalWake <- struct{}{}:
			411	+	default:
			412	+	}
			413	+	d.snapshotWorkerMu.Unlock()
			414	+	return true
			415	+	}
			416	+
			417	+	// captureSession rotates history tails and clones visible frames while holding
			418	+	// each pane lock. The returned capture contains only immutable state; encoding
			419	+	// and persistence are deliberately deferred to snapshotEncodeWorker.
			420	+	func (d Daemon) captureSnapshotState(sess session, generation uint64) (*snapshotCapture, bool) {
259		sess.mu.Lock()	421		sess.mu.Lock()
260	-	name := sess.name	422	+	capture := &snapshotCapture{
			423	+	session: sess,
			424	+	generation: generation,
			425	+	name: sess.name,
			426	+	createdAt: uint64(sess.createdAt),
			427	+	active: uint16(max(sess.active, 0)),
			428	+	}
261		ephemeral := sess.ephemeral	429		ephemeral := sess.ephemeral
262	-	createdAt := sess.createdAt
263	-	active := sess.active
264		fallbackCwd := sess.cwd	430		fallbackCwd := sess.cwd
265		tabs := append([]*tab(nil), sess.tabs...)	431		tabs := append([]*tab(nil), sess.tabs...)
266		sess.mu.Unlock()	432		sess.mu.Unlock()
267	-	if ephemeral \|\| name == "" {	433	+	if ephemeral \|\| capture.name == "" {
268	-	return true	434	+	return capture, false
269		}	435		}
270			436
271	-	snap := snapcodec.Session{Name: name, CreatedAt: uint64(createdAt), Active: uint16(max(active, 0))}	437	+	capture.tabs = make([]snapshotCaptureTab, 0, len(tabs))
272		for _, tb := range tabs {	438		for _, tb := range tabs {
273		tb.mu.Lock()	439		tb.mu.Lock()
274	-	tabSnap := snapcodec.Tab{	440	+	tabCapture := snapshotCaptureTab{
275	-	StableID: tb.stableID,	441	+	stableID: tb.stableID,
276	-	Cols: uint16(max(tb.size.Cols, 0)),	442	+	cols: uint16(max(tb.size.Cols, 0)),
277	-	Rows: uint16(max(tb.size.Rows, 0)),	443	+	rows: uint16(max(tb.size.Rows, 0)),
278	-	NextPaneID: uint64(max(tb.nextPaneID, 0)),	444	+	nextPaneID: uint64(max(tb.nextPaneID, 0)),
279	-	Tree: tb.tree.Clone(),	445	+	tree: tb.tree.Clone(),
280		}	446		}
281		if tb.tree != nil {	447		if tb.tree != nil {
282	-	tabSnap.Focus = tb.tree.Focus	448	+	tabCapture.focus = tb.tree.Focus
283		}	449		}
284		panes := make([]*pane, 0, len(tb.panes))	450		panes := make([]*pane, 0, len(tb.panes))
285		for _, p := range tb.panes {	451		for _, p := range tb.panes {

@@ -287,51 +453,274 @@ func (d Daemon) captureSession(sess session) bool {
287		}	453		}
288		tb.mu.Unlock()	454		tb.mu.Unlock()
289			455
290	-	paneIDs := make(map[layout.PaneID]struct{}, len(panes))	456	+	tabCapture.panes = make([]snapshotCapturePane, 0, len(panes))
291	-	paneSnaps := make([]snapcodec.Pane, 0, len(panes))
292		for _, p := range panes {	457		for _, p := range panes {
293		p.mu.Lock()	458		p.mu.Lock()
294	-	pid := 0
295		pty := p.pty	459		pty := p.pty
			460	+	pid := 0
296		if pty != nil {	461		if pty != nil {
297		pid = pty.Pid()	462		pid = pty.Pid()
298		}	463		}
299	-	paneSnap := snapcodec.Pane{	464	+	paneCapture := snapshotCapturePane{
300	-	ID: p.id,	465	+	id: p.id,
301	-	StableID: p.stableID,	466	+	stableID: p.stableID,
302	-	Scrollback: cloneRows(p.scrollback.Snapshot()),	467	+	history: p.history.SealAndView(),
303	-	Visible: cloneRows(p.screen.PrimaryVisibleRows()),	468	+	visible: p.screen.PrimaryVisibleFrame(),
304		}	469		}
305		p.mu.Unlock()	470		p.mu.Unlock()
306	-	paneSnap.Cwd = fallbackCwd	471	+	paneCapture.cwd = fallbackCwd
307		if d.procCwd != nil && pid > 0 {	472		if d.procCwd != nil && pid > 0 {
308		if cwd, err := d.procCwd(pid); err == nil && cwd != "" {	473		if cwd, err := d.procCwd(pid); err == nil && cwd != "" {
309	-	paneSnap.Cwd = cwd	474	+	paneCapture.cwd = cwd
310		}	475		}
311		}	476		}
312	-	paneSnap.Process = d.capturePaneProcess(pty, pid)	477	+	paneCapture.process = d.capturePaneProcess(pty, pid)
313	-	paneIDs[paneSnap.ID] = struct{}{}	478	+	tabCapture.panes = append(tabCapture.panes, paneCapture)
314	-	paneSnaps = append(paneSnaps, paneSnap)	479	+	}
			480	+	capture.tabs = append(capture.tabs, tabCapture)
			481	+	}
			482	+	return capture, true
			483	+	}
			484	+
			485	+	// captureSession remains the synchronous producer-facing trigger for callers
			486	+	// such as teardown and benchmarks; the actual encoding and Write stay async.
			487	+	func (d Daemon) captureSession(sess session) bool {
			488	+	markSnapshotDirty(sess)
			489	+	return d.scheduleSnapshot(sess)
			490	+	}
			491	+
			492	+	func (d Daemon) encodeSnapshotCapture(capture snapshotCapture) ([]byte, error) {
			493	+	snap := snapcodec.Session{Name: capture.name, CreatedAt: capture.createdAt, Active: capture.active, Tabs: make([]snapcodec.Tab, 0, len(capture.tabs))}
			494	+	for _, tabCapture := range capture.tabs {
			495	+	tabSnap := snapcodec.Tab{
			496	+	StableID: tabCapture.stableID, Cols: tabCapture.cols, Rows: tabCapture.rows,
			497	+	NextPaneID: tabCapture.nextPaneID, Focus: tabCapture.focus, Tree: tabCapture.tree,
			498	+	Panes: make([]snapcodec.Pane, 0, len(tabCapture.panes)),
			499	+	}
			500	+	paneIDs := make(map[layout.PaneID]struct{}, len(tabCapture.panes))
			501	+	for _, paneCapture := range tabCapture.panes {
			502	+	sealed, tail, historyErr := vt.MarshalSealedHistory(paneCapture.history)
			503	+	visible, visibleErr := vt.MarshalVisible(paneCapture.visible)
			504	+	if historyErr != nil \|\| visibleErr != nil {
			505	+	return nil, fmt.Errorf("terminal snapshot pane %s: history=%w visible=%w", paneCapture.id, historyErr, visibleErr)
			506	+	}
			507	+	tabSnap.Panes = append(tabSnap.Panes, snapcodec.Pane{ID: paneCapture.id, StableID: paneCapture.stableID, Cwd: paneCapture.cwd, SealedChunks: sealed, Tail: tail, Visible: visible, Process: paneCapture.process})
			508	+	paneIDs[paneCapture.id] = struct{}{}
315		}	509		}
316		if tabSnap.Tree != nil {	510		if tabSnap.Tree != nil {
317		pruneTreeToPanes(tabSnap.Tree.Root, paneIDs)	511		pruneTreeToPanes(tabSnap.Tree.Root, paneIDs)
318		}	512		}
319	-	tabSnap.Panes = paneSnaps
320		snap.Tabs = append(snap.Tabs, tabSnap)	513		snap.Tabs = append(snap.Tabs, tabSnap)
321		}	514		}
			515	+	return d.snapshotMarshal(snap)
			516	+	}
322			517
323	-	data, err := snapcodec.Marshal(snap)	518	+	func (d Daemon) finishSnapshotCapture(capture snapshotCapture, succeeded bool) {
324	-	if err != nil {	519	+	if capture == nil {
325	-	d.log.Warn("marshaling session snapshot failed", "err", err, "session", name)	520	+	return
326	-	return false	521	+	}
			522	+	capture.finishOnce.Do(func() {
			523	+	capture.session.snapshotMu.Lock()
			524	+	if capture.session.snapshotPendingCaptures > 0 {
			525	+	capture.session.snapshotPendingCaptures--
			526	+	}
			527	+	capture.session.snapshotPending = capture.session.snapshotPendingCaptures > 0
			528	+	if succeeded && capture.session.snapshotGeneration == capture.generation {
			529	+	capture.session.snapDirty.Store(false)
			530	+	} else if !succeeded {
			531	+	capture.session.snapDirty.Store(true)
			532	+	}
			533	+	capture.session.signalSnapshotChangedLocked()
			534	+	capture.session.snapshotMu.Unlock()
			535	+	})
			536	+	}
			537	+
			538	+	func (d *Daemon) startSnapshotEncodeWorker() {
			539	+	if d == nil {
			540	+	return
			541	+	}
			542	+	d.snapshotWorkerMu.Lock()
			543	+	if d.snapshotWorkerCancel != nil {
			544	+	d.snapshotWorkerMu.Unlock()
			545	+	return
			546	+	}
			547	+	// Shutdown owns the worker lifetime so it can flush captures after Serve
			548	+	// cancels its parent context. The worker is always stopped explicitly.
			549	+	workerCtx, cancel := context.WithCancel(context.Background())
			550	+	d.snapshotWorkerID++
			551	+	workerID := d.snapshotWorkerID
			552	+	d.snapshotWorkerCtx = workerCtx
			553	+	d.snapshotWorkerCancel = cancel
			554	+	d.snapshotWorkerDone = make(chan struct{})
			555	+	d.snapshotWorkerFlush = make(chan struct{})
			556	+	d.snapshotWorkerFinalWake = make(chan struct{}, 1)
			557	+	d.snapshotFinalJobs = make(map[session]snapshotCapture)
			558	+	d.snapshotFinalOrder = nil
			559	+	d.snapshotWorkerClosing = false
			560	+	d.snapshotWorkerInFlight = nil
			561	+	done := d.snapshotWorkerDone
			562	+	flush := d.snapshotWorkerFlush
			563	+	finalWake := d.snapshotWorkerFinalWake
			564	+	d.snapshotWorkerMu.Unlock()
			565	+	go func() {
			566	+	defer close(done)
			567	+	write := func(capture *snapshotCapture) bool {
			568	+	if workerCtx.Err() != nil \|\| !d.setSnapshotWorkerInFlight(workerID, capture) {
			569	+	d.finishSnapshotCapture(capture, false)
			570	+	return false
			571	+	}
			572	+	data, err := d.encodeSnapshotCapture(capture)
			573	+	if err == nil && workerCtx.Err() == nil {
			574	+	err = d.snaps.Write(capture.name, data)
			575	+	}
			576	+	d.clearSnapshotWorkerInFlight(workerID, capture)
			577	+	if err != nil && workerCtx.Err() == nil {
			578	+	d.log.Warn("writing session snapshot failed", "err", err, "session", capture.name)
			579	+	}
			580	+	d.finishSnapshotCapture(capture, err == nil && workerCtx.Err() == nil)
			581	+	return workerCtx.Err() == nil
			582	+	}
			583	+	drainFinal := func() bool {
			584	+	for {
			585	+	capture := d.takeFinalSnapshotCapture()
			586	+	if capture == nil {
			587	+	return true
			588	+	}
			589	+	if !write(capture) {
			590	+	return false
			591	+	}
			592	+	}
			593	+	}
			594	+	flushQueued := func() {
			595	+	for {
			596	+	select {
			597	+	case capture := <-d.snapshotJobs:
			598	+	if !write(capture) {
			599	+	return
			600	+	}
			601	+	default:
			602	+	_ = drainFinal()
			603	+	return
			604	+	}
			605	+	}
			606	+	}
			607	+	for {
			608	+	select {
			609	+	case <-workerCtx.Done():
			610	+	return
			611	+	case <-flush:
			612	+	flushQueued()
			613	+	return
			614	+	case <-finalWake:
			615	+	if !drainFinal() {
			616	+	return
			617	+	}
			618	+	case capture := <-d.snapshotJobs:
			619	+	if !write(capture) {
			620	+	return
			621	+	}
			622	+	}
			623	+	}
			624	+	}()
			625	+	}
			626	+
			627	+	func (d Daemon) takeFinalSnapshotCapture() snapshotCapture {
			628	+	d.snapshotWorkerMu.Lock()
			629	+	defer d.snapshotWorkerMu.Unlock()
			630	+	for len(d.snapshotFinalOrder) > 0 {
			631	+	sess := d.snapshotFinalOrder[0]
			632	+	d.snapshotFinalOrder[0] = nil
			633	+	d.snapshotFinalOrder = d.snapshotFinalOrder[1:]
			634	+	capture := d.snapshotFinalJobs[sess]
			635	+	delete(d.snapshotFinalJobs, sess)
			636	+	if capture != nil {
			637	+	return capture
			638	+	}
327		}	639		}
328	-	if err := d.snaps.Write(name, data); err != nil {	640	+	return nil
329	-	d.log.Warn("writing session snapshot failed", "err", err, "session", name)	641	+	}
			642	+
			643	+	func (d Daemon) setSnapshotWorkerInFlight(workerID uint64, capture snapshotCapture) bool {
			644	+	d.snapshotWorkerMu.Lock()
			645	+	defer d.snapshotWorkerMu.Unlock()
			646	+	if d.snapshotWorkerID != workerID \|\| d.snapshotWorkerCancel == nil \|\| d.snapshotWorkerCtx == nil \|\| d.snapshotWorkerCtx.Err() != nil {
330		return false	647		return false
331		}	648		}
			649	+	d.snapshotWorkerInFlight = capture
332		return true	650		return true
333		}	651		}
334			652
			653	+	func (d Daemon) clearSnapshotWorkerInFlight(workerID uint64, capture snapshotCapture) {
			654	+	d.snapshotWorkerMu.Lock()
			655	+	defer d.snapshotWorkerMu.Unlock()
			656	+	if d.snapshotWorkerID == workerID && d.snapshotWorkerInFlight == capture {
			657	+	d.snapshotWorkerInFlight = nil
			658	+	}
			659	+	}
			660	+
			661	+	func (d *Daemon) stopSnapshotEncodeWorker() {
			662	+	if d == nil {
			663	+	return
			664	+	}
			665	+	d.snapshotWorkerMu.Lock()
			666	+	cancel := d.snapshotWorkerCancel
			667	+	if cancel == nil \|\| d.snapshotWorkerClosing {
			668	+	d.snapshotWorkerMu.Unlock()
			669	+	return
			670	+	}
			671	+	// Stop accepting producers, then let the single worker drain its bounded
			672	+	// queue. A non-context-aware store may still block indefinitely, so the
			673	+	// injected clock bounds how long teardown waits before detaching it.
			674	+	d.snapshotWorkerClosing = true
			675	+	flush := d.snapshotWorkerFlush
			676	+	done := d.snapshotWorkerDone
			677	+	close(flush)
			678	+	timer := d.clock.NewTimer(snapshotFinalFlushTimeout)
			679	+	d.snapshotWorkerMu.Unlock()
			680	+
			681	+	select {
			682	+	case <-done:
			683	+	timer.Stop()
			684	+	d.finishStoppedSnapshotWorker(false)
			685	+	case <-timer.C():
			686	+	cancel()
			687	+	d.finishStoppedSnapshotWorker(true)
			688	+	}
			689	+	}
			690	+
			691	+	func (d *Daemon) finishStoppedSnapshotWorker(abandoned bool) {
			692	+	d.snapshotWorkerMu.Lock()
			693	+	inFlight := d.snapshotWorkerInFlight
			694	+	d.snapshotWorkerCtx = nil
			695	+	d.snapshotWorkerCancel = nil
			696	+	d.snapshotWorkerDone = nil
			697	+	d.snapshotWorkerFlush = nil
			698	+	d.snapshotWorkerFinalWake = nil
			699	+	d.snapshotWorkerClosing = false
			700	+	d.snapshotWorkerInFlight = nil
			701	+	queued := make([]*snapshotCapture, 0, len(d.snapshotJobs)+len(d.snapshotFinalJobs))
			702	+	for _, capture := range d.snapshotFinalJobs {
			703	+	queued = append(queued, capture)
			704	+	}
			705	+	d.snapshotFinalJobs = nil
			706	+	d.snapshotFinalOrder = nil
			707	+	for {
			708	+	select {
			709	+	case capture := <-d.snapshotJobs:
			710	+	queued = append(queued, capture)
			711	+	default:
			712	+	d.snapshotWorkerMu.Unlock()
			713	+	if abandoned {
			714	+	d.finishSnapshotCapture(inFlight, false)
			715	+	}
			716	+	for _, capture := range queued {
			717	+	d.finishSnapshotCapture(capture, false)
			718	+	}
			719	+	return
			720	+	}
			721	+	}
			722	+	}
			723	+
335		func (d Daemon) capturePaneProcess(pty interface{ ForegroundPgid() (int, error) }, shellPid int) snapcodec.Process {	724		func (d Daemon) capturePaneProcess(pty interface{ ForegroundPgid() (int, error) }, shellPid int) snapcodec.Process {
336		if d == nil \|\| pty == nil \|\| shellPid <= 0 \|\| d.procGroupArgv == nil {	725		if d == nil \|\| pty == nil \|\| shellPid <= 0 \|\| d.procGroupArgv == nil {
337		return nil	726		return nil

@@ -368,27 +757,14 @@ func (d *Daemon) snapshotSaver(ctx context.Context) {
368		sessions := d.sessionsSnapshotLocked()	757		sessions := d.sessionsSnapshotLocked()
369		d.mu.Unlock()	758		d.mu.Unlock()
370		for _, sess := range sessions {	759		for _, sess := range sessions {
371	-	if sess.snapEligible.Load() && sess.snapDirty.Swap(false) {	760	+	if sess.snapEligible.Load() && sess.snapDirty.Load() {
372	-	if !d.captureSession(sess) {	761	+	d.scheduleSnapshot(sess)
373	-	sess.snapDirty.Store(true)
374	-	}
375		}	762		}
376		}	763		}
377		t.Reset(snapshotInterval)	764		t.Reset(snapshotInterval)
378		}	765		}
379		}	766		}
380			767
381	-	func cloneRows(rows [][]renderer.Cell) [][]renderer.Cell {
382	-	if len(rows) == 0 {
383	-	return nil
384	-	}
385	-	out := make([][]renderer.Cell, len(rows))
386	-	for i, row := range rows {
387	-	out[i] = append([]renderer.Cell(nil), row...)
388	-	}
389	-	return out
390	-	}
391	-
392		func pruneTreeToPanes(n *layout.Node, panes map[layout.PaneID]struct{}) bool {	768		func pruneTreeToPanes(n *layout.Node, panes map[layout.PaneID]struct{}) bool {
393		if n == nil {	769		if n == nil {
394		return false	770		return false

A internal/usecase/daemon/snapshot_acceptance_test.go +249 -0 collapsed to keep the page light

load diff

M internal/usecase/daemon/snapshot_restore_test.go +151 -35

 	"github.com/bnema/vev/internal/usecase/layout"
 	snapcodec "github.com/bnema/vev/internal/usecase/snapshot"
 	"github.com/bnema/vev/pkg/renderer"
+	"github.com/bnema/vev/pkg/vt"
 )
+func terminalPane(t *testing.T, id, stableID, cwd string, historyRows, visibleRows [][]renderer.Cell) snapcodec.Pane {
+	t.Helper()
+	h := vt.NewHistory(vt.HistoryConfig{MaxRows: defaultScrollbackRows})
+	for _, row := range historyRows {
+		h.Append(row)
+	}
+	sealed, tail, err := vt.MarshalSealedHistory(h.SealAndView())
+	require.NoError(t, err)
+	frame := renderer.NewFrame(40, 24)
+	for y, row := range visibleRows {
+		copy(frame.Row(y), row)
+	}
+	visible, err := vt.MarshalVisible(frame)
+	require.NoError(t, err)
+	return snapcodec.Pane{ID: layout.PaneID(id), StableID: stableID, Cwd: cwd, SealedChunks: sealed, Tail: tail, Visible: visible}
+}
+func TestRestorePaneTerminalRejectsMissingOrMalformedCanonicalBlobs(t *testing.T) {
+	valid := terminalPane(t, "pane", "p", "/work", nil, nil)
+	tests := []struct {
+		name   string
+		mutate func(*snapcodec.Pane)
+	}{
+		{name: "missing tail", mutate: func(p *snapcodec.Pane) { p.Tail = nil }},
+		{name: "missing visible", mutate: func(p *snapcodec.Pane) { p.Visible = nil }},
+		{name: "malformed tail", mutate: func(p *snapcodec.Pane) { p.Tail = []byte("bad") }},
+		{name: "malformed visible", mutate: func(p *snapcodec.Pane) { p.Visible = []byte("bad") }},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			snap := valid
+			tt.mutate(&snap)
+			p := newPane("pane", newRestorePTY(), domain.Size{Cols: 40, Rows: 24})
+			require.Error(t, restorePaneTerminal(p, snap))
+		})
+	}
+}
+func TestRestoreSessionRejectsInvalidActiveTab(t *testing.T) {
+	factory := &restorePTYFactory{}
+	d := newTestDaemon(t, factory, stubClock{})
+	err := d.restoreSession(context.Background(), snapcodec.Session{Name: "bad-active", Active: 1, Tabs: []snapcodec.Tab{{
+		Cols: 80, Rows: 24, Tree: layout.NewTree("pane-1"), Panes: []snapcodec.Pane{emptyTerminalPane(t, "pane-1", "/tmp")},
+	}}})
+	require.Error(t, err)
+	require.Empty(t, factory.opens)
+}
+func TestRestoreSessionClosesOpenedPTYWhenTerminalRestoreFails(t *testing.T) {
+	factory := &restorePTYFactory{}
+	d := newTestDaemon(t, factory, stubClock{})
+	pane := emptyTerminalPane(t, "pane-1", "/tmp")
+	pane.Tail = []byte("malformed")
+	err := d.restoreSession(context.Background(), snapcodec.Session{Name: "bad", Tabs: []snapcodec.Tab{{
+		Cols: 80, Rows: 24, Tree: layout.NewTree("pane-1"), Panes: []snapcodec.Pane{pane},
+	}}})
+	require.Error(t, err)
+	require.Len(t, factory.opens, 1)
+	require.Equal(t, 1, factory.opens[0].pty.closeCount())
+}
 func TestRestoreSnapshotsRestoresLayoutCwdAndRows(t *testing.T) {
 	store := &restoreSnapshotStore{}
 	store.blobs = []ports.SnapshotBlob{{Name: "work", Data: mustSnapshotBytes(t, snapcodec.Session{
 				layout.NewLeaf("pane-2"),
 			}}},
 			Panes: []snapcodec.Pane{
-				{ID: "pane-1", StableID: "p_saved_1", Cwd: "/one", Scrollback: [][]renderer.Cell{cells("old1")}, Visible: [][]renderer.Cell{cells("vis1")}},
+				terminalPane(t, "pane-1", "p_saved_1", "/one", [][]renderer.Cell{cells("old1")}, [][]renderer.Cell{cells("vis1")}),
-				{ID: "pane-2", StableID: "p_saved_2", Cwd: "/two", Scrollback: [][]renderer.Cell{cells("old2")}, Visible: [][]renderer.Cell{cells("vis2")}},
+				terminalPane(t, "pane-2", "p_saved_2", "/two", [][]renderer.Cell{cells("old2")}, [][]renderer.Cell{cells("vis2")}),
 			},
 		}},
 	})}}
 	p := tb.panes["pane-2"]
 	tb.mu.Unlock()
 	p.mu.Lock()
-	require.Equal(t, "old2", rowText(p.scrollback.View().Row(0)))
+	require.Same(t, p.screen.History(), p.history)
-	require.Equal(t, 1, p.scrollback.Len())
+	require.Equal(t, "old2", rowText(p.history.View().Row(0)))
+	require.Equal(t, 1, p.history.Len())
 	require.Equal(t, "vis2", rowText(p.screen.PrimaryVisibleRows()[0][:4]))
-	copySnap := scopy.NewSnapshot(p.scrollback, p.screen.Frame)
+	copySnap := scopy.NewSnapshot(p.history, p.screen.Frame)
 	require.Equal(t, "old2", rowText(copySnap.Row(0)[:4]))
 	require.Equal(t, "vis2", rowText(copySnap.Row(1)[:4]))
 	p.mu.Unlock()
 func TestRestoreSnapshotsLeavesFloatingUninitializedAndInactiveTabsCold(t *testing.T) {
 	store := &restoreSnapshotStore{blobs: []ports.SnapshotBlob{{Name: "work", Data: mustSnapshotBytes(t, snapcodec.Session{
 		Name: "work", Active: 0, Tabs: []snapcodec.Tab{
-			{Cols: 80, Rows: 24, Tree: layout.NewTree("pane-1"), Panes: []snapcodec.Pane{{ID: "pane-1", Cwd: "/one"}}},
+			{Cols: 80, Rows: 24, Tree: layout.NewTree("pane-1"), Panes: []snapcodec.Pane{emptyTerminalPane(t, "pane-1", "/one")}},
-			{Cols: 80, Rows: 24, Tree: layout.NewTree("pane-1"), Panes: []snapcodec.Pane{{ID: "pane-1", Cwd: "/two"}}},
+			{Cols: 80, Rows: 24, Tree: layout.NewTree("pane-1"), Panes: []snapcodec.Pane{emptyTerminalPane(t, "pane-1", "/two")}},
 		},
 	})}}}
 	factory := &restorePTYFactory{}
 func processRestoreTestStore(t *testing.T, name string, proc *snapcodec.Process) *restoreSnapshotStore {
 	t.Helper()
+	pane := emptyTerminalPane(t, "pane-1", "/tmp")
+	pane.Process = proc
 	return &restoreSnapshotStore{blobs: []ports.SnapshotBlob{{Name: name, Data: mustSnapshotBytes(t, snapcodec.Session{
 		Name: name,
 		Tabs: []snapcodec.Tab{{
-			Cols: 80,
+			Cols:  80,
-			Rows: 24,
+			Rows:  24,
-			Tree: layout.NewTree("pane-1"),
+			Tree:  layout.NewTree("pane-1"),
-			Panes: []snapcodec.Pane{{
+			Panes: []snapcodec.Pane{pane},
-				ID:      "pane-1",
-				Cwd:     "/tmp",
-				Process: proc,
-			}},
 		}},
 	})}}}
 }
 						Cols: 80,
 						Rows: 24,
 						Tree: layout.NewTree("pane-1"),
-						Panes: []snapcodec.Pane{{
+						Panes: func() []snapcodec.Pane {
-							ID:      "pane-1",
+							pane := emptyTerminalPane(t, "pane-1", "/one")
-							Cwd:     "/one",
+							pane.Process = lessProc
-							Process: lessProc,
+							return []snapcodec.Pane{pane}
-						}},
+						}(),
 					},
 					{
-						Cols: 80,
+						Cols:  80,
-						Rows: 24,
+						Rows:  24,
-						Tree: layout.NewTree("pane-1"),
+						Tree:  layout.NewTree("pane-1"),
-						Panes: []snapcodec.Pane{{
+						Panes: []snapcodec.Pane{emptyTerminalPane(t, "pane-1", "/two")},
-							ID:  "pane-1",
-							Cwd: "/two",
-						}},
 					},
 				},
 			})}}},
 	d := newTestDaemon(t, factory, stubClock{})
 	d.ApplyConfig(domain.Config{Snapshot: domain.SnapshotConfig{RestoreProcessesSet: true, RestoreProcesses: []string{"pi"}}})
-	require.NoError(t, d.restoreSession(context.Background(), snapcodec.Session{Name: "agent", Tabs: []snapcodec.Tab{{Cols: 80, Rows: 24, Tree: layout.NewTree("pane-1"), Panes: []snapcodec.Pane{{ID: "pane-1", Cwd: "/tmp", Process: proc}}}}}))
+	pane := emptyTerminalPane(t, "pane-1", "/tmp")
+	pane.Process = proc
+	require.NoError(t, d.restoreSession(context.Background(), snapcodec.Session{Name: "agent", Tabs: []snapcodec.Tab{{Cols: 80, Rows: 24, Tree: layout.NewTree("pane-1"), Panes: []snapcodec.Pane{pane}}}}))
 	require.Len(t, factory.opens, 1)
 	require.Equal(t, []string{"pi --resume abc123\n"}, factory.opens[0].pty.writes)
 }
+func TestRestoreSessionPassesRestoreContextToPTYOpen(t *testing.T) {
+	factory := &restorePTYFactory{}
+	d := newTestDaemon(t, factory, stubClock{})
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	require.NoError(t, d.restoreSession(ctx, snapcodec.Session{Name: "restore-context", Tabs: []snapcodec.Tab{{
+		Cols: 80, Rows: 24, Tree: layout.NewTree("pane-1"), Panes: []snapcodec.Pane{emptyTerminalPane(t, "pane-1", "/tmp")},
+	}}}))
+	require.Len(t, factory.opens, 1)
+	require.Same(t, ctx, factory.opens[0].ctx)
+}
 func TestRestoreSnapshotsOpensCollapsedStackPanesWithValidPTYSize(t *testing.T) {
 	store := &restoreSnapshotStore{blobs: []ports.SnapshotBlob{{Name: "stacked", Data: mustSnapshotBytes(t, snapcodec.Session{
 		Name: "stacked",
 			Tree: &layout.Tree{Focus: "b", Root: &layout.Node{Kind: layout.Stack, Expanded: "b", Children: []*layout.Node{
 				layout.NewLeaf("a"), layout.NewLeaf("b"), layout.NewLeaf("c"),
 			}}},
-			Panes: []snapcodec.Pane{{ID: "a", Cwd: "/a"}, {ID: "b", Cwd: "/b"}, {ID: "c", Cwd: "/c"}},
+			Panes: []snapcodec.Pane{
+				emptyTerminalPane(t, "a", "/a"),
+				emptyTerminalPane(t, "b", "/b"),
+				emptyTerminalPane(t, "c", "/c"),
+			},
 		}},
 	})}}}
 	factory := &restorePTYFactory{}
 }
 func TestRestoreSnapshotsSkipsLiveCorruptAndEmpty(t *testing.T) {
-	valid := mustSnapshotBytes(t, snapcodec.Session{Name: "live", CreatedAt: 1, Tabs: []snapcodec.Tab{{Cols: 80, Rows: 24, Tree: layout.NewTree("pane-1"), Panes: []snapcodec.Pane{{ID: "pane-1", Cwd: "/ok"}}}}})
+	valid := mustSnapshotBytes(t, snapcodec.Session{
+		Name:      "live",
+		CreatedAt: 1,
+		Tabs: []snapcodec.Tab{{
+			Cols:  80,
+			Rows:  24,
+			Tree:  layout.NewTree("pane-1"),
+			Panes: []snapcodec.Pane{emptyTerminalPane(t, "pane-1", "/ok")},
+		}},
+	})
 	badVersion := append([]byte(nil), valid...)
 	binary.BigEndian.PutUint16(badVersion[4:6], 1)
 	tests := []struct {
 			Cols:  80,
 			Rows:  24,
 			Tree:  layout.NewTree("pane-1"),
-			Panes: []snapcodec.Pane{{ID: "pane-1", Cwd: "/bad"}},
+			Panes: []snapcodec.Pane{emptyTerminalPane(t, "pane-1", "/bad")},
 		}},
 	})}}}
 	factory := &restorePTYFactory{}
 		loadFn: func() ([]ports.SnapshotBlob, error) {
 			close(loaded)
 			<-releaseLoad
-			return []ports.SnapshotBlob{{Name: "work", Data: mustSnapshotBytes(t, snapcodec.Session{Name: "work", Tabs: []snapcodec.Tab{{Cols: 80, Rows: 24, Tree: layout.NewTree("pane-1"), Panes: []snapcodec.Pane{{ID: "pane-1", Cwd: "/restored"}}}}})}}, nil
+			return []ports.SnapshotBlob{{Name: "work", Data: mustSnapshotBytes(t, snapcodec.Session{
+				Name: "work",
+				Tabs: []snapcodec.Tab{{
+					Cols:  80,
+					Rows:  24,
+					Tree:  layout.NewTree("pane-1"),
+					Panes: []snapcodec.Pane{emptyTerminalPane(t, "pane-1", "/restored")},
+				}},
+			})}}, nil
 		},
 	}
 	factory := &restorePTYFactory{}
 	require.Equal(t, ports.ErrServerShutdown, perr.code)
 }
+func emptyTerminalPane(t *testing.T, id, cwd string) snapcodec.Pane {
+	t.Helper()
+	return terminalPane(t, id, "", cwd, nil, nil)
+}
 func mustSnapshotBytes(t *testing.T, s snapcodec.Session) []byte {
 	t.Helper()
 	b, err := snapcodec.Marshal(s)
 }
 type restorePTYOpen struct {
+	ctx  context.Context
 	dir  string
 	size domain.Size
 	env  []string
 	pty  *restorePTY
 }
-func (f *restorePTYFactory) Open(_ string, _ []string, env []string, dir string, sz domain.Size) (ports.PTY, error) {
+func (f *restorePTYFactory) Open(ctx context.Context, _ string, _ []string, env []string, dir string, sz domain.Size) (ports.PTY, error) {
 	f.mu.Lock()
 	defer f.mu.Unlock()
 	pty := newRestorePTY()
 	pty.writeErr = f.writeErr
-	f.opens = append(f.opens, restorePTYOpen{dir: dir, size: sz, env: append([]string(nil), env...), pty: pty})
+	f.opens = append(f.opens, restorePTYOpen{ctx: ctx, dir: dir, size: sz, env: append([]string(nil), env...), pty: pty})
 	return pty, nil
 }
 type restorePTY struct {
 	mu       sync.Mutex
 	writes   []string
+	closes   int
 	writeErr error
 	done     chan struct{}
 	once     sync.Once
 	}
 	return len(b), nil
 }
-func (p *restorePTY) Close() error                 { p.once.Do(func() { close(p.done) }); return nil }
+func (p *restorePTY) Close() error {
+	p.mu.Lock()
+	p.closes++
+	p.mu.Unlock()
+	p.once.Do(func() { close(p.done) })
+	return nil
+}
+func (p *restorePTY) closeCount() int {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	return p.closes
+}
 func (p *restorePTY) Resize(domain.Size) error     { return nil }
 func (p *restorePTY) Pid() int                     { return 0 }
 func (p *restorePTY) ForegroundPgid() (int, error) { return 0, nil }

M internal/usecase/daemon/snapshot_test.go +730 -8 collapsed to keep the page light

load diff

A internal/usecase/daemon/stalled_attachment_test.go +78 -0 collapsed to keep the page light

load diff

M internal/usecase/daemon/status.go +1 -24 collapsed to keep the page light

load diff

M internal/usecase/daemon/status_test.go +72 -20 collapsed to keep the page light

load diff

A internal/usecase/daemon/test_wait_helpers_test.go +55 -0 collapsed to keep the page light

load diff

M internal/usecase/daemon/theme_resolve.go +0 -1


@@ -28,7 +28,6 @@ func (d Daemon) applyHostThemeLocked(sess session, ac *attachedClient, t theme
28			28
29		if ac != nil {	29	if ac != nil {
30		ac.setTheme(t)	30	ac.setTheme(t)
31	-	ac.composed.invalidate()
32		}	31	}
33		for _, tb := range tabs {	32	for _, tb := range tabs {
34		tb.mu.Lock()	33	tb.mu.Lock()

A internal/usecase/daemon/transactional_resize.go +292 -0 collapsed to keep the page light

load diff

A internal/usecase/daemon/transactional_resize_test.go +403 -0 collapsed to keep the page light

load diff

A internal/usecase/daemon/transport_observability_test.go +390 -0 collapsed to keep the page light

load diff

M internal/usecase/daemon/wire.go +1 -24 collapsed to keep the page light

load diff

A internal/usecase/snapshot/benchmark_test.go +122 -0 collapsed to keep the page light

load diff

M internal/usecase/snapshot/codec.go +84 -313 collapsed to keep the page light

load diff

M internal/usecase/snapshot/codec_test.go +390 -312 collapsed to keep the page light

load diff

A internal/usecase/snapshot/preflight.go +449 -0 collapsed to keep the page light

load diff

A internal/usecase/snapshot/preflight_acceptance_test.go +45 -0 collapsed to keep the page light

load diff

M internal/usecase/snapshot/snapshot.go +10 -12 collapsed to keep the page light

load diff

M internal/usecase/visualsearch/model_test.go +2 -2


@@ -61,14 +61,14 @@ func TestVisualSearchNewSnapshotOwnsSourceRows(t *testing.T) {
61		require.Equal(t, "alpha", matches[0].Text)	61		require.Equal(t, "alpha", matches[0].Text)
62		}	62		}
63			63
64	-	func TestVisualSearchCloneSharesDocumentAndKeepsUIStateIndependent(t *testing.T) {	64	+	func TestVisualSearchCloneRetainsDocumentAndKeepsUIStateIndependent(t *testing.T) {
65		m := New(testSnapshot("alpha", "beta alpha"))	65		m := New(testSnapshot("alpha", "beta alpha"))
66		for _, r := range "alpha" {	66		for _, r := range "alpha" {
67		m.Insert(r)	67		m.Insert(r)
68		}	68		}
69		clone := m.Clone()	69		clone := m.Clone()
70			70
71	-	require.Same(t, &m.snapshot.Row(0)[0], &clone.snapshot.Row(0)[0])	71	+	require.Equal(t, m.snapshot.Row(0), clone.snapshot.Row(0))
72		clone.Down()	72		clone.Down()
73		clone.Insert('z')	73		clone.Insert('z')
74			74