perf(daemon): baseline large-history workloads

commit: b590931ca8a04cf8b7e5772d66d7d1c359783f9e
author: brice · 3mo ago · 2026-07-11 05:32 UTC
committer: brice · 3mo ago · 2026-07-11 06:15 UTC
parent: 2a2a3fb
signature: GPG VerifiedCD88F9E564E0102646FB6077A1A07C7C7C3E349D · arch-gpg Verified for the committer email

3 files · +156 -68

M docs/performance.md +53 -26

@@ -1,43 +1,70 @@
1	1		# Performance methodology (M5)
2	2
3		-	Enable daemon pprof explicitly with `VEV_PPROF_ADDR=127.0.0.1:6060 vev --daemon` (or by exporting the variable before launching clients). When unset, no debug HTTP server is started.
	3	+	This baseline measures in-process daemon work only. It deliberately does not
	4	+	measure coordinator behavior, chunk encoding, compression, GC tuning, or real
	5	+	network latency. Output and snapshot metrics count encoded output bytes and
	6	+	frames produced by the daemon; they are useful proxies for local rendering and
	7	+	IPC payload work, not measurements of a Unix socket, SSH, or UDP transport.
	8	+	Real Unix-socket, SSH, and UDP impairment/latency benchmarks are deferred.
4	9
5		-	Run microbenchmarks with:
	10	+	## Large-history daemon baseline
	11	+
	12	+	Run the full daemon matrix once when establishing a baseline:
6	13
7	14		```sh
8		-	go test ./pkg/renderer ./pkg/vt ./internal/adapters/ipc ./internal/usecase/daemon -run '^$' -bench=. -benchmem
	15	+	go test ./internal/usecase/daemon -run '^$' -bench '^BenchmarkDaemonHistory' -benchtime=1x -benchmem
9	16		```
10	17
11		-	Demo flood workloads are scripted in `scripts/bench-workloads.sh` for `yes`, `seq`, and `cat` styles for raw/tmux runs. vev does not currently support `vev new <name> -- command`, so automated vev workload command overrides are intentionally not documented. Capture bytes/frame and syscalls/frame externally (for example with `strace -c` around the client/daemon) and allocations with `-benchmem`/pprof.
	18	+	For repeatable comparison of the focused control workload, run three one-second
	19	+	samples (pin CPU/governor and record the host when comparing runs):
	20	+
	21	+	```sh
	22	+	go test ./internal/usecase/daemon -run '^$' -bench '^BenchmarkDaemonHistoryLivePaint/1tab-1pane-control$' -benchtime=1s -count=3 -benchmem
	23	+	```
	24	+
	25	+	The matrix exercises live paint, snapshot capture, copy-mode entry, copy search,
	26	+	and resize at `120x40` client geometry. Each pane is populated with 10,000
	27	+	full-width deterministic history rows containing a stable `needle-<tab>-<row>`
	28	+	prefix and patterned text. The layouts are:
	29	+
	30	+	- 1 tab x 1 pane (control)
	31	+	- 1 tab x 4 panes
	32	+	- 4 tabs x 1 pane
	33	+	- 4 tabs x 4 panes
	34	+	- 8 tabs x 1 pane
12	35
13		-	## Local baseline status
	36	+	The fixture starts no PTY readers, schedulers, clocks, or transports. It primes
	37	+	the renderer shadow before timing; live paint then alternates fixed writes. The
	38	+	reported `outputbytes/op`, `framepayloadbytes/op`, and `outputframes/op` values
	39	+	are encoded in-process output proxies. `snapshotbytes/op` is the serialized
	40	+	in-process snapshot size. They must not be presented as network throughput or
	41	+	latency.
14	42
15		-	Full end-to-end vev-vs-tmux workload baselines are deferred/waived for this non-interactive branch because they require an interactive terminal/daemon pair, external tracing tools around both processes, and (for automated vev flood workloads) future command override support. Until that exists, only start the daemon/client normally and do not claim a scripted vev flood command. tmux/raw comparison commands remain useful for shaping workloads:
	43	+	Capture a pre-change result outside the repository, including Go and kernel
	44	+	context:
16	45
17	46		```sh
18		-	VEV_PPROF_ADDR=127.0.0.1:6060 strace -ff -c -o vev-daemon.strace vev --daemon
19		-	strace -ff -c -o vev-client.strace vev new perf-manual
20		-	strace -ff -c -o tmux-yes.strace tmux new-session -d -s perf-yes 'timeout 30s yes'
21		-	strace -ff -c -o tmux-seq.strace tmux new-session -d -s perf-seq 'timeout 30s sh -c "while :; do seq 1 10000; done"'
22		-	strace -ff -c -o tmux-cat.strace tmux new-session -d -s perf-cat 'timeout 30s sh -c "while :; do cat /path/to/sample.txt; done"'
	47	+	{
	48	+	date -Is
	49	+	go version
	50	+	uname -a
	51	+	git rev-parse HEAD
	52	+	go test ./internal/usecase/daemon -run '^$' -bench '^BenchmarkDaemonHistory' -benchtime=1x -benchmem
	53	+	go test ./internal/usecase/daemon -run '^$' -bench '^BenchmarkDaemonHistoryLivePaint/1tab-1pane-control$' -benchtime=1s -count=3 -benchmem
	54	+	} > /tmp/vev-history-before.txt 2>&1
23	55		```
24	56
25		-	Record host CPU, OS/kernel, terminal emulator, vev commit, tmux version, terminal size, workload duration, raw command lines, syscall summaries, bytes written per rendered frame, and heap profiles or allocation counts before comparing.
	57	+	## Lower-level smoke baseline
26	58
27		-	## UDP mobile congestion behavior
	59	+	Use this command to check renderer, VT, IPC, and daemon microbenchmarks without
	60	+	claiming an end-to-end transport result:
28	61
29		-	The datagram transport uses one shared congestion-responsive byte pacer for initial sends and retransmits. Retransmissions therefore consume the same byte budget as new data instead of bypassing congestion feedback. Datagram clients use a one-state output window: the cumulative ACK releases that state, and newer repaint state coalesces while it remains unacknowledged rather than retaining an unbounded queue.
30		-
31		-	Microbenchmarks executed locally for this review using `go test ./internal/adapters/ipc ./internal/usecase/daemon -run '^$' -bench=. -benchmem`:
	62	+	```sh
	63	+	go test ./pkg/renderer ./pkg/vt ./internal/adapters/ipc ./internal/usecase/daemon -run '^$' -bench=. -benchmem
	64	+	```
32	65
33		-	\| benchmark \| methodology \|
34		-	\| --- \| --- \|
35		-	\| BenchmarkTransportSend \| writes encoded frames to an in-memory discard `net.Conn`; allocations reported here are Send-side frame assembly costs. \|
36		-	\| BenchmarkTransportRecvReuse \| reads a pre-encoded frame from a looping in-memory reader; no `Send` call, goroutine, channel, or frame production work runs inside the measured loop, so allocations reported here are Recv-side costs only. \|
37		-	\| BenchmarkPaintCachedSinglePaneDamage \| writes a small damage update into one pane, paints through the daemon compose→diff→send path, and advances the output ACK each iteration so the measured loop does not hit the unacked-output bailout. \|
	66	+	Enable daemon pprof explicitly when collecting a daemon profile:
38	67
39		-	\| workload \| vev baseline \| tmux baseline \| notes \|
40		-	\| --- \| --- \| --- \| --- \|
41		-	\| yes \| not run in this non-interactive review environment \| not run in this non-interactive review environment \| use same terminal size and duration \|
42		-	\| seq \| not run in this non-interactive review environment \| not run in this non-interactive review environment \| use same terminal size and duration \|
43		-	\| cat \| not run in this non-interactive review environment \| not run in this non-interactive review environment \| use same input file, terminal size, and duration \|
	68	+	```sh
	69	+	VEV_PPROF_ADDR=127.0.0.1:6060 vev --daemon
	70	+	```

@@ -1,43 +1,70 @@
1		# Performance methodology (M5)	1		# Performance methodology (M5)
2			2
3	-	Enable daemon pprof explicitly with `VEV_PPROF_ADDR=127.0.0.1:6060 vev --daemon` (or by exporting the variable before launching clients). When unset, no debug HTTP server is started.	3	+	This baseline measures in-process daemon work only. It deliberately does not
			4	+	measure coordinator behavior, chunk encoding, compression, GC tuning, or real
			5	+	network latency. Output and snapshot metrics count encoded output bytes and
			6	+	frames produced by the daemon; they are useful proxies for local rendering and
			7	+	IPC payload work, not measurements of a Unix socket, SSH, or UDP transport.
			8	+	Real Unix-socket, SSH, and UDP impairment/latency benchmarks are deferred.
4			9
5	-	Run microbenchmarks with:	10	+	## Large-history daemon baseline
			11	+
			12	+	Run the full daemon matrix once when establishing a baseline:
6			13
7		```sh	14		```sh
8	-	go test ./pkg/renderer ./pkg/vt ./internal/adapters/ipc ./internal/usecase/daemon -run '^$' -bench=. -benchmem	15	+	go test ./internal/usecase/daemon -run '^$' -bench '^BenchmarkDaemonHistory' -benchtime=1x -benchmem
9		```	16		```
10			17
11	-	Demo flood workloads are scripted in `scripts/bench-workloads.sh` for `yes`, `seq`, and `cat` styles for raw/tmux runs. vev does not currently support `vev new <name> -- command`, so automated vev workload command overrides are intentionally not documented. Capture bytes/frame and syscalls/frame externally (for example with `strace -c` around the client/daemon) and allocations with `-benchmem`/pprof.	18	+	For repeatable comparison of the focused control workload, run three one-second
			19	+	samples (pin CPU/governor and record the host when comparing runs):
			20	+
			21	+	```sh
			22	+	go test ./internal/usecase/daemon -run '^$' -bench '^BenchmarkDaemonHistoryLivePaint/1tab-1pane-control$' -benchtime=1s -count=3 -benchmem
			23	+	```
			24	+
			25	+	The matrix exercises live paint, snapshot capture, copy-mode entry, copy search,
			26	+	and resize at `120x40` client geometry. Each pane is populated with 10,000
			27	+	full-width deterministic history rows containing a stable `needle-<tab>-<row>`
			28	+	prefix and patterned text. The layouts are:
			29	+
			30	+	- 1 tab x 1 pane (control)
			31	+	- 1 tab x 4 panes
			32	+	- 4 tabs x 1 pane
			33	+	- 4 tabs x 4 panes
			34	+	- 8 tabs x 1 pane
12			35
13	-	## Local baseline status	36	+	The fixture starts no PTY readers, schedulers, clocks, or transports. It primes
			37	+	the renderer shadow before timing; live paint then alternates fixed writes. The
			38	+	reported `outputbytes/op`, `framepayloadbytes/op`, and `outputframes/op` values
			39	+	are encoded in-process output proxies. `snapshotbytes/op` is the serialized
			40	+	in-process snapshot size. They must not be presented as network throughput or
			41	+	latency.
14			42
15	-	Full end-to-end vev-vs-tmux workload baselines are deferred/waived for this non-interactive branch because they require an interactive terminal/daemon pair, external tracing tools around both processes, and (for automated vev flood workloads) future command override support. Until that exists, only start the daemon/client normally and do not claim a scripted vev flood command. tmux/raw comparison commands remain useful for shaping workloads:	43	+	Capture a pre-change result outside the repository, including Go and kernel
			44	+	context:
16			45
17		```sh	46		```sh
18	-	VEV_PPROF_ADDR=127.0.0.1:6060 strace -ff -c -o vev-daemon.strace vev --daemon	47	+	{
19	-	strace -ff -c -o vev-client.strace vev new perf-manual	48	+	date -Is
20	-	strace -ff -c -o tmux-yes.strace tmux new-session -d -s perf-yes 'timeout 30s yes'	49	+	go version
21	-	strace -ff -c -o tmux-seq.strace tmux new-session -d -s perf-seq 'timeout 30s sh -c "while :; do seq 1 10000; done"'	50	+	uname -a
22	-	strace -ff -c -o tmux-cat.strace tmux new-session -d -s perf-cat 'timeout 30s sh -c "while :; do cat /path/to/sample.txt; done"'	51	+	git rev-parse HEAD
			52	+	go test ./internal/usecase/daemon -run '^$' -bench '^BenchmarkDaemonHistory' -benchtime=1x -benchmem
			53	+	go test ./internal/usecase/daemon -run '^$' -bench '^BenchmarkDaemonHistoryLivePaint/1tab-1pane-control$' -benchtime=1s -count=3 -benchmem
			54	+	} > /tmp/vev-history-before.txt 2>&1
23		```	55		```
24			56
25	-	Record host CPU, OS/kernel, terminal emulator, vev commit, tmux version, terminal size, workload duration, raw command lines, syscall summaries, bytes written per rendered frame, and heap profiles or allocation counts before comparing.	57	+	## Lower-level smoke baseline
26			58
27	-	## UDP mobile congestion behavior	59	+	Use this command to check renderer, VT, IPC, and daemon microbenchmarks without
			60	+	claiming an end-to-end transport result:
28			61
29	-	The datagram transport uses one shared congestion-responsive byte pacer for initial sends and retransmits. Retransmissions therefore consume the same byte budget as new data instead of bypassing congestion feedback. Datagram clients use a one-state output window: the cumulative ACK releases that state, and newer repaint state coalesces while it remains unacknowledged rather than retaining an unbounded queue.	62	+	```sh
30	-		63	+	go test ./pkg/renderer ./pkg/vt ./internal/adapters/ipc ./internal/usecase/daemon -run '^$' -bench=. -benchmem
31	-	Microbenchmarks executed locally for this review using `go test ./internal/adapters/ipc ./internal/usecase/daemon -run '^$' -bench=. -benchmem`:	64	+	```
32			65
33	-	\| benchmark \| methodology \|	66	+	Enable daemon pprof explicitly when collecting a daemon profile:
34	-	\| --- \| --- \|
35	-	\| BenchmarkTransportSend \| writes encoded frames to an in-memory discard `net.Conn`; allocations reported here are Send-side frame assembly costs. \|
36	-	\| BenchmarkTransportRecvReuse \| reads a pre-encoded frame from a looping in-memory reader; no `Send` call, goroutine, channel, or frame production work runs inside the measured loop, so allocations reported here are Recv-side costs only. \|
37	-	\| BenchmarkPaintCachedSinglePaneDamage \| writes a small damage update into one pane, paints through the daemon compose→diff→send path, and advances the output ACK each iteration so the measured loop does not hit the unacked-output bailout. \|
38			67
39	-	\| workload \| vev baseline \| tmux baseline \| notes \|	68	+	```sh
40	-	\| --- \| --- \| --- \| --- \|	69	+	VEV_PPROF_ADDR=127.0.0.1:6060 vev --daemon
41	-	\| yes \| not run in this non-interactive review environment \| not run in this non-interactive review environment \| use same terminal size and duration \|	70	+	```
42	-	\| seq \| not run in this non-interactive review environment \| not run in this non-interactive review environment \| use same terminal size and duration \|
43	-	\| cat \| not run in this non-interactive review environment \| not run in this non-interactive review environment \| use same input file, terminal size, and duration \|

M internal/usecase/daemon/performance_benchmark_test.go +103 -12


@@ -36,6 +36,83 @@ func TestPerformanceFixtureCounters(t *testing.T) {
36	36		require.True(t, fixture.resized())
37	37		}
38	38
	39	+	func TestPerformanceFixtureLargeHistoryTopology(t *testing.T) {
	40	+	fixture := newPerformanceFixture(t, performanceConfig{size: domain.Size{Cols: 120, Rows: 40}, tabs: 4, panes: 4, historyRows: 10_000})
	41	+
	42	+	require.Len(t, fixture.sess.tabs, 4)
	43	+	for _, tb := range fixture.sess.tabs {
	44	+	require.Len(t, tb.panes, 4)
	45	+	for _, p := range tb.panes {
	46	+	require.Equal(t, 10_000, p.scrollback.Len())
	47	+	}
	48	+	}
	49	+	}
	50	+
	51	+	func BenchmarkDaemonHistoryLivePaint(b *testing.B) {
	52	+	benchmarkDaemonLargeHistory(b, "live-paint", func(f *performanceFixture) { f.paintLive() })
	53	+	}
	54	+	func BenchmarkDaemonHistorySnapshotCapture(b *testing.B) {
	55	+	benchmarkDaemonLargeHistory(b, "snapshot-capture", func(f *performanceFixture) { f.captureSnapshot() })
	56	+	}
	57	+	func BenchmarkDaemonHistoryCopyEnter(b *testing.B) {
	58	+	benchmarkDaemonLargeHistory(b, "copy-enter", func(f *performanceFixture) { f.d.enterCopyMode(f.sess, f.ac) })
	59	+	}
	60	+	func BenchmarkDaemonHistoryCopySearch(b *testing.B) {
	61	+	benchmarkDaemonLargeHistory(b, "copy-search", func(f *performanceFixture) { f.d.handleInput(f.sess, f.ac, []byte("/needle\r")) })
	62	+	}
	63	+	func BenchmarkDaemonHistoryResize(b *testing.B) {
	64	+	benchmarkDaemonLargeHistory(b, "resize", func(f *performanceFixture) { f.resize() })
	65	+	}
	66	+
	67	+	func benchmarkDaemonLargeHistory(b testing.B, workload string, run func(performanceFixture)) {
	68	+	b.Helper()
	69	+	for _, topology := range []struct {
	70	+	name string
	71	+	tabs, panes int
	72	+	}{
	73	+	{name: "1tab-1pane-control", tabs: 1, panes: 1},
	74	+	{name: "1tab-4panes", tabs: 1, panes: 4},
	75	+	{name: "4tabs-1pane", tabs: 4, panes: 1},
	76	+	{name: "4tabs-4panes", tabs: 4, panes: 4},
	77	+	{name: "8tabs-1pane", tabs: 8, panes: 1},
	78	+	} {
	79	+	b.Run(topology.name, func(b *testing.B) {
	80	+	fixture := newPerformanceFixture(b, performanceConfig{
	81	+	size: domain.Size{Cols: 120, Rows: 40}, tabs: topology.tabs, panes: topology.panes, historyRows: 10_000,
	82	+	})
	83	+	if workload == "copy-search" {
	84	+	fixture.d.enterCopyMode(fixture.sess, fixture.ac)
	85	+	}
	86	+	b.ReportAllocs()
	87	+	fixture.resetMetrics()
	88	+	b.ResetTimer()
	89	+	for b.Loop() {
	90	+	run(fixture)
	91	+	fixture.ac.ackOutputState(fixture.ac.output.next)
	92	+	}
	93	+	b.StopTimer()
	94	+	metrics := fixture.metrics()
	95	+	if workload == "live-paint" && metrics.outputFrames != uint64(b.N) {
	96	+	b.Fatalf("live paint emitted %d frames for %d operations", metrics.outputFrames, b.N)
	97	+	}
	98	+	benchmarkReportMetrics(b, metrics, b.N, 10_000)
	99	+	})
	100	+	}
	101	+	}
	102	+
	103	+	func benchmarkReportMetrics(b *testing.B, metrics performanceMetrics, operations, historyRows int) {
	104	+	b.Helper()
	105	+	if operations == 0 {
	106	+	return
	107	+	}
	108	+	perOperation := float64(operations)
	109	+	b.ReportMetric(float64(metrics.outputFrames)/perOperation, "outputframes/op")
	110	+	b.ReportMetric(float64(metrics.outputBytes)/perOperation, "outputbytes/op")
	111	+	b.ReportMetric(float64(metrics.outputPayloadBytes)/perOperation, "framepayloadbytes/op")
	112	+	b.ReportMetric(float64(metrics.snapshotBytes)/perOperation, "snapshotbytes/op")
	113	+	b.ReportMetric(float64(historyRows), "historyrows/pane")
	114	+	}
	115	+
39	116		func TestPerformanceFixturePaintLiveUsesPrecomputedAlternatingWrites(t *testing.T) {
40	117		fixture := newPerformanceFixture(t, performanceConfig{})
41	118

@@ -54,7 +131,10 @@ func TestPerformanceFixturePaintLiveUsesPrecomputedAlternatingWrites(t *testing.
54	131		// performanceConfig intentionally describes only an in-process daemon shape:
55	132		// no PTY readers, schedulers, clocks, or transports are started by this fixture.
56	133		type performanceConfig struct {
57		-	size domain.Size
	134	+	size domain.Size
	135	+	tabs int
	136	+	panes int
	137	+	historyRows int
58	138		}
59	139
60	140		type performanceMetrics struct {

@@ -83,7 +163,17 @@ func newPerformanceFixture(t testing.TB, config performanceConfig) *performanceF
83	163		if !config.size.Valid() {
84	164		config.size = domain.Size{Cols: 80, Rows: 24}
85	165		}
86		-	d, sess, ac, _ := newManualSessionWithPTYs(t, nil, nil)
	166	+	if config.tabs == 0 {
	167	+	config.tabs = 1
	168	+	}
	169	+	if config.panes == 0 {
	170	+	config.panes = 2
	171	+	}
	172	+	if config.historyRows == 0 {
	173	+	config.historyRows = 1
	174	+	}
	175	+	ptys := make([]ports.PTY, config.tabs)
	176	+	d, sess, ac, _ := newManualSessionWithPTYs(t, ptys...)
87	177		output := &countingOutputTransport{}
88	178		ac.tr = output
89	179		ac.size = config.size

@@ -102,7 +192,7 @@ func newPerformanceFixture(t testing.TB, config performanceConfig) *performanceF
102	192		}
103	193		WithSnapshotStore(fixture.snaps)(d)
104	194		for tabIndex, tb := range sess.tabs {
105		-	fixture.configureTab(tb, tabIndex, tabSize(config.size))
	195	+	fixture.configureTab(tb, tabIndex, config.panes, config.historyRows, tabSize(config.size))
106	196		}
107	197		fixture.activePane = fixture.findActivePane()
108	198

@@ -113,28 +203,29 @@ func newPerformanceFixture(t testing.TB, config performanceConfig) *performanceF
113	203		return fixture
114	204		}
115	205
116		-	func (f performanceFixture) configureTab(tb tab, tabIndex int, size domain.Size) {
	206	+	func (f performanceFixture) configureTab(tb tab, tabIndex, paneCount, historyRows int, size domain.Size) {
117	207		f.t.Helper()
118	208		tb.mu.Lock()
	209	+	defer tb.mu.Unlock()
119	210		tb.size = size
120	211		left := tb.focusedPane()
121		-	rightID := layout.PaneID(fmt.Sprintf("pane-%d-right", tabIndex+1))
122		-	right := newPane(rightID, nil, size)
123		-	tb.panes[rightID] = right
124		-	require.NoError(f.t, tb.tree.Split(left.id, layout.Right, true, rightID, domain.Rect{Width: size.Cols, Height: size.Rows}))
	212	+	for paneIndex := 1; paneIndex < paneCount; paneIndex++ {
	213	+	id := layout.PaneID(fmt.Sprintf("pane-%d-%d", tabIndex+1, paneIndex+1))
	214	+	tb.panes[id] = newPane(id, nil, size)
	215	+	require.NoError(f.t, tb.tree.Split(left.id, layout.Right, true, id, domain.Rect{Width: size.Cols, Height: size.Rows}))
	216	+	}
125	217		placements, ok := layout.Solve(tb.tree.Root, domain.Rect{Width: size.Cols, Height: size.Rows})
126	218		require.True(f.t, ok, "performance fixture layout must be solvable")
127		-	require.Len(f.t, placements, 2)
	219	+	require.Len(f.t, placements, paneCount)
128	220		f.d.applyLayoutLocked(tb)
129	221		for _, p := range tb.panes {
130	222		p.mu.Lock()
131		-	width, height := p.screen.Frame.Width, p.screen.Frame.Height
132		-	for row := range max(height-1, 1) {
	223	+	width := p.screen.Frame.Width
	224	+	for row := 0; p.scrollback.Len() < historyRows; row++ {
133	225		p.screen.Write([]byte(performanceFullWidthRow(width, tabIndex, row) + "\r\n"))
134	226		}
135	227		p.mu.Unlock()
136	228		}
137		-	tb.mu.Unlock()
138	229		}
139	230
140	231		// performanceFullWidthRow produces stable, full-width visible content without


@@ -36,6 +36,83 @@ func TestPerformanceFixtureCounters(t *testing.T) {
36		require.True(t, fixture.resized())	36		require.True(t, fixture.resized())
37		}	37		}
38			38
			39	+	func TestPerformanceFixtureLargeHistoryTopology(t *testing.T) {
			40	+	fixture := newPerformanceFixture(t, performanceConfig{size: domain.Size{Cols: 120, Rows: 40}, tabs: 4, panes: 4, historyRows: 10_000})
			41	+
			42	+	require.Len(t, fixture.sess.tabs, 4)
			43	+	for _, tb := range fixture.sess.tabs {
			44	+	require.Len(t, tb.panes, 4)
			45	+	for _, p := range tb.panes {
			46	+	require.Equal(t, 10_000, p.scrollback.Len())
			47	+	}
			48	+	}
			49	+	}
			50	+
			51	+	func BenchmarkDaemonHistoryLivePaint(b *testing.B) {
			52	+	benchmarkDaemonLargeHistory(b, "live-paint", func(f *performanceFixture) { f.paintLive() })
			53	+	}
			54	+	func BenchmarkDaemonHistorySnapshotCapture(b *testing.B) {
			55	+	benchmarkDaemonLargeHistory(b, "snapshot-capture", func(f *performanceFixture) { f.captureSnapshot() })
			56	+	}
			57	+	func BenchmarkDaemonHistoryCopyEnter(b *testing.B) {
			58	+	benchmarkDaemonLargeHistory(b, "copy-enter", func(f *performanceFixture) { f.d.enterCopyMode(f.sess, f.ac) })
			59	+	}
			60	+	func BenchmarkDaemonHistoryCopySearch(b *testing.B) {
			61	+	benchmarkDaemonLargeHistory(b, "copy-search", func(f *performanceFixture) { f.d.handleInput(f.sess, f.ac, []byte("/needle\r")) })
			62	+	}
			63	+	func BenchmarkDaemonHistoryResize(b *testing.B) {
			64	+	benchmarkDaemonLargeHistory(b, "resize", func(f *performanceFixture) { f.resize() })
			65	+	}
			66	+
			67	+	func benchmarkDaemonLargeHistory(b testing.B, workload string, run func(performanceFixture)) {
			68	+	b.Helper()
			69	+	for _, topology := range []struct {
			70	+	name string
			71	+	tabs, panes int
			72	+	}{
			73	+	{name: "1tab-1pane-control", tabs: 1, panes: 1},
			74	+	{name: "1tab-4panes", tabs: 1, panes: 4},
			75	+	{name: "4tabs-1pane", tabs: 4, panes: 1},
			76	+	{name: "4tabs-4panes", tabs: 4, panes: 4},
			77	+	{name: "8tabs-1pane", tabs: 8, panes: 1},
			78	+	} {
			79	+	b.Run(topology.name, func(b *testing.B) {
			80	+	fixture := newPerformanceFixture(b, performanceConfig{
			81	+	size: domain.Size{Cols: 120, Rows: 40}, tabs: topology.tabs, panes: topology.panes, historyRows: 10_000,
			82	+	})
			83	+	if workload == "copy-search" {
			84	+	fixture.d.enterCopyMode(fixture.sess, fixture.ac)
			85	+	}
			86	+	b.ReportAllocs()
			87	+	fixture.resetMetrics()
			88	+	b.ResetTimer()
			89	+	for b.Loop() {
			90	+	run(fixture)
			91	+	fixture.ac.ackOutputState(fixture.ac.output.next)
			92	+	}
			93	+	b.StopTimer()
			94	+	metrics := fixture.metrics()
			95	+	if workload == "live-paint" && metrics.outputFrames != uint64(b.N) {
			96	+	b.Fatalf("live paint emitted %d frames for %d operations", metrics.outputFrames, b.N)
			97	+	}
			98	+	benchmarkReportMetrics(b, metrics, b.N, 10_000)
			99	+	})
			100	+	}
			101	+	}
			102	+
			103	+	func benchmarkReportMetrics(b *testing.B, metrics performanceMetrics, operations, historyRows int) {
			104	+	b.Helper()
			105	+	if operations == 0 {
			106	+	return
			107	+	}
			108	+	perOperation := float64(operations)
			109	+	b.ReportMetric(float64(metrics.outputFrames)/perOperation, "outputframes/op")
			110	+	b.ReportMetric(float64(metrics.outputBytes)/perOperation, "outputbytes/op")
			111	+	b.ReportMetric(float64(metrics.outputPayloadBytes)/perOperation, "framepayloadbytes/op")
			112	+	b.ReportMetric(float64(metrics.snapshotBytes)/perOperation, "snapshotbytes/op")
			113	+	b.ReportMetric(float64(historyRows), "historyrows/pane")
			114	+	}
			115	+
39		func TestPerformanceFixturePaintLiveUsesPrecomputedAlternatingWrites(t *testing.T) {	116		func TestPerformanceFixturePaintLiveUsesPrecomputedAlternatingWrites(t *testing.T) {
40		fixture := newPerformanceFixture(t, performanceConfig{})	117		fixture := newPerformanceFixture(t, performanceConfig{})
41			118

@@ -54,7 +131,10 @@ func TestPerformanceFixturePaintLiveUsesPrecomputedAlternatingWrites(t *testing.
54		// performanceConfig intentionally describes only an in-process daemon shape:	131		// performanceConfig intentionally describes only an in-process daemon shape:
55		// no PTY readers, schedulers, clocks, or transports are started by this fixture.	132		// no PTY readers, schedulers, clocks, or transports are started by this fixture.
56		type performanceConfig struct {	133		type performanceConfig struct {
57	-	size domain.Size	134	+	size domain.Size
			135	+	tabs int
			136	+	panes int
			137	+	historyRows int
58		}	138		}
59			139
60		type performanceMetrics struct {	140		type performanceMetrics struct {

@@ -83,7 +163,17 @@ func newPerformanceFixture(t testing.TB, config performanceConfig) *performanceF
83		if !config.size.Valid() {	163		if !config.size.Valid() {
84		config.size = domain.Size{Cols: 80, Rows: 24}	164		config.size = domain.Size{Cols: 80, Rows: 24}
85		}	165		}
86	-	d, sess, ac, _ := newManualSessionWithPTYs(t, nil, nil)	166	+	if config.tabs == 0 {
			167	+	config.tabs = 1
			168	+	}
			169	+	if config.panes == 0 {
			170	+	config.panes = 2
			171	+	}
			172	+	if config.historyRows == 0 {
			173	+	config.historyRows = 1
			174	+	}
			175	+	ptys := make([]ports.PTY, config.tabs)
			176	+	d, sess, ac, _ := newManualSessionWithPTYs(t, ptys...)
87		output := &countingOutputTransport{}	177		output := &countingOutputTransport{}
88		ac.tr = output	178		ac.tr = output
89		ac.size = config.size	179		ac.size = config.size

@@ -102,7 +192,7 @@ func newPerformanceFixture(t testing.TB, config performanceConfig) *performanceF
102		}	192		}
103		WithSnapshotStore(fixture.snaps)(d)	193		WithSnapshotStore(fixture.snaps)(d)
104		for tabIndex, tb := range sess.tabs {	194		for tabIndex, tb := range sess.tabs {
105	-	fixture.configureTab(tb, tabIndex, tabSize(config.size))	195	+	fixture.configureTab(tb, tabIndex, config.panes, config.historyRows, tabSize(config.size))
106		}	196		}
107		fixture.activePane = fixture.findActivePane()	197		fixture.activePane = fixture.findActivePane()
108			198

@@ -113,28 +203,29 @@ func newPerformanceFixture(t testing.TB, config performanceConfig) *performanceF
113		return fixture	203		return fixture
114		}	204		}
115			205
116	-	func (f performanceFixture) configureTab(tb tab, tabIndex int, size domain.Size) {	206	+	func (f performanceFixture) configureTab(tb tab, tabIndex, paneCount, historyRows int, size domain.Size) {
117		f.t.Helper()	207		f.t.Helper()
118		tb.mu.Lock()	208		tb.mu.Lock()
			209	+	defer tb.mu.Unlock()
119		tb.size = size	210		tb.size = size
120		left := tb.focusedPane()	211		left := tb.focusedPane()
121	-	rightID := layout.PaneID(fmt.Sprintf("pane-%d-right", tabIndex+1))	212	+	for paneIndex := 1; paneIndex < paneCount; paneIndex++ {
122	-	right := newPane(rightID, nil, size)	213	+	id := layout.PaneID(fmt.Sprintf("pane-%d-%d", tabIndex+1, paneIndex+1))
123	-	tb.panes[rightID] = right	214	+	tb.panes[id] = newPane(id, nil, size)
124	-	require.NoError(f.t, tb.tree.Split(left.id, layout.Right, true, rightID, domain.Rect{Width: size.Cols, Height: size.Rows}))	215	+	require.NoError(f.t, tb.tree.Split(left.id, layout.Right, true, id, domain.Rect{Width: size.Cols, Height: size.Rows}))
			216	+	}
125		placements, ok := layout.Solve(tb.tree.Root, domain.Rect{Width: size.Cols, Height: size.Rows})	217		placements, ok := layout.Solve(tb.tree.Root, domain.Rect{Width: size.Cols, Height: size.Rows})
126		require.True(f.t, ok, "performance fixture layout must be solvable")	218		require.True(f.t, ok, "performance fixture layout must be solvable")
127	-	require.Len(f.t, placements, 2)	219	+	require.Len(f.t, placements, paneCount)
128		f.d.applyLayoutLocked(tb)	220		f.d.applyLayoutLocked(tb)
129		for _, p := range tb.panes {	221		for _, p := range tb.panes {
130		p.mu.Lock()	222		p.mu.Lock()
131	-	width, height := p.screen.Frame.Width, p.screen.Frame.Height	223	+	width := p.screen.Frame.Width
132	-	for row := range max(height-1, 1) {	224	+	for row := 0; p.scrollback.Len() < historyRows; row++ {
133		p.screen.Write([]byte(performanceFullWidthRow(width, tabIndex, row) + "\r\n"))	225		p.screen.Write([]byte(performanceFullWidthRow(width, tabIndex, row) + "\r\n"))
134		}	226		}
135		p.mu.Unlock()	227		p.mu.Unlock()
136		}	228		}
137	-	tb.mu.Unlock()
138		}	229		}
139			230
140		// performanceFullWidthRow produces stable, full-width visible content without	231		// performanceFullWidthRow produces stable, full-width visible content without

M internal/usecase/daemon/render_test.go +0 -30


@@ -648,36 +648,6 @@ func TestComposeTabFrameCachedTitleBarsDoNotAllocate(t *testing.T) {
648	648		require.Zero(t, allocs, "valid cached title-bar renders must not allocate")
649	649		}
650	650
651		-	func BenchmarkPaintCachedSinglePaneDamage(b *testing.B) {
652		-	win := newTab(nil, domain.Size{Cols: 81, Rows: 24})
653		-	left := win.focusedPane()
654		-	right := newPane("pane-2", nil, domain.Size{Cols: 40, Rows: 24})
655		-	win.tree.Root = &layout.Node{Kind: layout.Split, Dir: layout.Horizontal, Children: []*layout.Node{layout.NewLeaf(left.id), layout.NewLeaf(right.id)}}
656		-	win.tree.Focus = right.id
657		-	win.panes[right.id] = right
658		-	left.screen.Write([]byte("left"))
659		-	right.screen.Write([]byte("right"))
660		-	sess := &session{id: "s", name: "work", tabs: []*tab{win}, active: 0}
661		-	ac := &attachedClient{tr: discardTransport{}, output: newOutputStateStream(), size: domain.Size{Cols: 81, Rows: 26}}
662		-	ac.initOverlays()
663		-	sess.client = ac
664		-	ac.setSession(sess)
665		-	d := New(nil, stubClock{}, slog.New(slog.NewTextHandler(io.Discard, nil)))
666		-	d.paint(sess, ac, true)
667		-	left.screen.ClearDamage()
668		-	right.screen.ClearDamage()
669		-
670		-	b.ReportAllocs()
671		-	b.ResetTimer()
672		-	for i := 0; i < b.N; i++ {
673		-	right.screen.Write([]byte("x\b"))
674		-	d.paint(sess, ac, false)
675		-	// Mirror the client ACK pump so the steady state measures compose→diff→send,
676		-	// not the ack-gate bailout after maxUnackedOutputStates frames.
677		-	ac.ackOutputState(ac.output.next)
678		-	}
679		-	}
680		-
681	651		func TestComposeTabFrameStackUsesShellFallback(t *testing.T) {
682	652		win := newTab(nil, domain.Size{Cols: 20, Rows: 5})
683	653		p := win.focusedPane()


@@ -648,36 +648,6 @@ func TestComposeTabFrameCachedTitleBarsDoNotAllocate(t *testing.T) {
648		require.Zero(t, allocs, "valid cached title-bar renders must not allocate")	648	require.Zero(t, allocs, "valid cached title-bar renders must not allocate")
649		}	649	}
650			650
651	-	func BenchmarkPaintCachedSinglePaneDamage(b *testing.B) {
652	-	win := newTab(nil, domain.Size{Cols: 81, Rows: 24})
653	-	left := win.focusedPane()
654	-	right := newPane("pane-2", nil, domain.Size{Cols: 40, Rows: 24})
655	-	win.tree.Root = &layout.Node{Kind: layout.Split, Dir: layout.Horizontal, Children: []*layout.Node{layout.NewLeaf(left.id), layout.NewLeaf(right.id)}}
656	-	win.tree.Focus = right.id
657	-	win.panes[right.id] = right
658	-	left.screen.Write([]byte("left"))
659	-	right.screen.Write([]byte("right"))
660	-	sess := &session{id: "s", name: "work", tabs: []*tab{win}, active: 0}
661	-	ac := &attachedClient{tr: discardTransport{}, output: newOutputStateStream(), size: domain.Size{Cols: 81, Rows: 26}}
662	-	ac.initOverlays()
663	-	sess.client = ac
664	-	ac.setSession(sess)
665	-	d := New(nil, stubClock{}, slog.New(slog.NewTextHandler(io.Discard, nil)))
666	-	d.paint(sess, ac, true)
667	-	left.screen.ClearDamage()
668	-	right.screen.ClearDamage()
669	-
670	-	b.ReportAllocs()
671	-	b.ResetTimer()
672	-	for i := 0; i < b.N; i++ {
673	-	right.screen.Write([]byte("x\b"))
674	-	d.paint(sess, ac, false)
675	-	// Mirror the client ACK pump so the steady state measures compose→diff→send,
676	-	// not the ack-gate bailout after maxUnackedOutputStates frames.
677	-	ac.ackOutputState(ac.output.next)
678	-	}
679	-	}
680	-
681		func TestComposeTabFrameStackUsesShellFallback(t *testing.T) {	651	func TestComposeTabFrameStackUsesShellFallback(t *testing.T) {
682		win := newTab(nil, domain.Size{Cols: 20, Rows: 5})	652	win := newTab(nil, domain.Size{Cols: 20, Rows: 5})
683		p := win.focusedPane()	653	p := win.focusedPane()