Add a jank scenario and three jank plants, captured

[ramesh130]

Closes ramesh130/perfettoagent#6

A jank scenario for devicelab, three plants for it, and fifteen captures. devicelab/lab run jank [--plant <name>] opens the feed, runs its plan once without tracing as a warm-up, and checks it scrolled back to the top. It then traces the same plan: 3 rounds of 8 swipes, each round ending in two taps (pause, then play the watched row). Each trace is about 55 s and 29–32 MB.

Decisions

• The click needed a handler. Feed rows had no click handler, and the screen picker can't be hit at a device-independent point. So the demo gains tap-to-pause on the feed: tap anywhere on the list to pause the watched row, tap again to play it. That makes the middle of the screen the tap target on any phone (ui_tap_centre, a screen fraction like the swipes). It is a tap gesture plus a semantics action, not clickable, which would have merged every row into one accessibility node.
• The click is in the same trace. A case is one trace pair, so every regression has to fit in one run. The rests keep the tap frames apart from the scroll frames. There are 6 taps, so a single slow tap stands out as one of six.
• scenario_setup hook: it gets the demo to the feed and warms it up before a playing trace starts. open_feed and wait_session_state move to lib/playback.sh, shared with the leak hunt.
• No gfx (perfetto/sources/jank.pbtxt: view, dalvik, binder_driver, sched). android.frames.timeline still finds every frame.
• Plant sizes come from pilots (devicelab/plants/README.md has the tables):
  • feed-tap-sleep: 120 ms, as the roadmap says.
  • feed-grain-allocations: 1,000,000 boxed floats in each scrolled frame's draw, not 50k. At 50k, GC moved but no frame metric did. Work done between frames showed up only as fewer frames.
  • feed-row-remeasure: the padding breathes by 4 dp, so every visible row is re-measured every frame (ADR-0007). A re-measure of these small rows moved p50 but not p95, so each row's re-measure also fits its title in 0.05 sp steps.

Captures

All captures are on superplayer_verify_36 (API 36), from base 57e3144, clean tree. Runs are in devicelab/out/<run>-jank/ of the main checkout; pilots are in devicelab/out/jank-pilots/. Plant patch sha256s: sleep 27700ab54513…, grain ba6902ce089a…, remeasure 1a29bb777a35….

run	build	frames	app jank	any jank	p50	p95	p99	doFrame p95	asleep in onTouch	GCs	GC time
20260924T033342Z (B1)	clean	676	54.7%	94.4%	40.4	81.5	154.5	38.0	0	7	259
20260924T033648Z (S1)	feed-tap-sleep	733	45.6%	97.4%	34.4	72.6	127.3	36.3	724.7 (6)	8	436
20260924T033945Z (G1)	feed-grain-allocations	427	66.5%	96.7%	40.7	87.7	138.0	60.3	0	149	10356
20260924T034233Z (R1)	feed-row-remeasure	1643	80.3%	100.0%	62.9	92.7	124.1	69.3	0	203	13632
20260924T034924Z (C1)	clean	730	42.2%	95.9%	35.5	72.3	123.5	36.6	0	6	172
20260924T035208Z (B2)	clean	746	45.3%	96.9%	37.1	76.5	148.1	39.4	0	8	302
20260924T035447Z (S2)	feed-tap-sleep	710	37.5%	97.0%	33.7	65.2	122.4	31.9	726.3 (6)	6	186
20260924T035727Z (G2)	feed-grain-allocations	444	65.3%	96.4%	40.7	89.1	129.9	67.0	0	142	11044
20260924T040009Z (R2)	feed-row-remeasure	1729	73.5%	99.9%	62.5	88.5	124.2	65.5	0	196	12153
20260924T040632Z (C2)	clean	834	44.5%	96.8%	39.3	73.3	136.0	37.0	0	8	256
20260924T040915Z (B3)	clean	824	37.1%	97.2%	33.5	67.8	115.8	32.5	2.6 (0)	8	428
20260924T041151Z (S3)	feed-tap-sleep	789	42.0%	97.5%	33.0	71.8	130.3	34.3	724.4 (6)	8	406
20260924T041431Z (G3)	feed-grain-allocations	406	64.3%	95.8%	40.1	86.0	120.4	61.2	0	134	9484
20260924T041712Z (R3)	feed-row-remeasure	1796	73.0%	100.0%	62.1	84.2	115.5	62.6	0	196	11802
20260924T042341Z (C3)	clean	760	43.0%	97.8%	38.7	73.3	139.9	35.2	0	8	377

By hand:

• Sleep: six sleeps of 120–123 ms inside AndroidOwner:onTouch in every planted run, 724–726 ms in total. No clean run has one over 16 ms.
• Grain: 134–149 GCs taking 9.5–11.0 s, against 6–8 GCs taking 0.17–0.43 s. The main thread waits for GC (GC: Wait For Completion Alloc) only in planted runs. doFrame p95 is 60–67 ms against 32–39. App jank is 64–67% against 37–55%; that gap is 10 points, less than the widest clean pair.
• Remeasure: p50 is 62–63 ms against 33–40, and app jank 73–80%. p95 is 84–93 ms against 68–82: every planted run is above every clean run, but the smallest gap is 2.7 ms, inside the clean spread of 13.7 ms. p99 does not separate.
• Clean pairs: B1/C1, B2/C2, B3/C3.

Frames stdlib evidence (for perfettoagent#8)

Checked with perfettoagent's pinned trace processor on API 36 without gfx:

• Frame count and duration agree. android_frames finds the window's frames plus the few that never reached the timeline: 701 = 676 + 25 in B1. Its dur p95 matches the window layer's (81.2 vs 81.5 ms).
• The frame-to-timeline join is unreliable here. actual_frame_timeline_id usually points at a video SurfaceView layer's row (563 of 701). App jank read through that join is 12.7%, against 54.7% from the window layer. Add MediaRequest with content identity, ordered sources, and start position #8 must filter by layer.
• jank_type != 'None' is saturated at 92–100% on the emulator (Buffer Stuffing, Prediction Error). The usable signal is App Deadline Missed.

Review (two axes)

• Standards: everything fixed in 775b0a0 and the docs commit:
  • Fixes: the clickable accessibility merge, the pause race (the service player is now confirmed paused), duplicated feed-opening code, unexplained constants, the SQL headers and the GC scope (collections on HeapTaskDaemon, waits by thread), the tap-coordinates paragraph in the harness README, the docs index, the CLAUDE.md commands, and knob validation plus tests.
  • Left as judgement calls: the plausibility of the grain and title plants, and duplicated screen-size parsing in ui.sh.
• Spec:
  • Done: the plants README rows and sizing, the capture table mapped to the roadmap metrics, and the clean pairs.
  • Open, outside this repo: perfettoagent should record an ADR for the 1M allocation, the title fit, and "jank = App Deadline Missed on the window layer" (roadmap item 5).
  • Known limitation: rebase-merge rewrites shas, so 57e3144 is recorded in run.json, and its main equivalent will be noted on the issue.

Checks

• devicelab/test/selftest: devicelab 104, jank 20, leak-hunt 34, startup 16 passed.
• ./gradlew assemble check --max-workers=1: BUILD SUCCESSFUL, 1187 JUnit tests, 0 failures.
• Demo assembleDebug lintDebug spotlessCheck: green.
• Tree clean after the captures.

Skipped

• No new issue for the main-thread block when the pool recycles a player (Compose:onForgotten, 2–3 s per run); it is noted against Measure UI jank during playback in the feed demo #51 in devicelab/jank/README.md.
• No knobs in run.json (every capture used the defaults).
• p95 separation for the remeasure plant is thin, and is reported as such.

🤖 Generated with Claude Code

https://claude.ai/code/session_017etd1nNb2E2XPL6WTvkAbo