[vanceingalls]

What

Tests only. Gives us a way to see whether capture motion blur draws the right picture, so we can decide on shipping it default-on. Five bugs reached users that the old tests could not see (worker-start state #6967, nested canvas offset #6975, WebGL freeze #7060, content-box cut and SVG clip path #7080, scroll offset on first wrap). The old tests ask "does blur happen"; these ask "is the picture right".

Two pixel invariants, both against the same composition rendered with blur off:

1. Outside the trail, blur-on equals blur-off. Trail mask from blur-off neighbour frames (radius 2 frames), 16 px tile MAE under 1.0. Measured noise 0.0078.
2. Inside the trail, blur-on is close to the mean of 8 blur-off sub-frame samples across the shutter (renderSeek(t, { subFrameDivisions: 24 })). Threshold 36.97 (noise 10.99; smallest catching mutation 124.3).

Plus 14 fixtures, a preview liveness test, and a multi-worker render differential through the producer.

Fixtures, the bug each guards, and mutation proof

• nested-canvas-offset: fix(core): a canvas in a blurred composition stays where it was laid out #6975. Rule removed: 4376 differing px (max 239) in the canvas region; with the rule, 0. Chrome 153 and 156 identical.
• callbacks: fix(core): a motion blur worker starting mid-film keeps the right state #6967 rewind. rewindThroughEveryCall() removed: invariant 1 worst tile MAE 106.10 (engine); through the producer with 7 workers 8 of 8 mutated runs fail (97.02), 0 of 6 clean runs.
• overflow-wordmark, svg-clip-logo: fix(core): keep overflowing content and SVG clip paths in motion-blurred frames #7080 class. Content-box cut mutation 130.13, clip-path mutation 124.32 on invariant 2.
• scroll-list: scroll offset lost on a fresh first wrap (pinned, see below).
• webgl-canvas, iframe-in-root, custom-element-in-root, flip-card-3d: cases capture cannot take, which blur through the filter tier.
• held-card, css-keyframes, stateful-callback, dpr2-mover, portrait-mover: baselines for holds, CSS motion, ripples-style stateful callbacks, device pixel ratio and aspect.

Known bugs on main, pinned by symptom

Tests named KNOWN BUG on main: ... keep the suite green today and flip when the fix lands. A pin requires every bad tile inside the named region and the worst MAE within [0.5x, 2x] of the measured value, so a second defect or a changed symptom fails it.

• scroll-list loses its scroll offset on a fresh worker's first wrapped frame (fix drafted in perf(core): a render frame with no mover to blur skips capture blur #7038, to be split out).
• callbacks shows its label B before the call that sets it (events="on").
• preview shows a blank frame after play() on the capture tier.

The producer scroll-list case only bounds the race (zero bad tiles in frames 2-27, any later bad tile inside #list); the engine pins carry the deterministic "bug fixed" flip.

What this does not catch

• Blur that is missing or too weak: invariant 2's max-tile statistic passes a wrong shutter angle; the fixtures test's differs check and the existing pixels tests cover absence.
• A DOM change within 2 frames of a checkpoint (masked); checkpoints sit 3 frames from changes.
• Performance (these check pixels; perf(core): a render frame with no mover to blur skips capture blur #7038's lack of speedup would not show), Windows, Safari, Studio-embedded player, <video>.

Verification

Pins and thresholds were measured on macOS arm64 (Chrome 153 and 156) and then confirmed on Linux CI: three green runs of Test (runtime) and Producer: integration tests with durations within noise of main (see the comment below). Local: unit 16, differential 73, fixtures + liveness 44, producer differential 18.

Review notes: vault page section. Next: adopt Rames Jusso's 60-composition sweep (defects A-O) as fixtures.

🤖 Generated with Claude Code

[heygengenesis]

Verdict: REQUEST CHANGES · 1 blocker · 1 important · 1 nit · head ab76d6d

The first Linux run is red in both lanes, and every failure is held DOM text on a capture-tier fixture. Details are inline on thresholds.ts. The other two are timing and dead-field issues.

CI evidence and durations (run 41711675131)

• Test (runtime): 9 of 73 differential cases fail: callbacks, stateful-callback and overflow-wordmark outside cases at every checkpoint. Worst tiles have MAE 1.8–10 on the text rows. The pixels (18), fixtures (29) and liveness (15) files pass, and supported was true, so the suite really ran. Job time is 8m08s against a 10-minute cap; recent main runs take about 6.5–7 min. The differential file adds 48.6 s, liveness 13.5 s and fixtures 8.4 s.
• Producer: integration tests: 4m38s against a 20-minute cap. Three blur cases fail with the same signature: overflow-wordmark held frames 13–17 (MAE 2.2–3.2), callbacks frames 18–22 on the label (4.2–6.0) and scroll-list frames 2–6 in the list (2.2–7.2). The blur off twice noise cases pass, so this is not render nondeterminism. The 4th failure (coreRuntimeBrowser.test.ts memory, 4695 vs 800) is in a file this PR doesn't touch; I haven't checked whether it's flaky on main.
• Shutter centring in shutterTimes matches frames.ts:125 (t - shutter/2).

Audited: all 11 files.

— Genesis

[github-actions]

Edit accuracy: accurate 267 (base branch 267), smooth 109 of those

The gate passes.
Smoothness is reported in the artifact, not gated. A case fails only if it fails 2 of 3 runs.

Quarantined, measured but not gated (8)

• seqplainkeys-none-center-r0-nested-z100 (fixed by studio: undo/redo on a keyframed plain element sometimes leaves the box 230-400px off (edit-accuracy seqplainkeys-none-*) #6989): fail / pass / pass, passes
• seqplainkeys-none-center-r0-root-z100 (fixed by studio: undo/redo on a keyframed plain element sometimes leaves the box 230-400px off (edit-accuracy seqplainkeys-none-*) #6989): fail / pass / pass, passes
• seqplainkeys-none-center-r30-root-z100 (fixed by studio: undo/redo on a keyframed plain element sometimes leaves the box 230-400px off (edit-accuracy seqplainkeys-none-*) #6989): not run
• seqplainkeys-none-pct-r0-root-z100 (fixed by studio: undo/redo on a keyframed plain element sometimes leaves the box 230-400px off (edit-accuracy seqplainkeys-none-*) #6989): fail / pass / pass, passes
• seqplainkeys-none-pct-r30-root-z100 (fixed by studio: undo/redo on a keyframed plain element sometimes leaves the box 230-400px off (edit-accuracy seqplainkeys-none-*) #6989): not run
• seqplainkeys-none-px-r0-nested-z100 (fixed by studio: undo/redo on a keyframed plain element sometimes leaves the box 230-400px off (edit-accuracy seqplainkeys-none-*) #6989): fail / pass / pass, passes
• seqplainkeys-none-px-r30-nested-z100 (fixed by studio: undo/redo on a keyframed plain element sometimes leaves the box 230-400px off (edit-accuracy seqplainkeys-none-*) #6989): not run
• seqplainkeys-none-px-r30-root-z100 (fixed by studio: undo/redo on a keyframed plain element sometimes leaves the box 230-400px off (edit-accuracy seqplainkeys-none-*) #6989): not run

Unstable (2)

• seqrepeat-none-px-r0-root-z100: tracking 0.05, pressJump 0, drop 0, reload 0, render 0.03, renderKey -, undo true, teleport true / tracking 0.05, pressJump 0, drop 0, reload 0, render 0.03, renderKey -, undo true, teleport true / tracking 0.05, pressJump 0, drop 0, reload 0, render 0.03, renderKey -, undo true, teleport true
• seqplainkeys-none-px-r0-root-z100: tracking 0.06, pressJump 0, drop 0, reload 0, render 0.03, renderKey -, undo false, teleport true / tracking 0.06, pressJump 0, drop 0, reload 0, render 0.03, renderKey -, undo true, teleport true / tracking 0.06, pressJump 0, drop 0, reload 0, render 0.03, renderKey -, undo true, teleport true

[vanceingalls]

Linux CI results for this PR.

Three green runs of both lanes (Chrome 153 pin, ubuntu-latest):

• Test (runtime): 6m44s, 7m54s, 5m40s (main: 6m24s)
• Producer: integration tests: 5m28s, 5m51s, 6m19s (main: 5m57s)

What the first Linux run found, all fixed on this branch:

• The runner draws DOM text with subpixel (LCD) fringes and the capture canvas draws it gray, so every text tile differed by MAE 2–10 on Linux and Windows while macOS measured 0.008. Printed tiles from the runner showed chroma 131–145 on the blur-off arm and 10–26 on the capture arm. That is a host setting, not blur, so both arms now draw grayscale text (--disable-lcd-text in the engine launch; a fontconfig rgba=none override for the producer render). Thresholds are unchanged.
• The blur test helpers were shipping in the engine tarball and pushed it over its size budget; they are now excluded in files.
• The 7-Chrome render in the producer lane skewed the Chrome node-count assertions in coreRuntimeBrowser.test.ts (failed 2 of 3 early runs), so that file now runs alone after the rest, and the reachability runner hash is re-pinned.

Not from this PR: Test (support) (and the aggregate Test) fail on current main content too: scripts/check-core-skills.test.mjs expects "ask once for that total" but the Reverse skill text from 341fdd3 says otherwise. Regression shard 10 failed once with an FFmpeg encode error in sub-composition-video on the first run and passed after the rebase.

Not covered yet, as stated in the description: the shutter-angle and weak-blur cases, performance, Windows rendering, and Rames's 60 compositions.

[vanceingalls]

All three addressed (head 3438bb1/9be81b2, after the earlier commits):

• blocker (Linux text): it was host LCD text. A tile dump from the runner showed chroma 131–145 on the blur-off arm and 10–26 on the capture arm, so the DOM draws subpixel fringes and the capture canvas draws gray. Both arms now draw grayscale text: --disable-lcd-text in the engine launch, a fontconfig rgba=none override in the producer render. Thresholds unchanged. Three green runs of Test (runtime) and Producer: integration tests (5m40–7m54 and 5m28–6m19; main 6m24 and 5m57). The coreRuntimeBrowser failure was real interference: it failed 2 of 3 early runs, so the blur render now runs alone after the other files.
• important (liveness pin): replaced the rAF sampler with a MutationObserver that reads the output canvas the moment the root is wrapped. It detects the blank on all 10 capture fixtures in 8 of 8 local runs. Guards: unreadable == 0, wrapped == 10, blank > 0. It pins the structural state (output empty at the wrap), not a presented frame count, which the trace showed is vsync-dependent.
• nit (blurs): field and no-op filter removed; the PR description no longer claims a blocked-fixture path.

Also pushed: the blur helpers are out of the engine tarball (size budget), and the reachability runner hash is re-pinned. Test (support) is red on main's own skill text (check-core-skills.test.mjs vs 341fdd3), not this PR.

[heygengenesis]

Verdict: APPROVE · 0 blocker · 1 important · 0 nits · 3 resolved, 0 open, 1 new · ab76d6d → 9be81b2

— Genesis

[heygengenesis]

important: Flake gate hits its 30-minute job limit and gets cancelled on every head of this PR (5 of 5, most recently run 41738536260). It re-runs each of the 6 changed files 10 times. The 5 engine files take about 24 minutes (the differential file alone about 14), so this file never gets its turn, and its multi-worker render is never checked under CPU contention. Any later PR that touches these files will hit the same limit. The gate isn't part of the required Test aggregate, so this doesn't block merge. Either fit the PR inside the gate's budget (fewer or cheaper browser cases) or say in the description that this file isn't flake-gated.

[heygengenesis]

Verdict: APPROVE · 0 blocker · 0 important · 0 nits · 0 resolved, 0 open, 0 new · 9be81b2 → d50d6c7

Delta

git range-diff shows all 17 prior commits patch-identical after the rebase. The one new commit raises the Flake gate's timeout-minutes from 30 to 60 (.github/workflows/ci.yml). That matches runsFor in scripts/ci/flake-gate.mjs, which budgets by file count (10 runs each for ≤6 files), with each run capped at 120 s. The gate passed at this head in 34 min. CI is fully green. The Studio timeline viewport gate failed on its first attempt with a thumbnailRenders perf-ratchet miss in code this PR doesn't touch, and passed on rerun.

— Genesis