[vanceingalls]

What was wrong

With capture-tier blur on, a <canvas> inside the composition paints without its layout offset. In k3-promo, #hcanvas has top: 200px, and its "K3. Now Open" text and glow landed exactly 200 px too high for frames 411-449. It happens on whole and chunked renders, and on frames where no mover is drawn, so it comes from the capture itself, not the blur trail.

Layout is correct: offsetTop still reads 200 during capture. Only the paint is wrong.

What Chrome does

Measured by editing the film, frames 411-449 against blur off (Chrome 156):

• Broken (mean 17.2 dB): top: 200px, a top: 200px wrapper around the canvas, overflow: visible on the clip, explicit size, will-change: transform.
• Correct (mean 47.8 dB): transform: translateY(200px), transform: translateZ(0), transform: translateY(0.001px), translate: 0 0.001px.

So the offset needs a real transform on the canvas, and a zero one is enough.

The fix

While captured, nested canvases get translate: 0 0 0 under :where(), so the rule has no specificity. It is a separate property from transform, so a composition's own transforms still apply, and any composition rule that sets translate wins.

Evidence

Unedited k3-promo, Subtle, compared with blur off:

• Frames 411-449: mean 17.2 dB before, 47.8 dB after, and the frames match by eye.

• Frames 450-470: 51.9 dB before, 57.0 dB after.

• Other frames: unchanged against blur off.

• With .nudge { translate: 30px 0 } on the canvas, it stays nudged while captured, and the unedited film is pixel-identical to the rule without :where().

The unit test only pins the rule. A jsdom test cannot see the paint.

Limits

• translate makes the canvas a stacking context, so a non-positioned canvas could now paint over a later sibling. Only while captured.
• <video> was not checked for the same loss.

🤖 Generated with Claude Code

[heygengenesis]

Verdict: REQUEST CHANGES · 1 blocker · 0 important · 0 nits · b693e0c

The fix works, but the rule's specificity overrides a composition's own stylesheet translate on a canvas, which the PR says it won't do. ↓ inline

Scope

Read captureMotionBlur/index.ts end-to-end around wrap/unwrap, discover, findCaptureDefect and the backface check. Chrome serializes translate: 0 0 0 as 0px (probed on 152), so the 3D checks (individualIs3d, individualTurnsIn3d) do not misread nested canvases as 3D. CI was still running when this was posted; no job had failed yet.

— Genesis

[github-actions]

Edit accuracy: accurate 271 (base branch 271), smooth 111 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 (0)

Unstable (1)

• seqplainkeys-none-center-r0-nested-z100: 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 / tracking 0.06, pressJump 0, drop 0, reload 0, render 0.03, renderKey -, undo true, teleport true

[heygengenesis]

Verdict: APPROVE · 1 resolved, 0 open, 0 new · b693e0c → c0468ee