[vanceingalls]

What

data-motion-blur-tier="auto" now draws with the filter tier. Capture is opt-in with data-motion-blur-tier="capture" and still runs wherever html-in-canvas exists.

Why: Rames Jusso's sweep of 60 one-pattern compositions found 15 defects, and the worst ones are in the capture tier by design: anything in front of a mover is painted over (A), blend modes are ignored (B), a selector through the root's parent stops matching (H), and the first preview frame after play() is black (N). auto used to pick capture for any rotation, scale or left/top motion, which is what Desktop preview runs. The filter tier has none of those four.

Before

k3-promo frame 274 on main, auto drawing with capture: the swinging pendulum rods and the text carry streaks.

After

The same frame on this branch, auto drawing with the filter tier: the rods soften, the text blurs gently, and nothing in front of a mover is painted over (a scanline overlay in front of the bars at frame 178 survives here and is lost on main).

What it costs: please read before approving

• Rotation and zoom now soften instead of streaking. On k3-promo frame 274, main (capture) streaks the swinging pendulum rods and the "Every modality" text; this branch softens the rods and blurs the text gently. On the same film at frame 178 a scanline overlay in front of the bars survives on this branch and is lost on main, and on spacex-launch frame 71 a cursor missing on main is present here.
• The filter tier is not cheap on mover-heavy films. Render CPU seconds (user+sys of the render's process tree, 4 workers, software GPU, Chrome 153, Subtle, arms interleaved, three runs on spacex-launch, one on k3-promo; wall times were taken at a load average of 30 to 250 and are not quoted): spacex-launch (77 movers) no blur 184-193, auto (filter) 547-552, forced capture 302-320, so filter is 2.9x no blur and 1.75x capture. k3-promo (106 movers) 117 no blur, 236 filter, 205 capture. "About what a render without blur costs" holds only for the four-whip bench the guide used to cite. The cause is not isolated (likely the 32 or more feOffset copies per mover per frame on the software GPU). The guide now says all of this.
• Known regression until the follow-up PR that fixes it (to be linked): a position: fixed child of a mover that moves via left jumps into the mover, because the SVG filter makes the mover its containing block; in the sweep 3495 px differ from the 32-sample oracle for the filter tier against 721 for capture, and before this change auto kept such movers on capture. The guide lists this and the filter tier's other measured limits (::before/::after outside the box cut off, a mover's own clip-path or mask gets hard edges, :last-child on the root's children also matches the injected svg, a visibility toggle inside the shutter drawn at full strength, backdrop-filter differing from the oracle).

Vance saw these numbers and the held-merge note and asked on 2026-11-06 for everything in this plan to be merged. Merge order: #7271 (callbacks mute), #7292 (the fixed-child fix, which removes the regression above), then this PR.

Changes

• tier.ts chooseTier: auto and filter draw with the filter tier when the fallback is on; capture draws with capture where it can run; renderer-off stays off. The report reason code for auto is tier-auto.
• Deleted the translation-only hand-off and translationOnly from discovery, with its tests.
• hyperframes check: tier-auto is a chosen tier, so no fallback info; motion_blur_unverified now only for forced capture in a browser without html-in-canvas.
• Docs, the skill line and the Studio panel label (Motion blur (Beta)).
• Tests: tier.test.ts, checkMotionBlur.test.ts, index.test.ts, a fixtures case proving auto draws a rotating mover with the filter tier and a page that asks for capture keeps it, and a filter-tier arm (20 cases) in its own file, captureMotionBlur.differential.filter.browser.test.ts. The differential is now three files (default 51 cases, inside 22, filter 20) because the 98-case single file timed out the Flake gate's 120 s per-run limit; each file now runs in at most 37 s locally under one CPU burner per core. The filter arm skips the four fixtures that already draw with the filter tier. Filter-arm inside noise is at most 4.9 against a third of insideMaxMae (12.3), so no thresholds changed. Filter-arm inside noise is at most 4.9 against a third of insideMaxMae (12.3), so no thresholds changed. Four pixels cases that tested capture's defect detection through auto now pass "capture".
• With only the chooseTier line reverted, the new tier.test.ts case, two core index.test.ts cases and the new fixtures case fail.

Not in this PR

• htmlCompiler.ts detectRenderModeHints still routes every blur render through screenshot capture and its message says "draws each frame with html-in-canvas". No longer accurate for the default; follow-up.
• Defects D, E, F, I, O on the filter tier, and the capture-tier set (A, B, C, H, J, K, L, M, N), stay pinned by the sweep tests.

🤖 Generated with Claude Code

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated
hyperframes	🟢 Ready	View Preview	Nov 6, 2026, 4:36 AM

💡 Tip: Enable Automations to automatically generate PRs for you.

[github-actions]

Edit accuracy: accurate 267 (base branch 267), smooth 110 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): 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): 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): 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): 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

[heygengenesis]

Verdict: APPROVE · 0 blocker · 0 important · 1 nit · 603ffd46

This approval assumes the stated merge order. #7292 has to land first, because it removes the fixed-child regression (F) that this PR otherwise ships. #7292, #7293 and this PR merge cleanly together, and the core blur suites pass on the combined tree (215/215).

• nit (PR body): "the follow-up PR that fixes it (to be linked)" should now read #7292. The "Filter-arm inside noise…" sentence appears twice. After the rebase onto #7271, once-guard-label is in every differential arm, so the counts are now 54/24/22, not 51/22/20. The Flake gate still passed each file 6/6.

Scope

• I read the whole diff.
• I checked that reverting the chooseTier line fails 3 tests (tier + 2 index).
• translation-only has no readers left.
• The defect reason codes that the fixture edits stopped asserting (iframe-in-root, accelerated-canvas, backface-visibility) are still covered in htmlInCanvasDefects.test.ts and in the pixels cases that now pin "capture".
• The sweep pins its tiers explicitly, so the new default does not change it.
• CI is all green on 603ffd46. The merge state is only waiting on review.
• For the htmlCompiler.ts:359 follow-up: every data-motion-blur="on" render still takes the screenshot route, and the no-blur arm did not. So some of the 2.9× may come from the route rather than the filter tier. That is worth separating when the follow-up is measured.