[miguel-heygen]

What

A composition root with no data-start of its own now takes its start from the nearest composition host or timed wrapper above it, even when a neutral element (a <section>, a vfx <canvas>, an author's wrapper div) sits between them. A wrapper with its own data-start wins over the host above it, because that wrapper is what shows and hides the composition. Before, only the immediate parent counted, so a wrapped composition was timed on the root clock: it started at 0 instead of its host's start, and once root time passed its own duration its layers went hidden.

Why

#4334. With a host at 5 s and one neutral wrapper between it and the inner root, the resolver returned 0 for every host marker (data-composition-src, -id, -file); with no wrapper it returned 5.

Related work

Fixes #4334. Follows the maintainer's call on the issue: keep the resolver marker-only (no special case for canvas.hf-vfx-src); the exporter keeps copying data-start onto its wrapped element.

How

inheritedTimingHost in packages/core/src/runtime/startResolver.ts now walks to the nearest ancestor carrying a composition marker or a data-start (closest()) instead of checking only the parent. A timed wrapper counts because the runtime hides it outside its window: timing the composition from the host would start its clock early and show it mid-animation. CSS animations already pick their clock the same way (closest("[data-start]") in the CSS adapter). It stops at the first match, so an outer host or a sibling clock is never used, and explicit starts, references and fallbacks are untouched. isCompositionHost uses the same selector, so the marker list has one definition.

Test plan

• startResolver.test.ts: the full 3x3 matrix from the issue (each marker, 0/1/2 neutral wrappers) plus a nested-host case that must take the nearest host (3 s inside a host at 2 s = 5 s), plus a timed wrapper at 2 s inside a scene at 5 s that must give 7 s. 45 passed, 3 runs. With the old parent-only lookup restored, exactly the 7 new wrapper/nested cases fail; with a host-only closest(), the timed-wrapper case fails (expected 5 to be 7).
• Real render, wrappedMountTiming.integration.test.ts (Chrome + ffmpeg): a scene at 1 s holds, inside a <section>, a composition whose box fades in over its own first second. Halfway through the scene the frame is half grey (between 90 and 165 of 255), not fully faded in. 3 runs green; with the old lookup it fails (expected 255 to be less than 165).
• src/runtime in core: 104 files passed.
• oxfmt, both tsc --noEmit (core, producer), comment ratchet, producer test classification: clean.
• Comments follow CONTRIBUTING.md "Comments".

Why under 100 lines of source: the bug is one lookup in the one resolver every timing path goes through; the rest of the diff is the matrix and the render test.

[heygengenesis]

Verdict: REQUEST CHANGES · 1 blocker · 0 important · 0 nits · head db8ee7ab

• blocker — required Test reachability fails: Runner mapping needs review: packages/producer/scripts/test-classification.mjs. Editing that file changes its digest, and scripts/test-reachability.json:50 still pins the old one (88d7415a…). Re-pin it to 47a6204722443284067bf5c464601a252091aab3670ef1626e50ce507b8214fc, as #4610 did. With that one-line change, node scripts/check-test-reachability.mjs --report passes locally.

I found nothing wrong in the resolver change. A wrapped root now resolves the way a directly mounted one already did, and timeline.ts already used the nearest composition's start as its fallback.

Scope

Read all 4 files and the callers of resolveStartForElement, resolveHostSlotsForElement and resolveCompositionOrigin. Locally: startResolver.test.ts 44/44, and 7 fail with the old lookup restored. Core src/runtime passes 2557/2557. wrappedMountTiming.integration.test.ts passes with Chrome and ffmpeg. The other CI jobs were still running when I posted this.

[github-actions]

Edit accuracy: accurate 271 (base branch 271), smooth 188 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)

[miguel-heygen]

Re-pinned the classification digest in scripts/test-reachability.json (a66cd1f); node scripts/check-test-reachability.mjs passes, and fails with the old pin. a66cd1f also has 81a4848: a composition inside a wrapper with its own data-start now starts with that wrapper, not the host above it, since the wrapper is what shows and hides it (test: wrapper at 2 s in a scene at 5 s gives 7 s).

[heygengenesis]

Verdict: APPROVE — 1 resolved, 0 open, 0 new · db8ee7a → a66cd1f. The test-reachability.json digest is re-pinned, and I read the timed-wrapper change in 81a4848 and found no blocker. Locally: reachability passes, core src/runtime 2558/2558. Required CI is green.