Chained sub-composition slots: id-ref data-start ignored in render media scheduling

[valeriangalliat]

Describe the bug

Documented relative data-start on composition hosts (data-composition-src slots that chain with data-start="hook") is ignored during render media scheduling.

After sub-compositions are inlined, collectRenderMedia resolves each host's timeline window with parseFloat(data-start). Non-numeric id-refs (hook, performance, intro + 0.5) become 0, so every scene's embedded <video> is scheduled at global 0–2s instead of its resolved window (e.g. 2–4s, 4–6s).

Symptom: multi-scene exports show the first scene, then go black for the remainder. The same project with numeric slot starts (0, 2, 4) renders correctly — proving id-ref resolution is the variable.

This is the composition-host slot path (renderMediaCollector / resolveHostWindow). Open PR #3252 fixes compileTimingAttrs for sibling <video> elements but does not cover host slot scheduling.

Link to reproduction

https://github.com/ArcadeHQ/hyperframes-repros/tree/patch/chained-scene-slots

Steps to reproduce

1. git clone https://github.com/ArcadeHQ/hyperframes-repros && cd hyperframes-repros && git checkout patch/chained-scene-slots
2. npx hyperframes@0.8.4 render --quality draft --workers 1 --output renders/out.mp4
3. Sample frames: ffmpeg -ss 1 -i renders/out.mp4 -frames:v 1 renders/at-1.png (red), -ss 3 (black), -ss 5 (black)
4. Control: cp index-numeric.html index.html, re-render — red at 1s, blue at 3s, green at 5s

Expected behavior

Each scene's video is extracted and frame-injected at its resolved global window (0–2s red, 2–4s blue, 4–6s green).

Actual behavior

Only the first scene's footage is scheduled. The MP4 is red for ~2s, then black through the end.

Environment

✓ Version          0.8.4 (latest)
✓ Node.js          v24.19.0 (darwin arm64)
✓ FFmpeg           ffmpeg 8.1.2 at /opt/homebrew/bin/ffmpeg
✓ FFprobe          ffprobe 8.1.2 at /opt/homebrew/bin/ffprobe
✓ Chrome           cache: .../chrome-headless-shell/mac_arm-152.0.7977.42/...

Additional context

Related open PR: #3252 (timing compiler id-ref fix — different code path).

A fix likely belongs alongside id-ref-aware host timing (similar to resolving data-start references against sibling slots / data-composition-id targets in the inlined document).