fix(producer): export places a scene's media where preview plays it

[miguel-heygen]

What changes for a user

Exported video now places a sub-composition's media where Studio preview plays it:

• A split or head-trimmed scene (a host with data-playback-start) plays its video and audio from its in-point. Since fix(core): a split sub-composition plays its videos from its in-point in preview and timeline #5413, preview did this but export still played from 0, so a split scene exported seconds behind what was previewed.
• A scene written inline in index.html (a data-composition-id host with no separate scene file) that starts after 0 plays its media at its start. Export placed it at 0.
• A looping video inside a split scene stays inside its scene's time slot.

Projects whose scenes are separate files with no in-point export exactly as before.

Why

packages/producer/src/services/renderMediaCollector.ts is the one place every exported video, audio and image gets its window on the root timeline; frame extraction and the audio mix both read it. It now follows the rules preview uses, through core's helpers (compositionOriginSeconds, hostInPointSeconds, cutToHostSlots):

• every composition host counts ([data-composition-id], [data-composition-src], [data-composition-file], the set core's start resolver uses), not only inlined scene files, so an inline scene's start is no longer ignored; the root composition's start counts but its end no longer cuts media;
• each host's offset is its start minus its in-point; a clip is cut to its in-point hosts' slots, and a clip they cut away entirely is dropped;
• audio and video that start inside a cut begin at the slot's start, that much further into their source (rate curves shifted to match); a loop moves to its slot's start and keeps its phase;
• media is matched by the id the parsers report (render id, else author id) for video, audio and img[src] only, and an author id never takes over a render id's window.

Checked

Rendered on Linux through the producer (createRenderJob + executeRenderJob), main vs this branch. Fixtures use a 19 s video with a frame counter and an audio track that plays one tone per source second.

fixture	main	this branch
scene data-start=0, in-point 7	audio src 0..11, picture src 0..11	audio src 7..18, picture src 7..18 (SSIM 0.875-0.877 vs source t+7)
inline scene data-start=6, no in-point	audio src 0..11 from 0 s	silent 0-5 s, then src 0..5 at 6-11 s
looping clip in a scene at 6 s with in-point 6	loop visible from 0 s	black until 6 s, then the loop
no sub-composition		decoded video and audio frame hashes identical to main

• renderMediaCollector.test.ts: 18 pass, three runs in a row. The new cases fail against main or the earlier head with the values they got (for example an inline scene's video at 0..19, a looping video starting at -1, a scene video moved to 0..12 by a root image sharing its id). Two compile the real document through compileForRender; one checks windows against main's for a document main got right.
• Producer unit lanes (bun and vitest: 874 pass), producer typecheck, oxlint, oxfmt, fallow and the comment ratchet are clean.
• 73 producer regression fixtures compiled on main and on this branch give identical media windows; none has an inline scene starting after 0.

Not covered here

• An image inside a trimmed scene keeps its uncut start, and an image with no id inside an inline scene is not shifted.
• A looping video keeps its source phase from its original start rather than the cut point.
• When a render has to probe media in the browser, the probe stage resets a clip's own source offset to its authored value; a trimmed clip inside a trimmed scene could lose the shift there. Not exercised.
• A clip's fade-in is applied at the cut point, not at its own start.
• Not run: the Docker regression harness against committed baselines, the hyperframes render CLI path end to end (it calls the producer API above), Studio snapshot, macOS.

[github-actions]

Edit accuracy: accurate 2061 (base branch 2061), smooth 1583 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)