fix(producer): bound nested media by a host's data-duration when it has no data-end

[miga-heygen]

What

The render-media planner now closes a composition host (data-composition-file element) at the same window the runtime uses to hide its descendants. resolveHostEnd in packages/producer/src/services/renderMediaCollector.ts delegates to the runtime's own resolveAuthoredTimingWindow from @hyperframes/core: a positive data-duration closes the host at start + duration; data-end applies only when data-duration is absent or non-positive, and only when it lies past the host start.

Why

The compiler only stamps data-end on media tags, so a host authored with just data-duration was unbounded for the media planner. Audio and video nested in a shortened scene kept playing over the next scene (#3964).

The first cut of this fix read data-end first and data-duration second, which is the reverse of the runtime's precedence (resolveDurationForElement in packages/core/src/runtime/startResolver.ts). It also treated data-duration="0" as a real duration, which collapsed the host window to its start and dropped every nested clip. Sharing the runtime's resolver instead of re-implementing the rule removes both divergences and the local parseNumeric copy (parseFloat vs the runtime's Number parsing).

How

• resolveHostEnd returns resolveAuthoredTimingWindow({ start, duration, end })?.end ?? null.
• resolveHostWindow doc block rewritten to describe the shared rule.
• Tests in renderMediaCollector.test.ts:
  • host with data-duration and no data-end bounds nested media (including an id-ref data-start);
  • conflicting data-start=2 data-duration=2 data-end=6 bounds nested media at 4 (duration wins);
  • data-duration="0" and "-1" are treated as absent: unbounded when there is no data-end, otherwise fall back to data-end; nested clips are not dropped;
  • data-end at or before the host start is ignored rather than collapsing the window.
• htmlCompiler.test.ts: the template-wrapped sub-composition test now expects the 4s clip to close with its 2s host (end: 4, was 6).

Verification

• bun test packages/producer/src/services/renderMediaCollector.test.ts — 7 pass.
• bun test packages/producer/src/services/htmlCompiler.test.ts — 135 pass.
• bun run --cwd packages/producer typecheck — clean; oxlint and oxfmt --check clean on the touched files.
• Mutation checks: reverting to end-first precedence fails the conflicting-attribute test; removing the non-positive guard fails both it.each cases; reading data-end raw fails the end-before-start test.

Closes #3964

Supersedes #3965 — carries valeriangalliat's fix verbatim with signed commits so it can merge under the repo's signed-commit ruleset; adds duration-first precedence to match the runtime and a non-positive-duration guard, with tests.

— Miga

[terencecho]

APPROVE at a1f32d44 — planner's host-window resolver now delegates to the runtime's own; divergence removed by construction

Clean fix. 3 files, +96/-14, one commit co-authored by valeriangalliat + miguel-heygen (bot-carrier repackage of #3965 with signed commits to satisfy the repo's signed-commit ruleset).

Head verified: a1f32d4414d2e340b5683359d87aa3f99db12b26.

The change is the right shape

resolveHostEnd in packages/producer/src/services/renderMediaCollector.ts:44-56 now returns resolveAuthoredTimingWindow({start, duration, end})?.end ?? null, imported from @hyperframes/core. That is the same resolver the runtime uses for resolveDurationForElement — so the planner's host-close window and the runtime's descendant-hide window cannot drift by construction. The previous local parseNumeric (Number.parseFloat) is deleted alongside the divergent read-order (data-end first) and the data-duration="0" mishandling (which collapsed the window to hostStart and dropped every nested clip).

The docstring update on resolveHostWindow (:72-81) correctly cites the shared resolver as the new source-of-truth for the host end, and the doc block on resolveHostEnd names all three fold rules: positive-duration wins over conflicting end, non-positive-duration falls back to end, and end-before-start is ignored.

Test coverage matches the fix rules exactly

Four new tests in renderMediaCollector.test.ts:20-90 pin every rule the fix relies on:

• closes a host authored with data-duration but no data-end — the primary scenario (#3964): audio in a 2s-slot-over-4s-scene closes at 2s not 4s; a video whose own window is entirely after 2s is dropped from the plan; an id-ref data-start="hook" chains correctly.
• prefers data-duration over a conflicting data-end, like the runtime — precedence test: data-start=2 data-duration=2 data-end=6 → close at 4 (duration wins).
• it.each(["0", "-1"]) — non-positive-duration fallback: unbounded when no data-end, else fall back to data-end; nested clips are NOT dropped.
• ignores a data-end at or before the host start — end-before-start is treated as absent; window stays unbounded rather than collapsing.

The amended existing test in the same file (preserves an explicitly marked legacy-global media window) correctly reflects the new behavior: open-ended audio tracks close with the host (start 2 + duration 6 → end 8) rather than staying unbounded. The amended htmlCompiler.test.ts:1229-1238 template-wrapped sub-composition test correctly expects the 4s clip to close at end: 4 with its 2s host (was end: 6, which was the unbounded-scene-file bug).

Mutation checks (per PR body, verified by test shape)

• Reverting to end-first precedence → the prefers data-duration over conflicting data-end test fails (asserts end: 4, mutation would give 6).
• Removing the non-positive-duration guard → both it.each cases fail (open-clip would collapse to hostStart instead of staying unbounded at end: 4; capped-clip would collapse instead of falling back to end: 3).
• Reading data-end raw without the start-past-end guard → the ignores a data-end at or before the host start test fails (window would collapse from unbounded to 1, dropping the clip that ends at 5).

Each mutation → distinct red test. Coverage is honest.

CI

At head: 43/43 checks green, zero pending, zero failing. reviewDecision: REVIEW_REQUIRED reflects that this stamp is the one that clears the gate.

Substance is Val's fix (valeriangalliat) tightened for runtime-precedence parity and non-positive-duration handling, co-authored with Miguel. Miga carries it as a signed-commit repackage of #3965. Ship it.

— Review by tai (pr-review)