fix(lint): drop false media_in_subcomposition rule

[miguel-heygen]

What

Removes the media_in_subcomposition lint rule, which errored on every <video> / <audio> placed inside a sub-composition, and corrects the docs and pre-render guards that repeated the same false claim.

Why

The rule asserted that nested media is "never seeked/decoded and renders blank/black". That is not true of the runtime.

Media discovery is flat: the runtime collects elements with a single document.querySelectorAll("video, audio"), resolves each element's host composition via closest("[data-composition-id]"), and rebases its local data-start by the accumulated absolute start of every ancestor composition (packages/core/src/runtime/media.ts, startResolver.ts). Nesting depth never enters the lookup, so media seeks and decodes correctly at any depth.

The practical cost was worse than a noisy warning: the rule was an error, so authors were pushed to hoist media out of the sub-composition it belongs to, or to silence a check that was right about nothing.

How

• Delete the rule from packages/lint/src/rules/media.ts and flip its test to assert nested media is not flagged.
• Drop the now-dead media_in_subcomposition clause from the registry components test.
• Drop the equivalent pre-render guard from the faceless-explainer and pr-to-video assemble-index.mjs scripts, which enforced the same rule outside the linter.
• Correct the reference docs that documented the restriction: hyperframes-core SKILL, data-attributes, variables-and-media, composition-patterns, and hyperframes-cli lint-validate-inspect.

The one real constraint is preserved in the docs rather than dropped with the rule: a sub-composition timeline cannot reach host-root elements, so host-root media motion is still authored on the main timeline. That is a timeline-scoping rule, not a media-nesting rule, and it is what the deleted rule appears to have been a garbled version of.

Net: 49 insertions, 89 deletions.

Test plan

• Unit tests added/updated

• Manual testing performed

• Documentation updated (if applicable)

• packages/lint/src/rules/media.test.ts (21 passed): the former positive case now asserts nested <video> / <audio> produce no diagnostic.

• packages/cli/src/registry/registryComponents.test.ts (2 passed).

• Traced the runtime path end to end through the producer render to confirm nested media resolves its absolute start and decodes.

Not covered: no new test asserts the sub-composition timeline scoping constraint, which was already documented behaviour and is unchanged by this PR.

[jrusso1020]

Additive review at 47a6a722, concurring with @magi. I verified the central claim at source rather than restating Magi's cites.

The rule's premise is false — removal is correct. The runtime drives media at any nesting depth: flat discovery (runtime/media.ts:77, runtime/init.ts:1702 — querySelectorAll("video, audio"), not scoped to root), per-element host resolution (init.ts:724/1858/2927, startResolver.ts:120 — closest("[data-composition-id]")), and local data-start rebased by the ancestor composition offset (init.ts:569-578: "data-start is authored relative to the media element's OWN sub-composition"). That's exactly the machinery for seeking nested media at global time. Matches Magi's sub-composition-video regression fixture rendering with visual parity + 0.956 audio correlation, and @miguel's hands-on verification.

The correction is complete — the false claim is removed from every site, not just the lint rule (Rule 2). I grepped the head: no residual never seeks / renders blank / DIRECT child of host claim about sub-comp media remains in non-test source. Specifically purged/corrected:

• packages/lint/src/rules/media.ts — the media_in_subcomposition closure removed; media.test.ts flips to a positive assertion ("does not flag <video> inside a sub-composition (runtime drives nested media)").
• skills/{faceless-explainer,pr-to-video}/scripts/assemble-index.mjs — the pre-render guard ② (media-in-subcomp HARD FAIL) removed + renumbered, with a comment now citing runtime/{media,startResolver}.ts. These are the "pre-render guards" the description refers to.
• packages/cli/src/registry/registryComponents.test.ts — drops media_in_subcomposition from the invalid-media allowlist.
• skills/hyperframes-core/references/variables-and-media.md now states the correct behavior and preserves the one real constraint: a sub-composition timeline cannot target host-root elements (a timeline-selector scope rule), so host-root media motion belongs on the main timeline at global time. That's the kernel the old rule garbled — good that it's kept.

One thing gates merge — CI (Rule 5): Skills: manifest in sync is red at this head. The PR bumps four skill hashes in skills-manifest.json, but CI's regen disagrees — the committed manifest is out of sync; re-run the manifest generator and commit the result before merge. Separately, CLI: npx shim and Producer: … tests show failures but at 0s with no producer/CLI-runtime code in this diff, so they read as infra/setup rather than this change — worth a re-run to confirm.

Verdict: COMMENT — the fix is correct, verified at source, and complete across all sites; it's ready on the merits and only gated on regenerating the skills manifest (and confirming the two 0s failures are infra). Happy to approve once the manifest check is green.

— Jerrai