fix(engine): fail-fast the sub-composition timeline wait when a script 404s

[vanceingalls]

Why

pollSubCompositionTimelines waits for every [data-composition-id] host to register window.__timelines[id]. When the script carrying that registration fails to load (404 / request failure), the registration can never arrive — but the poll still burned the full playerReadyTimeout (45 seconds), then warned and shipped a silently animation-less render. The engine was logging the exact load failure and waiting anyway.

Wild scale (30 days, local renders, n=89,786): the capture-setup histogram decays smoothly — 402/503/364/282/191 per 5s bucket — then spikes to 705 at the 45s bucket. A textbook timeout wall: ~1,000 renders/month across 402 distinct users, ≈15 user-hours of pure waiting, each render also delivering broken animations with only a console warn.

Found while benchmarking comp 0768f038: its "45s init" was 99% this one poll phase (139ms → 45,203ms) — its animations.js was unreachable.

What

• Sessions record failed script resources (requestfailed + HTTP ≥ 400 responses with resourceType === "script" — extending the diagnostic listeners that already existed) in session.scriptLoadFailures.
• The poll fail-fasts: once any script has failed, the wait cuts to a 2s grace with a loud warning naming the URL(s). Late-registering fetch-async comps are untouched — no script failure means the full 45s still applies, and a registration landing inside the grace window still wins (unit-tested).
• Outcome telemetry: sub_timeline_wait ("ready" | "timeout" | "script_failure") on render_complete, threaded session → CapturePerfSummary → RenderPerfSummary (worst across sessions) → CLI. The wild rate becomes a direct filter instead of setup-histogram forensics.

Validation

• Discovery comp e2e: total render ~72s → 23.1s — poll cut at 2.1s, warning names the script, render completes; healthy comp reports ready. Telemetry smoke: subTimelineWait: "script_failure" / "ready" respectively.
• Canary suite 7/7, PSNRs identical to baseline.
• 4 new unit tests on the poll contract (ready / fail-fast / full-timeout-when-healthy / late-registration-inside-grace).
• Engine suite: 907 passed; the 14 failures are pre-existing on main at v0.7.42 (18 fail on a clean checkout — stash A/B verified; parseVideoElements/color-grading, untouched by this diff).
• tsc/oxlint/oxfmt clean. The now-orphaned pollPageExpression helper was removed.

Corpus note

258/1,762 corpus comps (14%) reference local scripts missing from the corpus fetch — historical eval init timings on those comps measured this timeout, not the engine. Capture-stage ratios remain valid (both paths paid it equally).

🤖 Generated with Claude Code