fix(core): resolve long nested media in local time

[miguel-heygen]

Summary

Long videos starting at local 0s inside delayed sub-compositions no longer disappear from late snapshots. Media starts before the host’s absolute window now inherit the host offset even when their long duration overlaps that window; legacy media authored with a start already inside the global host window keeps its existing behavior.

Fixes reported:1787898835.985789 and reported:1787908114.998239.

• https://slack.com/archives/C0BGC335AQY/p1787898835985789
• https://slack.com/archives/C0BGC335AQY/p1787908114998239

Test plan

• packages/core/src/runtime/init.test.ts (91 passed, including legacy global-PIP coverage)
• packages/cli/src/commands/snapshot.test.ts (33 passed)
• Core and CLI typechecks
• oxlint and oxfmt --check on changed files
• Exact Terra point fixture: 40s, 41s, 80.4s, and 111s snapshots are all non-black
• Exact Terra 18-frame sweep: every in-host sample from 42.353s through 112.941s is non-black; only pre-host and post-host samples are black

---

[vanceingalls]

Correctness read:

The fix drops authoredEnd-based overlap and disambiguates purely on authoredStart's position relative to the host window. That's the right call — a long local clip can end past the host's absolute window without changing what the start means, so the duration was a false signal. Three shape checks:

• Composition-local (host@39.233 + video@0/dur=80): authoredStart=0 < inheritedStart=39.233 → startsInsideHostWindow=false → returns inheritedStart + authoredStart = 39.233. Video visible through the host tail. Matches the added test at init.test.ts:1187.
• Legacy root-global PIP (host@45.4 + video@45.4): authoredStart >= inheritedStart and authoredStart < hostEnd → true → returns authoredStart=45.4. Legacy path preserved.
• hostDuration null/0: hostEnd=null → guard collapses to authoredStart >= inheritedStart, same as the pre-fix branch. No behavior change on unknown-duration hosts.

The retained inline comment on lines 663-669 explains why duration was unsound — first-principles, no ticket citation. Good comment.

CI note: Tests on windows-latest shows red, but the vitest report inside it is 196 files / 2852 tests passing with Exited with code 0; the wrapper exits 1 after an LFS-pointer checkout phase warning ("Encountered 68 files that should have been pointers"). Reads as runner LFS-state flake, not a regression from this change — and this check isn't in the required set (mergeable: MERGEABLE). Worth a rerun to green-up the record if you'd like.

Review by Via

[vanceingalls]

R3 re-stamp at d55f9f2 — test-only additive delta (single initSandboxRuntimeModular test covering long nested media staying visible across host schedule window at 11 checkpoints from 42.353s to 112.941s + resolver=39.233s). Same fix as approved at 4d9d5ba, no production code touched. All 30 required checks green.

— Via