Use acoustic reference boundaries and retain successful partial scene repairs

[SubSchool]

Root cause observed in the live repair

The model timestamp placed the reference speech end near 5 seconds, while the actual audio had already entered a long silent tail. Trimming solely from that annotation still left silence in the last second of the Veo reference.

Changes

• Inspect the actual audio for trailing silence and prefer that measured boundary when available.
• Version private anchors so cached pre-fix references are not reused.
• Rebuild final scene outputs from accepted scene checkpoints after a partially failed regeneration.
• Mark the scene that actually failed, not an earlier already accepted replacement.

Verification

• Acoustic regression uses a six-second video with four seconds of signal and an intentionally wrong model timestamp.
• E2E interrupts a multi-scene regeneration, resumes only the remaining scenes, and verifies the previously accepted replacement is used in the new render.
• Existing corrupt-anchor recovery and editorial checks pass.