fix(h3): no desfasar el soundtrack cuando el plan empieza a mitad de canción

[cursor]

El concat final de _run_minimax_h3_story_video (H3 Legacy / proyectos H3 sin registry) remuxeaba la canción o el diálogo desde 0:00 aunque el plan saltara la intro.

Los slices de drive por plano ya partían de audio_origin_sec (_audio_timeline_start). El MP4 final no pasaba ese offset a concatenate_multi_clip_videos. El rejoin y el Director nativo sí lo hacen.

Disparador: videoclip o short_film_audio con MiniMax H3 Legacy (o lote Quick Video, que usa ese modelo) cuyo primer plano empieza después de 0 s — intro instrumental, entrada a mitad de tema. Los labios se generan al verso; el montaje suena el instrumental y -shortest recorta el final.

Fix: pasar audio_start_sec=audio_origin_sec en el concat final.

Tests: pytest tests/test_director_minimax_h3.py::test_h3_legacy_assembly_offsets_source_audio_to_plan_start — passed. También los tests de concat H3 existentes y test_rejoin_offsets_source_audio_to_first_planned_time.

No reabre #16, #17 ni #18.

  

---

Note

Medium Risk
Touches final A/V mux timing for H3 multiclip outputs; wrong offset would still misalign music and dialogue, but the fix aligns assembly with existing per-clip slicing and other pipeline paths.

Overview
Fixes soundtrack desync when MiniMax H3 Legacy (or registry-less H3) assembles a music video or short_film_audio project whose first shot starts after 0 s (e.g. skipping an instrumental intro).

Per-clip generation already drove from audio_origin_sec via _audio_timeline_start, but the final concatenate_multi_clip_videos call in _run_minimax_h3_story_video remuxed the full track from 0:00—so lip-sync matched the verse while the output still played the intro, and -shortest could trim the end. The change passes audio_start_sec=audio_origin_sec on that final concat, matching rejoin and native Director behavior.

Tests update FakeWgp.concatenate_multi_clip_videos stubs for the new argument and add test_h3_legacy_assembly_offsets_source_audio_to_plan_start (plan starting at 12.5 s asserts assembly uses the same offset).

^{Reviewed by Cursor Bugbot for commit a47cd57. Configure here.}

[github-actions]

PR Review — Loreframe Studio

Risk: low
Scope: 3 file(s); +96/-6; backend services

Automated review from scripts/analyze_pr.py. This is a heuristic pass (no LLM) so humans still own the merge decision.

Findings

• No heuristic issues. Still run the CI checklist below.

Changed files

• modified: app/services/director_pipeline.py, tests/test_director_h3_identity_continuity.py, tests/test_director_minimax_h3.py

CONTRIBUTING checklist

• python scripts/verify_clean_repo.py
• python -m compileall -q app/services app/launch.py scripts
• cd ui && npm run build if the UI changed
• No weights, CivitAI sidecars, or generated guides
• Stays local-first (no required accounts / telemetry)

Posted by the repo PR review workflow. Re-runs on each push to the PR.

[cursor]

Risk: medium. No pude aprobar: GitHub bloquea la autoaprobación de este PR. Se asignó revisor para revisión humana.

  

_{Sent by Cursor Approval Agent: Pull Request Router and Approver}