Wave 21 — durable music and localization verticals (partial)

[alpesdigital-adm]

Wave 21: F5.001–F5.005 partial verticals

New/API-first music analysis and montage, canonical localization profiles/variants, bounded translation preflight/provider worker, and subtitles-only media approval through the real snapshot/proxy worker. No legacy runtime, main merge, deploy, VPS, paid provider call, or TODO checkbox completion.

Root-verified local evidence

• Unit suite: 2235/2235; typecheck, architecture, ESLint, domain language, DB/API/parity/infra/platform checks and production build pass; both audits report zero vulnerabilities.
• PostgreSQL 16 disposable bootstrap: all 211 migrations from zero; structural validator 282 tables / 1302 indexes / 966 FKs.
• Real Chrome against production Next: human login, locale profile and variant creation via HTTP, invalid locale rejection and fail-closed provider preflight.
• Concrete localization worker -> persisted snapshot -> normal proxy materializer/FFmpeg/storage/critic -> approval: 100 frames at 25fps, 960x540, AAC, 4 seconds, manifest/file SHA equality, visible subtitle ink and decoded original-audio correlation. Root inspected exported cue PNG. Initial source/alignment setup is arranged, not an ingest/paid-translation proof.
• Positive PostgreSQL test rejects valid blocked critic, tampered review hash, stale variant head, actor/path replay mismatch; equivalent review retry retains first observation and changed render identity fails.
• PostgreSQL preflight atomic consumption, music authenticated worker/replay/rights rotation, atomic montage publication; real PCM/music MP4 suite 12/12.
• Supervised local DB shutdown: zero client connections and zero remaining workspaces in final test DB. All subordinate agents completed.

Compatibility / limitations

Published 349 capabilities and 609 schemas were deep-compared unchanged; reviewed additions yield 369 capabilities / 638 schemas. UI parity report regenerated.
Only subtitles-only on a single identity clip at rate=1 is supported. Translated word placements are proportional within measured source blocks, not measured translated speech alignment. Local voice, uploaded localized audio, TTS/lip-sync/avatar, edited timeline mapping and full F5 acceptance remain open. Musical downbeats/sections are not claimed as measured downbeat detection. Local content-addressed storage used, not MinIO, for new local media proof.

A prolonged environment interruption invalidated one regression run; music suites were repeated successfully. An attempted historical prisma-proxy-review.integration.mjs fails before its assertions because its old fixture omits canonical operation progress; it is NOT counted green. The new real localization PG test carries the relevant retry regression instead.

Two commits: 2d31497 implementation/tests/CI; d116817 evidence and partial status. This is a draft for independent review and CI, not a production delivery claim. Audited progress remains 380/1259 (30.18%).