[Release acceptance] Prove music quality Gate on the real provider

[Jah-yee]

Current scope

The RunPod deployment gap from the original issue is obsolete: scripts/deploy-music-serverless.ts, bun run deploy:music-serverless, the GPU image workflow, worker tests, and deployment documentation now exist. Do not reactivate the recovered 623-line deploy script.

The remaining valuable work is a current music-quality evaluation gate against today's music-engine contract.

Deliverable

• A compact, versioned fixture corpus that exercises melody preservation, contour, timing, style conditioning, output validity, and deterministic/mock behavior.
• Objective checks where possible, perceptual rubric and sampled artifacts where objective scores are insufficient.
• One local command and one scheduled/explicit CI job with retained reports.
• Model/version/seed recorded in every report.
• Threshold changes reviewed as product/music decisions, not silently updated to make CI green.

Non-goals

• Restoring deleted scripts verbatim.
• Treating BPM/key estimates alone as proof of musical quality.
• Running paid GPU evaluation on every ordinary PR.

Acceptance

A known-good release and a deliberately degraded fixture produce meaningfully different scores, reports are reproducible, and the operator can identify which musical dimension regressed.

rc.2 status

Merged integrated PR #440 and final release-tail PR #454 now supply the online technical Gate, versioned receipts, bounded retries, frozen golden cases, provider evidence artifact, and operator runbook. Keep this issue open only for real-provider product-level acceptance in #455/#450 and independent review in #456. Close it once the frozen known-good/degraded checks discriminate correctly, all three canary profiles pass, and the human melody/musicality decision is recorded; further model tuning belongs in a new post-release issue.

[Jah-yee]

rc.2 status update: PR #433 implements the online technical layer and part of this issue's repository deliverable: versioned generation/conditioning receipts, deterministic Web/Worker WAV quality gates, bounded retries, tests/music-quality-golden/cases.json, model/version/seed evidence, sanitized observability, and an operator runbook that keeps objective gating separate from musical evaluation. Candidate worker and contract suites are green.

Do not close this issue yet. The remaining acceptance is the product-level offline layer: run the frozen music corpus against the real production RunPod/provider, retain scored artifacts, include a deliberately degraded fixture, and record a human musicality/melody-preservation decision. That work is now part of release owner canary #450; after that evidence exists, reconcile it here and close only if the known-good/degraded discrimination is demonstrated.

[Jah-yee]

@moapacha owns execution of this acceptance: provision/run the real-provider canary, preserve sanitized model/input/output/Gate evidence, investigate degraded results, and prepare the smallest fix or configuration correction needed to make known-good output pass and known-bad output fail. Do not stop after producing samples; drive failures to an assigned remediation and rerun.

@koriyoshi2041 owns independent verification: listen to every required artifact, inspect Gate measurements and delivered-byte provenance, challenge false-positive/false-negative behavior, and issue an explicit acceptance or release-blocking finding. You may broaden inspection and prepare a focused fix, but a different maintainer must approve release-critical changes.

This issue is complete only when real-provider musicality, continuity, level, clarity, fragment/dropout rejection, bounded regeneration, persistence, and no-delivery-on-failed-Gate are evidenced for the exact release SHA.