vocal2midi-rs-rewrite

name	vocal2midi-rs-rewrite
description	Coordinate one named Vocal2Midi Linux model shard under the py2rs model control plane. Use when selecting, re-cutting, routing, implementing, reviewing, or promoting Qwen3-ASR, Romaji ASR, HubertFA, GAME, or RMVPE work.

Vocal2Midi Model Rewrite Coordinator

Use rewrite-in-rust/models/control-plane.yaml as the model-only control plane. It owns shard order, granularity, review cadence, serial execution, and shared prerequisite state. It never replaces the five independent model manifests with a combined unit-status inventory.

Selection Gate

Map the request to exactly one manifest:

• Qwen3-ASR -> models/qwen3-asr/manifest.yaml
• Romaji ASR -> models/romaji-asr/manifest.yaml
• HubertFA/HFA -> models/hubert-fa/manifest.yaml
• GAME -> models/game/manifest.yaml
• RMVPE -> models/rmvpe/manifest.yaml

If no model is named, request one. Do not infer priority across models or read the closed backend rewrite-in-rust/manifest.yaml.

Required Context

Read, in order:

1. the model control plane;
2. the selected model manifest;
3. rewrite-in-rust/models/README.md, resources.md, notes.md, and the glossary;
4. only the selected manifest's named source, dependency, bootstrap, test, and historical evidence paths.

The historical records are evidence for decisions, not permission to reuse old live statuses. Existing model statuses were reset to planned; current owner remains legacy-python until a fresh review and promotion decision.

Discovery And Routing

Route to vocal2midi-rs-dep-bootstrap before writer work when a unit crosses an asset/model-format/native boundary, needs a crate or shared capability, lacks a tensor-interface or lifecycle smoke, mixes separately verifiable behavior, or has an unclear rollback route. Dependency discovery may re-cut only the selected model manifest.

Use standard capability-first dependency policy: direct crate coverage first, then a focused adapter, then narrow hand-written behavior when evidence proves that is smaller or safer. py2rs-crate-recon is required for agent-mode dependency research. Do not add Cargo dependencies during preference capture or create private copies of shared capabilities.

Evidence Policy

Raw model asset/import/execution units may use behavior_verification.mode: model_evidence with the explicit same-weights/equivalent-configuration comparison exception. They still require asset provenance, load errors, tensor names/ranks/shapes/dtypes, cancellation, cleanup, repeated-call behavior, routing, and rollback evidence.

Audio preprocessing, adapters, public payloads, errors, lifecycle, and production routing use behavior_verification.mode: legacy_seam, exact comparison, and a behavior_reviewer before later review or promotion. A Rust-only contract cannot replace a missing legacy seam.

State Loop

1. Confirm one selected model and one unit.
2. Check the shared dependency registry and dependency-recon status.
3. Define the seam, observations, verification commands, review roles, and rollback before implementation.
4. Send one accepted unit to the writer; keep it active only while work is underway.
5. Writer verification advances it to reimplemented, never verified.
6. Add only writer-verified units to the serial three-unit review batch.
7. Run legacy-seam R0 first where required, then the union of non-behavior roles.
8. Promote one unit only after its verdicts and rollback evidence pass.

The control plane starts with no open review batch. Flush a batch when it reaches three units, a shard completes, a high-precision boundary is reached, or promotion is requested.

Non-Negotiables

• One model, one unit, one writer at a time.
• Never read or modify the closed backend manifest for model work.
• Never treat historical evidence as current verification.
• Never let a writer review or promote its own work.
• Never promote while the shared foundation is blocked or rollback is unclear.
• Keep model assets ignored unless repository policy and licensing change.
• Use the project uv Python 3.12 environment for checks.

Completion Response

Report the selected model/manifest/unit, route taken, control-plane state changes, historical evidence used, checks, review roles remaining, blocker(s), and rollback route.