vocal2midi-rs-unit-writer

name	vocal2midi-rs-unit-writer
description	Implement exactly one confirmed unit from one named Vocal2Midi model manifest as the writer role after dependency/bootstrap acceptance.

Vocal2Midi Model Unit Writer

Implement one confirmed unit in one selected model shard. Do not scan other model manifests, edit the closed backend manifest, or repair a blocked shared foundation by creating a private alternative.

Preconditions

Read the control plane, selected manifest, unit sources/tests/assets, dependency recon report, bootstrap note, shared dependency registry, and rollback route. Confirm:

• the unit exists and is the only active unit;
• dependency_recon is complete or an accepted policy outcome;
• the accepted seam and behavior-verification mode are concrete;
• model evidence covers asset/load/tensor/lifecycle requirements, or legacy seam evidence covers exact public behavior;
• canonical shared dependencies exist and are not blocked;
• no implementation path points to /tmp or an agent-private copy.

Workflow

1. Set the selected unit from planned to active only when implementation starts.
2. Keep changes inside the unit's accepted asset, model, adapter, lifecycle, or promotion boundary.
3. Add focused import/load, tensor-interface, lifecycle, or legacy-seam checks before changing production routing.
4. Reuse completed v2m-core capabilities and canonical shared dependencies; do not duplicate them in a model shard.
5. Preserve public payloads, structured errors, cancellation, cleanup, and rollback named by the manifest.
6. Run the unit's writer verification and the manifest validator.
7. Set the unit to reimplemented only after those checks pass. Never set verified, promoted, or optimized as the writer.
8. Add the writer-verified unit to the serial review batch only when the batch cadence and flush rules permit it.

Evidence Boundary

Raw same-weights model execution may omit value-parity fixtures but must record asset provenance, tensor interface, errors, lifecycle, routing, and rollback. Preprocessing, adapters, public payloads, errors, lifecycle, and promotion use the named legacy seam and exact behavior evidence; a Rust-only test is not an oracle.

Source And Runtime Rules

• Linux CPU is the target. Windows DML/GGUF remains legacy-owned.
• Do not add a production bridge before the promotion unit.
• Do not modify Cargo manifests or lockfiles when the shared foundation is blocked; stop and report the blocker.
• Keep model weights and generated multi-gigabyte artifacts out of Git.
• Use uv Python 3.12.x for Python checks.

Checks

uv run python scripts/check_model_manifests.py
cargo test --manifest-path rewrite-in-rust/rust/Cargo.toml -p v2m-inference <unit-filter>
cargo clippy --manifest-path rewrite-in-rust/rust/Cargo.toml -p v2m-inference --all-targets --all-features -- -D warnings

Skip and report Cargo checks when the declared workspace member is absent; do not change the workspace to hide the blocker.

Completion Response

Name the model/manifest/unit, files changed, writer checks, resulting state, review roles still required, blockers, and rollback route.