vocal2midi-rs-dep-bootstrap

name	vocal2midi-rs-dep-bootstrap
description	Prepare dependency, asset, tensor-interface, seam, and rollback evidence for one named Vocal2Midi Linux model unit under the py2rs control plane.

Vocal2Midi Model Dependency Bootstrap

Prepare exactly one unit in exactly one named model manifest. This is a prompt-mode decision skill: it explains capability and seam choices, while repository scripts validate deterministic structure.

Required Context

Read:

• rewrite-in-rust/models/control-plane.yaml;
• the selected model manifest only;
• rewrite-in-rust/models/README.md, resources.md, notes.md, and glossary;
• the unit's Python entrypoint, tests, assets, existing dependency record, bootstrap note, and historical records;
• pyproject.toml, Rust manifests, lockfiles, and vendored source indexes as relevant to the selected capability.

Do not read the closed backend manifest or other model manifests.

Dependency Gate

The unit must record dependency_recon before writer work:

• agent: run fresh-context py2rs-crate-recon for each observable capability;
• manual: require user-supplied registry, feature, dependency-path, and rejection evidence;
• disabled: require the explicit risk acknowledgement in notes.md.

Check rewrite-in-rust/dependencies/shared-dependencies.yaml before adding or hand-writing a shared capability. A blocked canonical path stops dependent implementation; do not create a second copy under /tmp, an agent directory, or another model shard.

Use the standard ladder: direct crate coverage when fixtures prove it, partial coverage plus a semantic-delta adapter when the Python source explains the gap, and narrow hand-written Rust only for uncovered or safer bounded behavior. Record why a full hand-written replacement is smaller or safer.

Model Evidence

For model_evidence units, prove:

• exact asset provenance, format, config, tokenizer, and license metadata;
• import/load success and every accepted tensor name, rank, shape, and dtype;
• representative empty/short/batched input handling where applicable;
• structured load and execution errors;
• cancellation, repeated calls, cleanup, and resource lifecycle;
• production handoff and legacy rollback.

Do not create model-output value-parity fixtures for the explicit same-weights exception. This does not exempt preprocessing, adapters, public payloads, errors, lifecycle, or routing units from exact legacy-seam evidence.

Re-cut Rules

Split or defer the selected unit when it mixes separately verifiable capabilities, lacks a stable public seam, needs an unowned shared data structure, or has no rollback route. Keep the selected model's other shards untouched.

Durable Output

Update only the selected manifest and its unit-scoped dependency/bootstrap artifacts when implementation work is authorized. Record:

• capability list and source paths;
• applied rewrite profile and deviations;
• crate-recon report/status;
• canonical shared dependency ids and owners;
• asset/tensor/seam smoke commands;
• behavior-verification mode and observations;
• inventory impact (confirmed, split, merged, deferred, or replaced);
• rollback route and concrete blockers.

Do not mark a unit reimplemented or verified here. The writer owns implementation state; independent review owns verification.

Checks

uv run python scripts/check_model_manifests.py

Run focused import/load and tensor-interface checks only when their canonical Rust source exists. If Cargo metadata fails because the declared shared model foundation is absent, report that blocker without changing Cargo files.

Completion Response

Name the selected model/unit, capability decisions, recon mode/report, shared dependency status, seam, evidence commands, inventory impact, blockers, and rollback route.