v3 Slice 1: Scaffold + spine - package, registries, tagged-text extractor, status taxonomy

[mark-allwyn]

Parent

#13

What to build

The foundational skeleton every other slice imports. Not a full vertical slice on its own - it stands up the package, the extension points, and the shared primitives so the tracer bullet stays thin.

• A new benchpress package at the repo root (the v2 system stays under legacy/), with a CLI stub exposing the command surface (generate, run, score, stats, export, audit) - subcommands may be stubs that exit cleanly.
• Core data types: Item, Part, PartResult, ItemResult, ModuleMeta (plain serializable dataclasses matching the PRD shapes).
• The three registries that provide extensibility (generalizing the legacy CHECKERS dict): a module registry, a part-scorer registry keyed by part type, and a metric registry - each with a registration decorator, populated by import side-effect.
• The tagged-text extractor: parse a model response into a dict of label -> raw field, tolerant (case-insensitive labels, markdown-tolerant, last occurrence wins, extra prose ignored).
• The status-taxonomy classifier: map (stop_reason, stop_details, extraction outcome) -> one of ok | refusal | invalid_answer | truncated | api_error.
• Root config model registry reused from the existing config.yaml shape (provider, model, company, type, launch_date, api_key_env, params).

Acceptance criteria

• benchpress imports cleanly; pytest exits 0 at the repo root (a trivial placeholder test is fine)
• Registering a module / part-scorer / metric makes it discoverable via its registry
• Tagged-text extractor unit-tested: case-insensitive labels, markdown fences, repeated labels (last wins), missing label -> absent key
• Status classifier unit-tested across all five statuses, including the refusal stop_reason path
• No secrets committed (.env stays gitignored)

Blocked by

None - can start immediately