feat(train): add generation and full-pool retrieval eval

[ppodolsky]

Summary

• add a full-pool retrieval evaluator with deterministic tie handling, rank-tail statistics, distractor shards, and duplicate-document validation
• add free-running generation evaluation that shares prompt framing with supervised training and reports EOS, repetition, containment, and source-overlap signals
• make retrieval rank semantics consistent in the existing evaluator and document both workflows

Verification

• cargo test --workspace
• cargo fmt --all -- --check
• strict clippy via the repository pre-push hook
• real Metal checkpoint smoke test produced a schema-v3 156-document report; the complete three-checkpoint matrix is running separately