evals: implement an offline human-review benchmark report

[WalksWithASwagger]

Summary

Implement an offline human-review benchmark report against the ratified model-eval contract.

Context

• Blocked by decision(evals): ratify a privacy-safe model-output benchmark contract #335.
• Roadmap phase: Phase 2 - model quality
• Relevant files: a small report module such as lib/eval_report.py, a CLI command under lib/commands/, generate.py, focused fixtures/tests, and docs/MODEL-EVALS.md
• Current behavior: run and review metadata cannot be aggregated into a stable benchmark comparison.
• Desired behavior: a read-only offline command joins approved benchmark manifests, run metadata, and human evaluations into deterministic JSON and Markdown summaries.

Implementation Notes

• Read only local approved fixture/manifest paths.
• Aggregate by model, style, prompt, decision, and rubric score.
• Report missing outputs and unreviewed items explicitly rather than dropping them.
• Keep generated media and private benchmark data gitignored.

Acceptance Criteria

• A read-only command joins the approved benchmark manifest, run.json, and evaluations.json.
• It emits deterministic JSON and Markdown summaries.
• Summaries include model, style, prompt, decision, rubric dimensions, missing outputs, and unreviewed items.
• Fixture tests cover aggregation, malformed input, missing reviews, and stable ordering/output.
• No provider or network call occurs.
• No generated image or private benchmark artifact is committed.

Tests/Evals

• Use only the synthetic records approved by the benchmark-contract issue.
• Snapshot only compact deterministic report text, not images or whole portal pages.

Verification

• Run the focused eval-report test module.
• python3 -m ruff check lib tests
• npm test
• Run a dry fixture report twice and confirm byte-identical JSON and Markdown.

Agent Instructions

• Use branch codex/issue-<this-issue>-offline-eval-report.
• Do not start until the eval contract is approved and merged.
• Do not call providers, rank models automatically, or alter defaults.
• Keep input/output paths explicit and local.

Human Checkpoints

• Stop if the ratified schema cannot represent a required review dimension without a policy change.

Out of Scope

• Generating benchmark outputs.
• Automated vision judging or model ranking.
• Portal UI for evaluations.
• Changing provider/model defaults.

Linear

Not applicable. Rafiki delivery is GitHub-only; do not create or update a Linear issue.