eval(bakeoff): shared driver for the #1269 model bake-off

[jasonssdev]

Summary

The shared driver the #1269 pre-registration requires before the first bake-off run. evals/ only; no model was run.

1. Latency and identity stamping (all eight harnesses in the run plan)

• query_attribution stores elapsed_s per answer and run_latencies_s per run; decision_revisions wraps its client in a timer and stores call_latencies_s / run_latencies_s; edge_typing, contradictions and adjudication persist the per-run latencies they already collected. Old stored rows still load.
• evals/harness_stamp.py adds a "stamp" to every runs JSON: harness commit and dirty flag, model name and Ollama digest (loopback /api/tags only), and sha256(prompt)[:16] per prompt sent (ids like contradiction/system), read after any treatment swap. The definition matches feat: keep LLM prompts in one folder, one file each, versioned by content hash #1277's prompt hash and production's SUBJECT_PROMPT_VERSION / JUDGE_PROMPT_VERSION (asserted in the self-test).
• query_sufficiency gains --arms quote, so the bake-off pays only for the shipped prompt.

2. evals/bakeoff/

• bakeoff_spec.py: candidates and every bar as data, each citing the pre-registration comment.
• bakeoff_bars.py: "no worse than baseline", vetoes, the waiver for bars the baseline already fails, the no-headroom rule, the knockout.
• bakeoff_metrics.py: one extractor per harness; on stored results they reproduce the pre-registered baselines (edge accuracy 0.357, attribution 0.933, B1–B8 pass, auto_merge FAIL).
• bakeoff_ollama.py: license, native-context and ollama ps memory gates, loopback only, never pulls.
• run_bakeoff.py: --plan, --run, --evaluate-only, --self-test; resumable (done cells skipped, failed retried); the second baseline session only runs in a later invocation; an unworked adjudication queue blocks rather than drops a candidate; Write takes the top 2 Generate survivors.
• Interpretation choices are listed in evals/bakeoff/README.md.

Related issue

Refs #1269

Type of change

• feat — new feature
• fix — bug fix
• docs — documentation only
• refactor — no behavior change
• test — tests only
• chore / ci — tooling, build, or CI
• Breaking change

How was this tested?

• New or extended model-free self-tests: stamp (27 checks), bars (40), metrics (40), spec (122), ollama (33), runner (52). The sweep runs 52/52 green.
• 19 mutations of comparisons and rules (strict bounds, a veto that never drops, disabled waiver/no-headroom, knockout that never drops, inverted resume/deferral, Write picking the worst, a 15-char prompt hash, transcribed bar values, …) all turn the self-tests red.
• Rebased onto main after eval: #1258 hard negatives for adjudication and a direction metric for edge typing #1272 and eval(resolution): reproduce #1223 on its field shapes; refute a one-sentence treatment #1275 (one import conflict in edge_typing, both imports kept). ruff check, ruff format --check, mypy . and evals/run_self_tests.py (52/52) pass on the rebased tree; pytest --cov passed before the rebase (96.46%) and runs in CI.

Checklist

• My commits follow Conventional Commits.
• I added or updated tests for the change.
• I updated docs where behavior, interfaces, or the knowledge model changed.
• Lint, format, type check, and tests pass locally (ruff, mypy, pytest).
• Output remains OKF-conformant and derived stores stay reconstructible from the bundle + sources.
• The change is consistent with the project's guiding principles (local-first, provenance, freshness, human-in-the-loop).