add stage-separated AMFV-Bench v0 eval

[VibeCodingScientist]

Summary

• Frozen AMFV-Bench v0: 48 NICE-grounded cases, stage-separated decomposer / retrieval / verifier metrics, and a deterministic amfv-eval scorer (no network, no LLM).
• Scorer rejects polarity and identifier swaps, counts omitted verdicts and skipped population-mismatch cases as misses, and does not reward decoy retrieval on insufficient cases.
• Gold Likert, scope, and evidence spans are clinician work. Rubric: datasets/amfv_datasets/eval/taxonomy.md.

v0 is a seed (seed-v0-clinician), not locked gold. If this lands I will (1) lock the 48 labels under a named annotator id after a full clinical pass, (2) independently re-grade or recruit a second clinician and adjudicate disagreements per taxonomy.md, and (3) stay on as an annotator for later splits (high-stakes traps, not silent growth of v0). Gold Likert / scope / evidence is clinician work; please do not mint verdicts from an LLM judge.

Replaces closed #19 so the red-team scorer fixes ship in one reviewable PR.

Test plan

• uv sync --dev
• uv run pytest datasets/test
• uv run ruff check datasets && uv run ruff format --check datasets
• uv run amfv-eval --predictions datasets/test/fixtures/eval/predictions_demo.jsonl (demo file supports a paediatric claim on purpose; scope_abstention should be 0)
• Skim datasets/amfv_datasets/eval/taxonomy.md and 2–3 gold cases for clinical sense

Made with Cursor

[CLAassistant]

All committers have signed the CLA.