One-command reproduction and the funnel figure (#25)

[NingyuSUN]

Two of the four items of #25. 0.8.0 is already released; the validation dossier remains.

One command — uv run --frozen --with matplotlib==3.11.2 python tools/reproduce.py

• It regenerates every committed benchmark table, summary and figure offline, in 18 steps:
  • the three real-data cases;
  • ten LLM-benchmark result sets;
  • task files, semantic units and the error taxonomy;
  • four figures.
• Each output is compared with the repository: text with line endings normalised, everything else byte for byte. Locally all 18 steps reproduce, in about 2.5 minutes.
• Model calls, agent runs and downloads are not repeated: their outputs are committed and pinned.
• --write updates the committed files, --only runs selected steps, and --list lists them.
• A new CI job, reproduce, runs it on every change. tests/test_reproduce.py fails if any results directory or figure is not covered by a step.
• It caught a real drift: the ClinVar and VBO figures were still captioned v0.7.0 after the 0.8.0 release. They are regenerated here.

Funnel figure — docs/assets/ai_validation_funnel.svg, shown in the README and the docs homepage

• It follows the single-cell held-out answers (6 models × 46 clusters) stage by stage:
  • first check: 198 pass, 63 have an identifier or marker error and are sent back, 15 conflicts go to a person;
  • after feedback: 57 of the 63 are corrected and admitted, so 255 are admitted and 21 go to a person;
  • against the authors' terms: 205 of the admitted agree, 50 disagree;
  • expert audit: marked not yet run (Expert review of the ClinVar packet #23, Calibrated triage and audit of auto-admitted records #24).
• celltype_loop.score now records these stages in each summary. The three single-cell summaries are regenerated; only the new funnel field and line are added.

Checks: ruff, mypy, pytest (98.71%), mkdocs build --strict, and the full reproduction.