Reproducible benchmark, funnel figure and validation dossier (0.8.0) #25

Part of the AI validation roadmap (docs/AI_VALIDATION_ROADMAP.md).

Goal. One reproducible benchmark and one document a pharma reader can assess.

Done when

• One command regenerates every benchmark table and figure from pinned snapshots, byte-for-byte (tools/reproduce.py, run in CI; One-command reproduction and the funnel figure (#25) #41)
• A funnel figure (docs/assets/ai_validation_funnel.svg, One-command reproduction and the funnel figure (#25) #41; the expert-review and audit stages await Expert review of the ClinVar packet #23, Calibrated triage and audit of auto-admitted records #24): claims extracted → rejected or routed at each stage → expert review → audited error of what was admitted
• A validation dossier (docs/VALIDATION_DOSSIER.md, Validation dossier (#25) #42) structured like FDA's draft AI credibility framework (January 2025): question of interest, context of use, model risk, credibility evidence, limitations
• Release 0.8.0 (Release 0.8.0 #40)