[cschanhniem]

Summary

• New module: src/openamp_foundry/evidence/benchmark_registry.py
• 5 BenchmarkCard instances covering all major pipeline benchmarks:
  • BMC-0001: AMP Activity Precision@k (leakage_aware_split, 3 metrics, 3 limitations)
  • BMC-0002: Charge-Matched AMP Activity (charge_stratified, 2 cheap enemies)
  • BMC-0003: Calibration Accuracy (random_70_30, calibration_error+brier_score)
  • BMC-0004: Family-Stratified AMP Precision (family_stratified, 3 cheap enemies, 3 limitations)
  • BMC-0005: Cheap Enemy Comparison (4 cheap enemies including conjunction, anti-hype primary benchmark)
• API: get_card(bmc_id) → card|None; validate_registry() → list[errors] (CI-callable)
• All 5 cards: ≥2 cheap enemies, ≥2 known limitations, validate() clean
• 63 tests covering registry structure, all-cards-valid, cheap enemy requirements, known limitations, get_card(), benchmark types covered, by-ID lookup
• BASELINE 6709→6772
• Closes Phase C C1 — benchmark governance is now machine-enforceable via validate_registry()