[cschanhniem]

Implements C7 from NEXT_100_PR_MAP: family-stratified precision@k.

What

• family_stratified_report.py: compute_family_stratified_report() + format_family_stratified_report() + FamilyStratifiedReport
• Detects when a single AMP family dominates the top-k ranked sequences (>60% dominance threshold)
• Per-family precision@k computed for each family in the top-k window
• family_inflation_likely=True when dominant family fraction >= 0.60 AND >= 0.50 overall threshold
• 63 tests; BASELINE 7150 to 7213

Why

A model that learns to rank one family highly (e.g. all defensins) will appear to have good precision@k while actually just fitting within-family patterns. This report makes that family-inflation artifact visible.

Disconfirming pass

• Report is evidence only, does not block ranking.
• References BMC-0004 (family-stratified split) for publication-grade testing.
• No safety/release policy touched.