[cschanhniem]

Summary

The rich selectivity scorer (v0.5.16, PR #130) was benchmarked and proven significant (detection AUROC=0.7138, CI 0.63-0.80) but was not connected to any production path where candidate selection actually happens. Every candidate was still scored with the old selectivity_proxy (detection AUROC=0.5744, CI 0.50-0.66, NOT significant) for hemolysis risk in:

• score_candidates() in pipeline.py
• expert composite scorer (as hemolysis_safety component, AUROC=0.565, not significant)
• pilot_priority formula in selection/pilot.py
• pilot panel report in reports/pilot_panel.py
• evidence certificates (via raw_scores in pipeline output)

This PR wires rich_selectivity into all five paths.

Changes

1. pipeline.py: rich_selectivity now computed for every candidate and included in raw_scores alongside selectivity_proxy (kept for backward comparability).

2. scoring/expert.py: rich_selectivity replaces hemolysis_safety (AUROC=0.565, not significant) as the expert composite hemolysis-risk component (weight 0.10). The old selectivity_proxy remains as the expert's activity-ranking selectivity component (weight 0.20) because it carries AMP-vs-decoy signal (AUROC=0.7729). Old hemolysis_safety preserved as hemolysis_safety_legacy in expert extras.

3. selection/pilot.py: _pilot_priority() uses rich_selectivity instead of selectivity_proxy for the safety tiebreaker, with backward-compat fallback to selectivity_proxy when rich_selectivity is absent.

4. reports/pilot_panel.py: pilot CSV and markdown include a rich_selectivity column alongside selectivity_proxy for expert comparison.

5. benchmark/retrospective.py: expert ablation benchmark references rich_selectivity instead of hemolysis_safety in per-component AUROC.

6. tests/test_rich_selectivity_integration.py: 14 new integration tests verifying rich_selectivity flows through pipeline scores, evidence certificates, expert composite weights/components, pilot priority formula (with fallback), and pilot panel report output.

Honest Results

Expert composite AUROC drops slightly: 0.7119 → 0.7097 (−0.0022) on AMP-vs-decoy. This is expected and acceptable:

• rich_selectivity is anti-AMP by design (AUROC=0.1973 for AMP-vs-decoy) because it penalises high hydrophobicity and charge that define AMPs.
• The tradeoff is correct: the expert now includes a significant hemolysis detector (CI excludes 0.5) instead of the old non-significant one (hemolysis_safety, AUROC=0.565).
• The old selectivity_proxy (weight 0.20 in expert) compensates for the anti-AMP bias of rich_selectivity (weight 0.10), keeping the net AUROC drop minimal.

Verification

• uv run pytest -q — 1571 passed, 7 skipped
• ruff check src tests — clean
• git diff --check — clean
• make metrics-snapshot — regenerated with updated expert AUROC
• pytest tests/test_docs_consistent.py — 12 passed

Constraint: wet-lab hemolysis assay remains mandatory; computational selectivity scores are triage signals only, not biological proof.
Confidence: high — rich selectivity detection AUROC=0.7138 (CI 0.63-0.80, significant); integration verified by 1571 passing tests.
Tested: uv run pytest -q (1571 passed, 7 skipped), ruff check (clean), git diff --check (clean), make metrics-snapshot.

Co-authored-by: OmX omx@oh-my-codex.dev