[cschanhniem]

Summary

Two targeted improvements to close consistency gaps identified in the post-PR-51 audit:

1. safety_score() — charge_density_ph74 consistency

safety.py was using the count-based charge_density proxy while activity.py uses charge_density_ph74 (pH 7.4 adjusted). For His-containing sequences (pKa 6.0), these differ significantly — up to 89% for histidine-rich sequences.

Fix: charge_density = abs(features.get("charge_density_ph74", features["charge_density"])) — uses pH 7.4 value preferentially, falls back to proxy for backward compatibility. Now consistent with activity scorer.

2. _pilot_priority() — stronger cytotoxicity penalty

The old formula gave cytotoxic-risk candidates (selectivity_proxy < 0.5) a soft reduced bonus but no explicit penalty. SEED-004 temporin (proxy=0.30) was only 0.035 points below a fully selective equal-ensemble candidate.

Fix: Add cytotox_penalty = 0.05 × (0.5 − proxy) / 0.5 for proxy < 0.5. SEED-004 now swings by 0.055 total (was 0.035) — still dominated by ensemble but enough to consistently deprioritize HIGH_CYTOTOX_RISK candidates when ensemble scores are close.

proxy	old net	new net	swing
1.00	+0.050	+0.050	0
0.50	+0.025	+0.025	0
0.30	+0.015	−0.005	−0.020 more
0.00	+0.000	−0.050	−0.050 more

Test plan

• 1172 tests pass
• New test: charge_density_ph74 preferred over proxy in safety scorer
• New tests: cytotox penalty boundary (0.50), max (0.00), intermediate (0.25)
• Updated tests: max delta 0.05→0.10, SEED-004 demerit 0.035→0.055
• Pre-push hook green

🤖 Generated with Claude Code