[cschanhniem]

Summary

Bug fix: ensemble.py silent crash on missing score key

# Before (raises KeyError if scores is missing a weighted key)
value = sum(scores[name] * weight for name, weight in weights.items()) / total_weight

# After (degrades gracefully, treats missing scorer as 0 contribution)
value = sum(scores.get(name, 0.0) * weight for name, weight in weights.items()) / total_weight

If any upstream scorer fails to produce a key that appears in weights, the old code would raise KeyError and crash the entire pipeline run (no JSONL output, no certificates, no manifest). With .get(), the missing scorer contributes 0.0 to the ensemble — a safe degradation that still produces ranked output so a human can review.

New tests

• test_missing_score_key_defaults_to_zero — weights has "safety" but scores doesn't; verifies result is 0.5714 (weighted avg with safety=0.0) not a crash
• test_extra_score_key_not_in_weights_is_ignored — extra keys in scores (e.g. boman_activity) that aren't in weights are ignored; result is unchanged
• test_eligibility_novelty_at_exact_minimum_passes — eligibility filter uses >= so a candidate with novelty=0.0 and min_novelty=0.0 must be selected; confirms boundary operator is inclusive

Test plan

• pytest tests/test_scoring_ensemble.py tests/test_pipeline_filters.py — 41 passed
• Full suite pytest — 777 passed
• Pre-push hook passed

Context

Iteration #24 of the pre-wet-lab quality loop. The ensemble missing-key crash is a real reliability risk: if a physchem feature is NaN or a scorer throws, the whole run dies silently instead of producing degraded but usable output.

🤖 Generated with Claude Code