[cschanhniem]

Summary

The standard triage benchmark (v0.5.12) uses random background peptides as decoys, which are trivially distinguishable from AMPs due to protein-like composition. This inflates selective_vs_decoy and hemolytic_vs_decoy AUROCs, making scorers appear to triage well when they are really only solving the easy AMP vs random problem.

This PR adds a strict triage benchmark that replaces random decoys with composition-matched scrambled versions of selective AMPs (same amino acids, permuted order). This destroys amphipathic helical phase, hydrophobic moment, and charge distribution while preserving all composition-based features.

Key Honest Finding

No scorer triages correctly with composition-matched decoys. The standard triage success of selectivity_proxy and expert_composite was an artifact of trivially distinguishable decoys.

Scorer	Std sel_vs_dec	Strict sel_vs_dec	Std correct	Strict correct
selectivity_proxy	0.782	0.500	YES	NO
expert_composite	0.757	0.510	YES	NO
ensemble	0.848	0.572	NO	NO
triage_score	0.863	0.674	NO	NO

• selectivity_proxy collapses to exactly 0.5000 on selective_vs_decoy (purely composition-driven)
• Ensemble drops from 0.848 to 0.572 (most signal was composition, not order)
• selective_vs_hemolytic is identical between standard and strict (expected, both classes are real AMPs)
• Ensemble admits 7 scrambled decoys into top-20 (vs 0 with random decoys)

Implication

The pipeline triage signal is almost entirely composition-driven. The real bottleneck (selective-vs-hemolytic discrimination) requires structural or contextual features beyond what current 1D physicochemical scorers can capture. Any future virtual assay layer must demonstrate order-dependent triage signal on this strict benchmark before claiming to improve candidate selection.

Changes

• benchmark/triage.py: run_strict_triage_benchmark() + scramble helpers
• benchmark/metrics_snapshot.py: strict_triage section in snapshot
• tests/test_triage_benchmark.py: 12 new tests (structure + findings)
• outputs/metrics_snapshot.json: regenerated with full 2000-iteration bootstrap
• docs/METRICS_CURRENT.md: strict triage results table, interpretation, changelog
• docs/ROADMAP.md: v0.5.14 entry + v1.1 triage item updated

Verification

• uv run pytest -q: 1516 passed, 7 skipped
• git diff --check: clean
• tests/test_docs_consistent.py: all 9 doc-consistency guards pass
• make metrics-snapshot: regenerated with strict_triage section

Tested: uv run pytest -q (1516 passed, 7 skipped), git diff --check (clean)
Confidence: high - composition-matched decoy AUROCs are deterministic (seed=42)
Co-authored-by: OmX omx@oh-my-codex.dev