[cschanhniem]

Implements C4 from NEXT_100_PR_MAP: similarity-neighbor distribution report for novelty honesty.

What

• src/openamp_foundry/evidence/similarity_neighbor_report.py: compute_similarity_report() + format_similarity_report() + SimilarityNeighborReport dataclass
• SIMILARITY_SHORTCUT_THRESHOLD = 0.80 — fraction above which a sequence is considered high-similarity to references
• SHORTCUT_FRACTION_THRESHOLD = 0.50 — if >50% of positives are high-similarity to known AMPs, novelty_shortcut_likely=True
• Character-overlap similarity (positional match / max length); notice included that alignment tools are needed for publication-grade analysis
• 63 tests in tests/evidence/test_similarity_neighbor_report.py
• BASELINE: 6898 → 6961

Why

Benchmark improvements can be driven by the model learning that benchmark positives look like known AMPs in sequence space, not by discovering genuinely novel candidates. This report makes that shortcut visible and auditable.

Disconfirming pass

• Similarity metric is positional only; noted in NOTICE. Does not replace alignment.
• Report flags the shortcut; does not block anything — it is evidence, not policy.
• No safety/release policy touched.