Implement negative structural evidence

[claude]

Summary

• Add dampened negative factor to similarity propagation that penalizes entity pairs whose functional neighbors have no matching counterpart, preventing false merges between identically-named entities in different structural contexts
• Key design: weighted average mismatch with binary relation threshold, max(neg_a, neg_b) bilateral factor, name-similarity seed for inner match check, and separate positive_base tracking for monotone structural propagation
• Removes xfail markers from test_identical_names_different_contexts_no_merge and test_similar_names_disjoint_neighborhoods_no_match; all 54 tests pass

Closes #23

Test plan

• All 54 existing tests pass (uv run pytest)
• test_identical_names_different_contexts_no_merge — two different "Dr. James Chen" in unrelated labs no longer merge
• test_similar_names_disjoint_neighborhoods_no_match — "Elena Vasquez" / "Lena Vasquez" with disjoint neighborhoods no longer match
• Positive-evidence monotonicity preserved (renamed test with neg disabled)
• Existing merge scenarios (Sharma name variations, shared entities across clusters, synonym relations) still pass

🤖 Generated with Claude Code

[claude]

Reviewed the full diff and source files. No issues found — approving.

What I checked:

• Negative factor correctness: name_seed is fixed (prevents circular reinforcement), positive_base is monotone non-decreasing, and since neg is constant per pair (computed from fixed name_seed), combined = base * neg is monotone after the initial drop. The abs(combined - old) > epsilon change correctly handles that initial drop.
• Forward vs inverse functionality swap between _build_weighted_adjacency and _build_forward_adjacency is correct per the PARIS design in negative_evidence.md.
• max(neg_a, neg_b) bilateral strategy is well-motivated for asymmetric news coverage.
• Test changes: xfail removals correct, NovaTech assertion relaxed with clear justification, monotonicity test correctly isolates positive propagation via neg_alpha=0.0, dead anchor variable and unused import pytest cleaned up.
• No dead code, no shims, no half-finished pieces. Reads as if negative evidence was always part of the design.

LGTM.

[monneyboi]

The negative evidence implementation reintroduces a binary relation similarity threshold (rel_threshold=0.8 in _one_sided_negative), but positive propagation still uses continuous rel_sim as a multiplier. This creates an inconsistent mental model: functionality pooling and negative evidence treat relations as "same or not" via a threshold, while positive propagation treats similarity as a continuous weight.

Requested change: Unify to a single rel_threshold parameter on match_graphs that flows to all three consumers:

1. compute_functionality — pool relations above threshold (already does this)
2. propagate_similarity (positive) — gate relation pairs: if rel_sim >= threshold, treat as same relation and propagate (without multiplying by rel_sim); below threshold, skip entirely. This reverts the "initial nongated" change from 695408b.
3. compute_negative_factor — check relation match using the same threshold (already does this, but hardcoded)

The mental model becomes: the threshold defines equivalence classes over free-text relations — essentially reducing them to a discrete enum. Above threshold = same relation, below = different. One threshold, one concept, used everywhere.

This means:

• Remove the separate rel_threshold param from compute_negative_factor / _one_sided_negative
• Add a rel_threshold param to propagate_similarity
• In the propagation loop, change strength_sum += rs * weight * neighbor_confidence to just strength_sum += weight * neighbor_confidence (only reached when rs >= rel_threshold)
• match_graphs passes the single rel_threshold (currently rel_cluster_threshold) to all three functions

[claude]

Reviewed the full diff (both commits) against match.py, the design doc, and the prior review. The threshold unification from monneyboi's feedback is fully addressed. No issues found.

What I checked:

• Threshold unification complete: rel_cluster_threshold flows from match_graphs → propagate_similarity → compute_negative_factor / _one_sided_negative. Positive propagation gates with rs < rel_threshold and drops the continuous rs * multiplier. One concept of relation equivalence throughout — functionality pooling, positive gating, negative evidence all use the same threshold.
• Convergence is guaranteed despite abs() change: positive_base is monotone non-decreasing, and the negative factor is computed from name_seed (fixed), not propagated confidence. So combined = base * neg converges once base does — no oscillation possible. The abs(combined - old) > epsilon correctly detects the initial drop when negative evidence first activates.
• Forward vs inverse functionality swap between _build_weighted_adjacency and _build_forward_adjacency is correct per PARIS semantics: positive evidence uses inverse functionality (target match → source match), negative evidence uses forward functionality (source should have specific target).
• max(neg_a, neg_b) bilateral strategy is well-motivated — asymmetric news coverage means missing neighbors on one side shouldn't compound.
• NovaTech assertion relaxation is correct: with one neighbor (founder) that doesn't match and forward functionality ~1.0, score = 1.0 × 0.7 = 0.7 < 0.8 is the right behavior — identical names with contradicting functional neighbors should be suppressed. The test's primary assertion (sharma ≠ vasquez) still holds.
• No dead code, no shims, no half-finished pieces. Clean integration.

LGTM — no issues to flag.