Fix negative evidence cascade suppressing same-name merges

[claude]

Summary

• Fix self-reinforcing negative evidence cascade that suppresses all same-name merges in a connected component when unrelated entities share neighbors via relation-similar edges (GH issue Negative evidence cascade suppresses legitimate same-name merges #30)
• Two targeted changes to the propagation inner loop: best-counterpart gating for negative evidence, and merged-neighbor deduplication for positive evidence
• Three new tests covering CEO succession, shared employee bridge, and regulator/regulated entity patterns

Root cause

The all-pairs negative evidence loop generated contributions from every low-confidence cross-pair, even when both neighbors had good matches elsewhere. For example, Park₁↔Chen₂ (confidence 0) generated negative evidence for Nextera₁↔Nextera₂ despite Park₁ matching Park₂ and Chen₂ matching Chen₁. Since seed = 1.0 zeroes out positive evidence (pos × (1 - seed) = 0), only negative evidence counted — Nextera dropped below 0.5 and cascaded to suppress Park₁↔Park₂ and Chen₁↔Chen₂.

A secondary issue: after progressive merging, multiple relation-similar edges to the same merged neighbor (e.g. "is CEO of" and "is outgoing CEO of") were counted independently, inflating positive evidence and causing false merges between unrelated entities sharing a merged neighbor.

Changes

worldgraph/match.py — propagation inner loop:

1. First pass computes best counterpart confidence for each neighbor of ca and cb
2. Negative evidence gated: only counted when both neighbors lack a better alternative (best counterpart ≤ 0.5)
3. Merged-neighbor (ra == rb) contributions deduplicated by canonical ID — one structural fact per merged entity

docs/negative_evidence.md — documents best-counterpart gating and merged-neighbor dedup

Tests — 3 new tests encoding the cascade patterns from the issue:

• test_predecessor_successor_at_same_company_no_match (test_propagation.py)
• test_shared_employee_bridge_no_company_merge (test_integration.py)
• test_regulator_and_regulated_entity_stay_separate (test_integration.py)

Test plan

• All 63 tests pass (60 existing + 3 new)
• New tests fail before the fix, pass after
• Existing negative evidence tests still pass (cross-cluster isolation, disjoint neighborhoods, shared anchor)

Closes #30

🤖 Generated with Claude Code