fix: skip impossible file-anchored dedup scoring

[GautamBytes]

Summary

• skip cross-file rationale/document dedup pairs before LSH/Jaro scoring
• apply same grouping to --dedup-llm tiebreak path

[GautamBytes]

@safishamsi please review it once whenever feasible!

[GautamBytes]

@safishamsi just pinging for review!!

[graphify-labs]

Looks safe to merge — no coupling regressions and no blocking issues, checked against the code graph (not a self-assessment).

Graphify reviewed this change.

Looks safe to merge — no coupling regressions and no blocking issues, checked against the code graph (not a self-assessment).

---

Graphify review — findings

This PR refactors the fuzzy-dedup phase in graphify/dedup.py to group candidates before running them through LSH, so that file-anchored non-code nodes (rationale/document) are only compared against nodes sharing the same source_file, while other node types keep global comparison. It introduces helper functions _is_file_anchored_noncode and _fuzzy_candidate_groups, restructures the main loop to iterate over these groups with a per-group LSH index and a shared seen_pairs set to avoid re-evaluating pairs. The test file tests/test_dedup.py is updated with numerous new/modified test cases and rationale fixtures covering cross-file merge behavior and determinism.

Worth a look

• Empty source_file ("") anchored nodes grouped together across unrelated files — graphify/dedup.py · Escalate · medium
  • agreed by 2 of 2 members but NOT verified (no proof, no reproducing execution) — consensus is not a verdict; needs human review

Analysis details — impact, health, verification

Impact & health

Graphify review

Impact — 307 functions depend on the 155 node(s) this change touches.

Health — this change adds coupling hotspots:

• worse: deduplicate_entities() — 51 callers, 22 callees
• worse: _llm_tiebreak() — 1 callers, 12 callees

Verification — 307 functions in the blast radius were not formally verified this run (proofs are advisory here).

Gate & verification

graphify gate

PASS — objectively clean (no health regressions, tests not run — proofs not run this pass (advisory)). Grounded, not self-assessed.

Advisory (not blocking):

• verification_scope: 172 function(s) in the blast radius were not formally verified this run

· 2 more finding(s) on lines outside this diff (see the check run).