Problem
matchingPairs in src/TypeScriptDuplicateFinder.ts gates pairs purely on the Jaccard similarity(left, right) >= threshold. A small candidate can clear a high ratio on very few shared fingerprints — e.g. two 6-node candidates sharing 5 fingerprints hit ~0.83 Jaccard but represent almost no shared structure. --min-nodes (default 20) mitigates this by raising the candidate-size bar, but ratio-only matching still admits high-ratio/low-absolute-overlap pairs near the size floor.
Proposed change
Add an absolute shared-fingerprint floor applied in addition to the Jaccard threshold:
- New option
--min-shared-fingerprints <n> (parsed in src/Options.ts), default off / 0.
- In the pair gate (
src/TypeScriptDuplicateFinder.ts, the score >= threshold site, ~L100), also require the intersection size to be at least the floor. The intersection count is already computed inside similarity (shared); expose it or inline the check so no second pass is needed.
Acceptance
- Two tiny same-shape candidates that pass Jaccard but share fewer than the floor are dropped.
- A large genuine duplicate (high absolute overlap) is unaffected.
- Flag off -> behavior byte-for-byte unchanged.
Related
False-positive-reduction set (#19-#24); codex-surfaced. Complements the kind-diversity floor (#20): that bounds intra-candidate uniformity, this bounds inter-candidate overlap magnitude.
Problem
matchingPairsinsrc/TypeScriptDuplicateFinder.tsgates pairs purely on the Jaccardsimilarity(left, right) >= threshold. A small candidate can clear a high ratio on very few shared fingerprints — e.g. two 6-node candidates sharing 5 fingerprints hit ~0.83 Jaccard but represent almost no shared structure.--min-nodes(default 20) mitigates this by raising the candidate-size bar, but ratio-only matching still admits high-ratio/low-absolute-overlap pairs near the size floor.Proposed change
Add an absolute shared-fingerprint floor applied in addition to the Jaccard threshold:
--min-shared-fingerprints <n>(parsed insrc/Options.ts), default off / 0.src/TypeScriptDuplicateFinder.ts, thescore >= thresholdsite, ~L100), also require the intersection size to be at least the floor. The intersection count is already computed insidesimilarity(shared); expose it or inline the check so no second pass is needed.Acceptance
Related
False-positive-reduction set (#19-#24); codex-surfaced. Complements the kind-diversity floor (#20): that bounds intra-candidate uniformity, this bounds inter-candidate overlap magnitude.