GoldenMatch

Changed

• Native Fellegi-Sunter (FS) block scorer is now authoritative by default (reference mode). _fs_native_enabled() flips from opt-in to default-on: when the native ext is importable, the probabilistic path uses the native Rust FS kernel (rapidfuzz-rs decides comparison levels); the numpy vectorized path becomes the reproducible fallback via GOLDENMATCH_FS_NATIVE=0 (also the automatic fallback for TF-adjustment / non-native-scorer fields or a missing wheel). Part of the Rust-is-the-reference direction (docs/design/2026-07-01-rust-is-the-reference-roadmap.md). Scoped to the probabilistic path only (opt-in type: probabilistic matchkeys / probabilistic routing); the default weighted path is unaffected. Measured F1-neutral on the probabilistic bench panel (gm_prob_native vs gm_probabilistic): febrl3 and synthetic_person identical, historical_50k −0.0007 (within noise); dblp_acm not measured (Leipzig CSVs gitignored in CI). Boundary-level score differences are possible where a rapidfuzz-rs vs rapidfuzz-py similarity sits exactly on a comparison-level threshold — the native result is now the reference; GOLDENMATCH_FS_NATIVE=0 restores the prior numpy operating point.

Fixed

• Auto-config no longer commits a standalone exact matchkey on a shared locality attribute (#1351). A high-density column mis-classified as an identifier (e.g. a zip whose cardinality inflates on the 1k-row profiling sample) could back an exact matchkey and collapse everyone sharing a value into one cluster (~55% over-merge on real circulation data). A new discriminative-power veto (build_matchkeys → autoconfig_discriminative.should_veto_exact) demotes a proposed exact key to blocking-only when records sharing its value do NOT co-agree on other identity fields — measured from the data, not from cardinality (which cannot separate zip from npi: both moderate-cardinality, opposite correct answers). Name-typed basket fields co-agree FUZZILY (SequenceMatcher ≥ 0.85) so corrupted duplicates keep their key (e.g. febrl3's soc_sec_id, whose duplicates carry corrupted names); structured ids compare exactly. Veto-only (never promotes; classification/blocking untouched), fail-safe keep on thin support / df=None; kill-switch GOLDENMATCH_DISCRIMINATIVE_VETO=0. Auto-config accuracy gate held (febrl3 / ncvr_synthetic / historical_50k F1 unchanged).