goldenmatch 0.5.0 → 0.6.0

package/CHANGELOG.md

@@ -4,6 +4,60 @@ All notable changes to goldenmatch-js are documented in this file.
 
 Format follows [Keep a Changelog](https://keepachangelog.com/). Versioning follows [Semantic Versioning](https://semver.org/) (strict after v1.0.0).
 
+## [0.6.0] - 2026-05-10
+
+Indicator-aware refit parity with Python `goldenmatch` v1.9 + v1.10.
+
+### Added
+
+- `IndicatorContext` memoization layer (`src/core/indicators.ts`) and 5 pure
+  complexity indicators ported from Python `core/indicators.py`:
+  `computeColumnPriors`, `estimateSparseMatchSignal`,
+  `computeCorruptionScore`, `estimateFullPopHits`,
+  `computeCrossBlockingOverlap`, plus `computeIdentityCollisionSignal`
+  used by the collision-aware refit rule.
+- 7 new indicator-aware refit rules in `autoconfigRules.ts`:
+  `ruleUniformHeavyBlocking`, `ruleBlockingFieldNullHeavy`,
+  `ruleRecallGapSuspected`, `ruleCollisionSignalTooHigh`,
+  `ruleSparseMatchExpand`, `ruleCrossBlockingDisagreement`,
+  `ruleCorruptionNormalize`.
+- `DEFAULT_RULES_V1_10` — 14-rule list mirroring Python's `DEFAULT_RULES`
+  order. The legacy `DEFAULT_RULES_V1_7_V1_8` 7-rule list is still exported
+  for callers that opt into base-only behavior.
+- `RuleContext.indicators` optional field carries the per-iteration
+  `IndicatorContext`; rules that need indicator signals are silent no-ops
+  when callers run the legacy v1.7/v1.8 rule list.
+- `RefitPolicy.propose(profile, current, history, indicators?)` — fourth
+  positional argument (back-compat: defaults to `null`).
+
+### Changed
+
+- `autoConfigureRows` rewrite: matchkey naming now matches Python
+  (`fuzzy_match` for weighted, `exact_<col>` for exact). Scorer selection
+  follows Python's `_SCORER_MAP` (e.g. `name → ensemble`,
+  `email → exact`). Adaptive threshold uses Python's formula plus the
+  post-build data-quality adjustment (avg_null > 0.15 → −0.05;
+  avg_len < 5 → +0.05).
+- `buildBlocking` aligned with Python: prefers high-cardinality
+  exact-eligible columns (email/phone/zip/identifier/year) for static
+  blocking, falls back to multi-pass name blocking
+  (`soundex` + `substring:0:5` + `token_sort + substring:0:8`).
+- Controller provisions a fresh `IndicatorContext` per iteration and
+  threads it into `policy.propose()` for v1.10 rule consumption.
+
+### Parity status
+
+- Controller stoppoint parity: 6/6 datasets pass shape-level assertions,
+  2/6 (`dirty_people`, `mixed_blocking`) byte-equal on the normalized
+  committed config. The remaining 4 diverge because Python's iteration
+  path hits a `ModuleNotFoundError` on subsequent iterations and falls
+  back to a virtual v0 entry (out-of-scope to replicate in TS).
+- Indicators parity: 8/8 fixture datasets pass at 4-decimal tolerance
+  on the 5 indicators. Identity-collision signal is unit-tested only —
+  the TS pure-JS token-sort approximation diverges numerically from
+  Python's `rapidfuzz.token_sort_ratio` at sub-rule precision, but the
+  rule-firing boundary (rate > 0.75) is preserved.
+
 ## [0.5.0] - 2026-05-10
 
 Auto-config controller parity with Python `goldenmatch` v1.7 + v1.8.