GoldenMatch

Changed

• Zero-config now routes probabilistic-shaped datasets to Fellegi-Sunter
  by default (#1874). A dataset with no surviving strong-identity exact
  matchkey (identifier/email/phone) and 2+ fuzzy fields is served by the
  EM-weighted FS path instead of exact+weighted matchkeys -- measured F1
  lifts on error-heavy PII (historical_50k 0.62 -> 0.78) with no regression
  where a strong key survives. Kill-switch:
  GOLDENMATCH_AUTOCONFIG_ROUTE_PROBABILISTIC=0.

• Config-healer loop cost is now bounded and tunable (#1404). heal (and
  the review_config / suggest_from_result verify path it drives) could run up
  to step_cap × (1 + max_verify) full pipeline passes per call — ~45 on the
  defaults — plus a full goldencheck blocking_risk variant scan (O(distinct²)
  per string column) re-run on the unchanged frame every iteration. Three cost
  levers, all defaulting to byte-identical behavior:

  • The goldencheck variant scan is memoized for the whole heal loop via a
    new variant_risk_cache() scope (core/suggest/adapter.py) keyed on the
    data-column set — it runs once over the frame instead of once per iteration.
    Output is unchanged; only the cost moves.
  • Verify fan-out is tunable: review_config/suggest_from_result/heal
    take max_verify, and GOLDENMATCH_SUGGEST_MAX_VERIFY sets it globally
    (default 8). Since the healer applies only the top surviving suggestion,
    max_verify=1 verifies just that candidate — the cheapest mode.
  • Marginal-gain early-stop: heal(min_health_gain=…) /
    GOLDENMATCH_HEAL_MIN_HEALTH_GAIN stops the loop once cluster-health gain
    flattens (fail-open — a result without clusters never triggers a stop). Off
    by default. GOLDENMATCH_HEAL_STEP_CAP also exposes the outer cap.

Added

• DedupeResult.identity_summary (#1913). The per-run identity-
  resolution summary (entities created/absorbed/merged) is now surfaced on
  the public result -- None when identity resolution is disabled. Backs the
  in-Postgres gm_resolve write path in goldenmatch-pg.
• Small-N Fellegi-Sunter routing floor (#1947). Below
  GOLDENMATCH_FS_ROUTE_MIN_ROWS rows (default 500) a probabilistic-shaped
  dataset stays on the robust weighted path -- FS EM is data-starved at small
  N and under-merges fuzzy-close variants. 0 disables the floor. Every
  dataset that validated the FS default is far above it, so routing there is
  unchanged.

Fixed

• Net-zero-evidence filter kills scale-growing FS over-merge (#1899,
  default ON). Pairs whose only agreement is on absent (unobserved) fields
  no longer accrue spurious match weight; ported to the numpy and native
  kernels.
• FS at 1M rows (#1896). A pairs-budget blocking gate stops a low-
  cardinality pass from compounding into a megablock, and an Arrow pair-
  stream plus EM block-sample cut the FS memory peak.
• Unobserved record_embedding masked out of the weighted score
  (#1859). A record with no embedding no longer contributes max agreement.
• Arrow-native auto-config blocking profile emitter (#1946). The
  controller's sample iterations no longer force an arrow->polars round trip
  just to count rows, so zero-config runs correctly (no degraded RED-sentinel
  config) on a base, polars-free install. Byte-identical with polars present.