GoldenMatch

Fixed

• Linear identity golden-record resolution: incremental identity resolution
  now builds each golden record from the row payload index prepared once per
  batch, rather than filtering the entire input frame once per cluster. This
  removes the quadratic singleton-heavy archive ingestion path.

Added

• Anomaly diagnostics with prefilled GitHub issue prompts (in-tree:
  goldenmatch.core._diagnostics_report + goldenmatch.core.diagnostics; no
  new package or dependency). When GoldenMatch hits a state that is probably
  its own bug, it emits an actionable message with a prefilled issue URL.
  Fires only on anomalies -- never on expected fallbacks or user errors:
  (1) a native wheel-skew slow path (the kernel symbol is missing from the
  installed wheel -- the #688 class); (2) an unexpected crash at dedupe_df /
  match_df (re-raised unchanged, with a traceback + PII-safe environment in the
  prompt; ControllerNotConfidentError, bad config, FileNotFound, ValueError
  and other by-design/user errors are never prompted); (3) the config linter
  itself crashing; (4) a broken native install (module present but fails to
  load, vs plain "not installed"). Sends nothing anywhere -- it is a better error
  message, not telemetry. Silence with GOLDEN_DIAGNOSTICS=0. Diagnostics is
  never load-bearing: the reporter never raises.

• Precision-anchor threshold raise (closes the #1207 over-merge, #1319): a
  new default auto-config rule, rule_precision_anchor_threshold_raise, that
  raises the weighted threshold to 0.9 on the precision-collapse shape
  (>= 95% of scored mass above the threshold on a name-dominated weighted
  matchkey with a strong exact identity anchor, the TF table live, and the
  threshold below 0.9). Two commit-dynamics fixes make the raise actually land
  (either alone still commits the over-merging config): the scoring-health
  unimodality (dip) gate now requires at least 30 scored pairs before a flat
  dip reads RED (_MIN_DIP_SUPPORT -- a flat dip over fewer pairs is sampling
  noise), and when the rule has fired, pick_committed rank-demotes entries
  whose config the rule's trigger still flags. Measured on the crafted #1319
  fixture: precision 0.009 -> 0.9868 at recall 1.0; NCVR results unchanged.

Fixed

• from_splink recognizes IS NOT NULL blocking guards (#1783): compound
  CustomRule blocking rules carrying trailing AND l.col IS NOT NULL guard
  conjuncts were dropped whole as unrecognized — on a 1M production dedupe the
  converted model blocked on 3 of 6 keys, costing ~28 points of pairwise
  recall. Guards on key columns are now recognized and ignored exactly
  (GoldenMatch blocking already implements the guard semantics: a null key
  component forms no block), reported as info. A guard on a column outside the
  blocking key still converts but warns (guard dropped, candidates are a
  superset of Splink's; strict=True gates on it), and a guards-only rule
  keeps the existing unrecognized-drop path.
• The #1318 TF name downweight now reaches the default (bucket) scoring
  path (#1781): the bucket backend's fast path resolved plugin scorers via a
  bare plugin.score_pair, so the per-dataset TF table behind
  GOLDENMATCH_TF_NAME_WEIGHTING (default-on) never reached
  name_freq_weighted_jw on the default path -- the flag was a silent no-op
  there. MatchkeyField.tf_freqs is now threaded through the bucket fast
  path's plugin branch (with a TypeError back-compat fallback for legacy
  plugins without the keyword); built-in scorer branches are untouched.
• The >= 100k RED-refuse gate enforces again after every controller branch:
  a latent n_rows shadow in AutoConfigController.run() (the
  suspicious-tight-blocking GREEN branch rebound the full-frame height to the
  sample height) silently disabled the REFUSE_AT_N refuse gate, so runs that
  should raise ControllerNotConfidentError on a RED config could slip
  through.