goldenmatch v2.5.0

Added

• MemoryStore: Postgres backend for Learning Memory (#1338). MemoryStore(backend='postgres', connection=<dsn>, table_prefix='goldenmatch_') and MemoryConfig(backend='postgres', connection, dataset=<tenant>, table_prefix) persist dedupe corrections + learned adjustments in Postgres, isolated per tenant: corrections filtered by dataset, adjustments keyed (dataset, matchkey_name), and MemoryLearner(dataset=…) so learn() never pools corrections across tenants. Adds LearnedAdjustment.dataset and MemoryConfig.table_prefix (regex-guarded). Reuses the existing postgres extra (psycopg[binary]) — no new hard deps. SQLite remains the default; every existing call path, signature default, and behavior is unchanged (the dialect refactor is behavior-preserving; full SQLite memory suite green).

Changed

• Auto-config: name_freq_weighted_jw now downweights agreements on high-frequency name values using a per-dataset frequency table (data-driven, applied across the whole score range), so identical common surnames (e.g. two "Smith") score below identical rare surnames - a higher matchkey threshold then separates same-name strangers from true matches (#1207, PR2a). Default-on; kill-switch GOLDENMATCH_TF_NAME_WEIGHTING=0 restores the static-census behavior. Validated by the CI accuracy gates (#528/DQbench/Febrl/NCVR); this is an accuracy change, not a measured-local win.
• Auto-config: per-identifier blocking union on null-sparse multi-source data (#1207, PR1).
  When no single exact key clears the null-rate gate, build_blocking now emits a
  per-identifier blocking union (one pass per strong identifier + name/geo) instead of a
  single high-null compound that capped recall. Default-on; no behavior change when a
  low-null single exact key exists. Strong-id passes use a non-null scale gate (the runtime
  blocker filters null block keys) with a #876 perfect-surrogate exclusion. Measured blocking
  recall 1.0 vs name-only 0.004 on a planted-dup fixture, no regression on the auto-config suite.
  Scope: auto_configure_df switches to learned blocking at total_rows >= 50_000, so the
  union applies below that threshold (or when learned blocking is off) today; the >=50k
  learned-blocking interaction is a tracked follow-up.