goldenmatch v3.4.0

Added

• Embeddings as first-class Fellegi-Sunter scorers. embedding /
  record_embedding fields train and score on the FS path (vectorized EM
  E-step + block scoring); they were unusable before (both EM training and the
  scalar scorer raised Unknown scorer). Model-backed scorers always run
  vectorized — GOLDENMATCH_FS_VECTORIZED=0 only affects string scorers. The
  TUI engine now routes probabilistic matchkeys through the block-scorer
  selector, so it gets native/vectorized FS too.

Fixed

• Term-frequency (tf_adjustment) now applies on the scalar FS path, matching
  the vectorized path — the same config no longer scores differently depending on
  route (a model-backed scorer or GOLDENMATCH_FS_VECTORIZED=0 forces scalar).
• Distributed and chunked lanes score probabilistic matchkeys. They
  previously dropped FS pairs silently, then (mid-cycle) failed loudly. Both
  lanes now score FS against ONE shared EMResult: the distributed driver
  trains once before dispatch (or loads mk.model_path driver-side;
  GOLDENMATCH_DISTRIBUTED_FS_TRAIN_ROWS bounds the training sample), and the
  chunked lane trains once on the first chunk. The loud
  NotImplementedError remains only for the bare scoring kernel invoked with
  no model source.
• Memory-bounded FS scoring for strategy-generated blocks. Probabilistic
  matchkeys with lsh / ann / learned / canopy / sorted_neighborhood
  blocking now score through score_probabilistic_external_blocks (one block
  resident at a time, exclude-set Arc handle built once, oversized auto-split,
  native/vectorized selection) instead of the batched scorer's all-units
  accumulation. GOLDENMATCH_FS_DEFAULT_BUCKET=0 still selects the legacy
  batched scorer and now logs a warning (documented in tuning).
• The columnar opt-in no longer demotes FS to the batched path. The
  columnar branch is structurally weighted-only, so probabilistic matchkeys
  keep the memory-bounded bucket route under GOLDENMATCH_COLUMNAR_PIPELINE=1.
• FS bucket lane auto-splits oversized blocks (default
  skip_oversized=False path) instead of scoring them whole, and the
  vectorized scorer refuses impossible dense allocations with an actionable
  error (GOLDENMATCH_FS_VEC_MAX_ELEMS).
• Missing values are unobserved evidence in FS scoring (null field values
  carry no likelihood contribution instead of folding into total disagreement);
  persisted v1 models without a training manifest are rejected on reuse.
• Multi-pass EM conditions per pair and blocking fields keep the fixed
  prior (repairs a recall collapse on near-unique blocking fields).
• Per-field blocking transform chains (BlockingKeyConfig.field_transforms);
  the Splink converter maps mixed SUBSTR blocking rules exactly.