v0.4.3

[0.4.3] - 2026-08-13

Added

• Corpus-backed, syntax-agnostic v3 dictionary fragment discovery with exact
  longest-match usage accounting
• Exact v3 dictionary objective scoring for overlapping whole entries and
  fragments
• Optional split v3 dictionary-matrix architecture with an independent
  HIER4/HIER8 dictionary namespace, diagonal paths, nested nodes, runtime
  loading, static code generation, and backward-compatible shared tables
• Regression coverage for fragment/whole coexistence, fragment-for-whole
  swaps, split dictionary matrices, generated/static round-trips, and loader
  validation

Changed

• V3 fragment refinement now uses deterministic dictionary-first admission:
  payload training ranks by active non-overlapping use count inside isolated
  dictionary gain, while embedded/amortized objectives may use potential counts
  for shadowed-entry swaps
• Only the strongest positive dictionary-gain class is exact-v3 evaluated in a
  refinement round; an improving fragment is committed immediately and corpus
  proposals are rebuilt against the new longest-match parse
• V3 training now evaluates shared and split dictionary-matrix layouts and
  selects the lower requested objective instead of forcing either architecture
• Dictionary matrix DIRECT cells address dictionary indices directly, so
  split tables avoid duplicating dictionary entries in the main symbol table
• Trainer diagnostics now distinguish round-summed gain-positive proposal
  observations from the small number of candidates actually exact-scored
• V3 dictionary reports now show final effective uses and a clearly labelled
  isolated admission gain instead of stale analyzer gain values

Fixed

• Removed the pathological v3 fragment-refinement behavior that exact-scored
  hundreds of thousands of corpus proposals with repeated full-corpus scans
• Prevented shadow-only occurrences from inflating payload fragment admission
  counts
• Escaped control and non-printable bytes in fragment diagnostics so entries
  such as newline-prefixed log fragments cannot break trainer output
• Preserved old shared v3 table loading while extending reserved v3 metadata for
  split dictionary matrices

Verified

• Full test suite passes after a clean rebuild
• Release build path passes make, make test, make package, and
  make examples
• On trainings/extra/logs_corpus.txt (1,886,013 bytes), refinement processes
  only 3 exact candidates across 3 rounds, adds 2 fragments, raises dictionary
  byte coverage to 84.9%, and reduces the payload objective from 4,772,661 to
  3,725,400 bits
• Shared/split selection remains corpus-dependent: split wins before fragment
  refinement on the logs corpus, while the refined distribution correctly
  selects the lower-payload shared HIER8 layout