bitermplus: v0.11.0

Bug fixes

• Coherence: last document's words were silently skipped due to a wrong boundary (w_rj = W
  instead of n_dw.indptr[d+1]). Could also cause out-of-bounds memory access when vocabulary size
  exceeded total non-zeros. Now correct for all documents.
• BTMClassifier mixed-case input: documents with capital letters or punctuation produced
  all-zero topic vectors because raw whitespace splitting did not match the lowercased,
  token-pattern-filtered vocabulary. fit() and transform() now tokenize via the fitted
  CountVectorizer's analyzer.
• set_params(vectorizer_params=None): crashed on the next fit() with TypeError. The
  constructor no longer rewrites None to {}.
• BTM.transform(): invalid infer_type now raises ValueError listing valid options instead
  of silently producing a memoryview assignment error.
• seed=0: previously fell back to time(NULL) (second-level granularity, so two models
  created in the same second got identical seeds). Now uses OS entropy via
  np.random.default_rng(None).

Performance

• Coherence: removed a redundant inner loop that re-ran the same idempotent check M times per
  word. ~20× faster on the coherence hot path with default M=20.

API

• BTMClassifier.fit() and fit_transform() now accept a verbose parameter (default False).
  Previously the tqdm progress bar was hardcoded on with no way to suppress it. Note: training is
  now silent by default; pass verbose=True to restore the progress bar.

Internal

• Removed dead code (_parse_words in get_vectorized_docs, unused time cimport).
• Added regression tests for the coherence boundary, mixed-case input,
  set_params(vectorizer_params=None), and invalid infer_type.

Full Changelog: v0.10.0...v0.11.0