v1.0.0

This is a substantial modernization release with breaking API and dependency changes. Pin ucluster<1.0 if you need the old behavior.

Removed

• TransformerCluster has been removed. FuzzyClusterer is now the transformer-based clusterer; there is no separate class. Update imports: from ucluster import TransformerCluster → from ucluster import FuzzyClusterer.
• tf-cluster VisiData command has been removed (redundant with fuzzy-cluster).
• FastText is no longer a dependency. The "train word vectors from scratch on each run" pipeline is gone. If you depended on per-corpus FastText training, pin ucluster<1.0.

Changed

• FuzzyClusterer now uses the paraphrase-multilingual-MiniLM-L12-v2 sentence-transformer model instead of training FastText from scratch. The model is downloaded from Hugging Face Hub on first use (~470MB) and cached. First run now requires network access.
• FuzzyClusterer.outlier_probabilities() semantics have changed. It previously returned hdbscan.HDBSCAN.outlier_scores_ (the GLOSH algorithm — per-point outlier-ness even for in-cluster members, effectively unbounded). It now returns 1 - probabilities_ (bounded in [0, 1], with 1.0 for noise points). The numbers are not comparable to the old GLOSH scores.
• FuzzyClusterer.__init__ signature changed: the dims parameter (FastText embedding dimension) is gone. A new model parameter accepts a sentence-transformer model name or path.
• HDBSCAN now comes from sklearn.cluster.HDBSCAN (upstreamed in scikit-learn 1.3) instead of the standalone hdbscan package. Same algorithm, no separate Cython wheel to compile.
• VisiData dependency bumped from ^2.11 to >=3.0. The plugin works on both 2.x and 3.x runtime, but the lockfile pins 3.x.
• NLTK punkt resource is now downloaded as punkt_tab (NLTK 3.9+ split). Existing ~/nltk_data/tokenizers/punkt installs will trigger a one-time re-download.
• Python floor raised from ^3.10 (which was equivalent to >=3.10,<4) to a plain >=3.10. No upper bound.
• Packaging migrated from old-style Poetry ([tool.poetry]) to PEP 621 ([project]) with hatchling as the build backend. Use uv sync instead of poetry install.

Removed installation pain

• The Mac M-series + OpenBLAS + Conda dance is gone. uv sync works out of the box on Apple Silicon and x86 because scikit-learn, torch, and sentence-transformers all ship pre-built wheels.
• env.yml (Conda environment) has been deleted.
• poetry.lock has been deleted; uv.lock takes its place.

Plugin

• VisiData plugin bumped to 2.0.0 to reflect the removed tf-cluster command.

v0.3.2b3

Update dependencies

Initial release

v0.1.0

README fixes