v0.2.1

What's changed

Fix

• Replaced naive whitespace tokenization in compute_minhash with word bigrams (unigrams + bigrams) as default — detects word-order and co-occurrence drift, not just bag-of-words changes
• Added char_shingle tokenizer option for structured strings and IDs (3-char shingles)
• Kept whitespace as a legacy opt-in for backwards compatibility

Consistency guards

• build_baseline raises ValueError if minhash records with mixed tokenizers are merged
• compute_signals raises ValueError if current sketch and baseline were built with different tokenizers

Chore

• Single-sourced version in lakesense/__init__.py via hatchling dynamic; removed duplicate from pyproject.toml
• Added .claude/settings.local.json to .gitignore

Breaking change

Existing minhash baselines stored under the old whitespace tokenization must be rebuilt — comparing them against new word_ngram sketches will raise a ValueError.