[jasperblues]

Closes #434. Port of liberation-data/drivine4j#616 (drivine4j PR #627), as liberation-data/drivine#643 ports it. Not to merge until train 2026.11.1 is published, and not before drivine#643.

text_query.py split a full-text query into words with \W. Python's \w is what str.isalnum holds, plus _. So it kept ², ½ and ⅻ in a word, and split a Hindi word at its vowel sign, which isalnum doesn't hold.

A term's expression reads the text with [[:alnum:]_], which on Postgres, and through sqlite_patterns (Java's \p{Alnum}), is Unicode's Alphabetic plus decimal digits. So किताब never matched, and ² stayed in a word where the text, and Neo4j's analyzer, end one at it.

A query's words are now split by _gap(): not Alphabetic, Nd or _, built from sqlite_patterns.property_ranges("Alnum") and class_text. It is built on the first query rather than on import, because it reads every code point (about 85 ms).

Goldens: unchanged.

Tests

• tests/integration/test_bm25.py, and its sync copy, on Neo4j, Postgres, Postgres indexed, SQLite and Oracle: किताब is found, and "m²x" finds "an m² x here". many's pinned score is 0.0989 for the two new seed texts.
• tests/unit/test_text_query.py:
  • the words of किताब, bⅻb, m²x, 1½2, ٣٤ and m\²x;
  • SQLite's [[:alnum:]_] and a query's words hold the same characters.

Verification

• Two fresh-reviewer passes. Pass 1 added the [[:alnum:]_] pin and the docs notes (docs/02-connecting.md, docs/07-schema.md); pass 2 came back clean.
• Unit tests: 2,484 passed. ruff and pyright are clean, and the sync code is current.
• Integration, on the test box with Oracle: the BM25 tests, their sync copy, the search tests, and the SQLite and Postgres store tests. 223 passed, and agreement with Neo4j is unchanged at 17/21.