[kvlonge]

Why

Native discovery matches Python discovery byte for byte on every corpus checked, and it is faster on cold and edit compiles. It still deferred to Python on Python 3.13 and 3.14, because it carried only the 3.12 Unicode table and the 3.12 inspect.cleandoc. This change makes it exact on every supported Python and makes it the default.

Changes

• Generate one str.isalnum() table per Unicode version of a supported CPython (15.0.0, 15.1.0, 16.0.0). Select the table by unicodedata.unidata_version on each call. Reproduce inspect.cleandoc per minor version: 3.13 changed it to strip spaces only. An unknown Python or Unicode version still defers to Python.
• SQLBUILD_COMPILER_ENGINE now defaults to native. python remains available as the oracle, and the CI differential still compares the two engines.
• Compiler and Rules stores remain namespaced by engine, so the first compile after upgrading is cold once. Semantic fingerprints do not change.
• Tests and docs that named the Python engine's cache paths now use the active engine's directories.
• A CI job on Python 3.13 and 3.14 runs the text oracles and a bounded engine differential.
• A new E2E test shows the unset default runs native discovery with output identical to Python.

Verification

• make rust-check, make check-ci and make test.
• Text oracles (every code point; seeded cleandoc bodies) pass on Python 3.12, 3.13 and 3.14.
• The compiler differential is identical (python vs native) on the per-PR corpus on 3.12, 3.13 and 3.14, on the dense 3,000-model corpus with and without custom rules, and on a private real project.
• Fresh-process compile timing on a real project (five alternating rounds per engine): discovery takes about 30% less time on cold and edit compiles. Whole compiles are faster or unchanged; no-change compiles reuse the previous compile in both engines.
• Focused E2E runs pass for engine switching, compile reuse, cache invalidation, the Rules cache, scope and skills.