[kvlonge]

Why

SQL tests were split into CTEs twice: by a Python extractor before macro expansion (with a Polyglot fallback) and a native one after, which deferred to Python whenever adapter rules differed. They disagreed often, and the native one panicked on __udf followed by a no-break space.

Changes

• One native extractor handles both passes on every engine under the adapter's lexical rules. The Python extractor, its Polyglot fallback and the deferral are removed. Expanded-test cache bumped to v3.
• CTE-name errors are checked once on the authored block, before declaration scoping and expansion, and point at the exact file:line:column. Discovery records each block's authored span, so locations survive dedent, tabs and cases.
• Behaviour changes, all engines:
  • CTE names must be unquoted ASCII identifiers.
  • Keywords match ASCII case-insensitively only.
  • Helper and expected CTEs may not call __udf/__table_fn (now also in macro tests).
  • Each malformed reference call in them is its own located P012.
  • Expected projections need an alias unless they are a bare or qualified column. Implicit aliases (1 order_id) count, but a trailing keyword or operand does not (a IS NULL, CASE … END).
  • Parenthesised or WITH expected branches are rejected; t.* and DISTINCT * count as SELECT *.
  • Duplicate actual/expected CTEs report as duplicate CTE names.
  • The no-break-space panic is a normal error.
• Testing docs and skill copy describe the rules.

Verification

• Differential fuzz vs the previous passes (20,000 tests, 7 adapter syntaxes): every mismatch is attributed to an agreed class, 0 unexplained, no panics.
• Rust rule and authored-span tests; offset-mapping tests; DuckDB E2E on python, native and native-preview (21 cases, including exact locations and repeated P012); failure-corpus cases.
• cargo test/clippy/fmt (1.99.0), compiler differential 224/224 (preview) and 218/218 (shipped), e2e harness, make check-ci, make test.