[kvlonge]

Why

SQLClean 0.9.0 lays out clauses, parenthesised groups and windows the way SQLQuack 0.6.0 does, so the two move together. SQLQuack 0.6.0 also gives each output column its expression's byte range, which column lineage source highlighting needs next.

Changes

• SQLClean moves to 0.9.0 (6e18af66) and every SQLQuack pin to 0.6.0 (8d28aa4c). cargo tree -d --depth 0 lists no sqlquack-* or sqlclean duplicate; no other lockfile entry changes.
• Test expectations follow SQLQuack 0.6.0: a one-line JOIN LATERAL … ON TRUE, the parse error Expected an expression after LIMIT, and a bare * now sits at its own position in the rule SQL tree, so a custom rule's unqualified * projection reports the star instead of its enclosing select or call.
• DuckShop edge case: not applicable. The user-visible change is formatter layout and a reported position, both output formatting.

BEGIN_COMMIT_OVERRIDE
feat(format): lay out clauses, parenthesised groups and windows as SQLClean 0.9.0 does
fix(rules): a custom rule's unqualified * projection reports the star's own line and column
END_COMMIT_OVERRIDE

Verification

• Same machine, 3,000 models, 3 alternating runs, base = origin/main build (load 2.3–3.5 on 8 cores):
  • dense cold: wall 7.93 → 7.92 s (0.998×), CPU 24.89 → 24.76 s (0.995×); warm 1.071× wall; one-model edit 1.021× wall, 0.987× CPU.
  • fresh cold: wall 6.81 → 6.79 s (0.996×), CPU 20.42 → 19.98 s (0.979×); warm 0.992×; edit 1.027× wall, 1.062× CPU.
• Diagnostics before and after are identical by code on DuckShop (0), the deploy project, the three website examples and the three playground templates. Formatting the website examples still preserves their compiled contract (4 passed).
• cargo test -p sqlbuild-rules and -p sqlbuild-analysis --lib (353 passed); format, rules, lint and rule-engine Python suites; compiler differential with CI's coverage gates (391 projects as expected).
• make check-ci passes.