[kvlonge]

Why

• A set operation after a parenthesised subquery in an expression was attached to the subquery: WHERE a.x < (SELECT …) UNION ALL SELECT … parsed as a.x < ((SELECT …) UNION ALL SELECT …). The same happened in HAVING, JOIN ON, BETWEEN and the SELECT list, so downstream tools reported false column-count errors and mis-formatted the query.
• Scope building for set operations copied the whole left branch at every level, which made validation and analysis quadratic in the number of branches.

Changes

• fix: a bracketed query takes a following set operation only when it is the first operand inside another bracket (x IN ((SELECT 1) UNION (SELECT 2))) or when CTAS opts in.
• perf: nested plain set-operation branch scopes store a shallow node. Lineage and nullability get the full branch query through the new set_operation_branch_queries, and Scope::output_columns follows the left branch.

Verification

• Regression tests across 7 dialects (parse tree, regenerated SQL, validation) for the misparsed shapes and the unchanged bracketed forms.
• Scaling tests: linear scope size and a timing ratio for 100 vs 400 branches; identical diagnostics for column-count mismatches at the start, middle and end of long chains.
• Differential checks: old vs new on 330+ query × dialect cases, byte-identical, apart from the intended fixes.
• Crate suite passes apart from failures that also occur on main.