[ravenac95]

After a parenthesized query in FROM, the parser folded any following UNION/INTERSECT/EXCEPT (and a trailing LIMIT) into the derived table:

SELECT 1 AS s, * FROM (SELECT x FROM a) UNION ALL SELECT 2, x FROM b
-- was reprinted as
SELECT 1 AS s, * FROM ((SELECT x FROM a) UNION ALL SELECT 2, x FROM b)

That changes the query: the outer projection is applied to the whole union. Found in production as TYPE_MISMATCH: UNION query has different number of fields: 6, 7. qualify_tables' set-operation-subquery normalization undid the wrapper, but only when the outer query is a bare SELECT *, which hid the bug in the common case. Upstream 0.14.1 has the same behaviour.

Change

• The parenthesized-query branch no longer parses a set operation after its closing paren; that set operation belongs to the enclosing query.
• Inside nested parens (FROM ((SELECT 1) UNION (SELECT 2))), the nested-paren branch already parsed the set operation. It now wraps it in the same Subquery the inner branch used to return. Its ORDER BY/LIMIT/OFFSET handling relies on that shape: without it, SELECT * FROM (((SELECT 1) UNION SELECT 2) ORDER BY x LIMIT 1 OFFSET 1) moved the modifiers outside the outer parens (caught by sqlglot_identity).
• qualify_tables::test_can_disable_set_operation_wrapper_normalization asserted the misparsed shape with the normalization turned off. It now expects the correct one, with no wrapper left to normalize.

Tests

• tests/unaliased_derived_table_set_operation.rs (new):
  • the bug: projection-plus-star, inner LIMIT with an outer LIMIT, a comma join, and EXCEPT, in Trino and generic;
  • parenthesized set operations that must keep their shape, including the ORDER BY … LIMIT … OFFSET one above.
• With the sqlglot fixtures extracted: sqlglot_transpilation, sqlglot_identity, sqlglot_dialect_identity, sqlglot_parser and sqlglot_transpile all pass in full. cargo test --lib (1343), identity_roundtrip and query_analysis pass.
• kuma's query-rewriter suite, including its column and lineage fixtures, passes against main with this PR, Reprint Trino/Presto TIMESTAMP(x) as CAST(x AS TIMESTAMP) #16 and Keep Presto/Trino two-argument MAX(x, n)/MIN(x, n) as aggregates #17 combined.

🤖 Generated with Claude Code

https://claude.ai/code/session_017CxHoCFbTWJA9bcZWC9ydu