[kvlonge]

Why

Semantic validation rebuilt work that analysis and its own reference pass had already done: a scope tree per SELECT for the semantic checks, SELECT copies to strip WITH clauses, repeated whole-tree walks, and a deep comparison of whole derived-table queries per column lookup. On a 3,000-model project this roughly tripled native analysis time.

Changes

• Per-SELECT semantic checks run on the reference-validation scope tree instead of building a scope per SELECT. A SELECT inside a CTE body now sees the CTEs defined before it, which removes a false E201 when a CTE shadows a table of the same name (covered by a DuckDB-checked test). DML and other statements keep the standalone path.
• One traversal collects set-operation ORDER BY findings, SELECT counts, projection stars and lambdas; lambda binding is skipped when there are none.
• Validation scope views take ownership of their queries instead of copying them; dotted-column normalisation is skipped when every qualifier names a source.
• Derived-table child-scope lookups are cached in lineage and nullability analysis.
• Version 0.13.8.

Verification

• cargo test --lib -p polyglot-sql: 1343 passed; make test-rust-feature-gates and make check-consistency pass; query-analysis, regression and type-truth suites pass (the existing analyze_query_set_operation_name_alignment_and_cast_consensus_454 failure on main is unchanged).
• Downstream, diagnostics are identical on a semantic corpus and generated 3,000-model projects, with native analysis roughly 40% faster.