[kvlonge]

Why

The existing scale benchmark primarily padded SQL with comments and did not exercise wide declarations, dense tests/audits, or large set-operation ASTs. Profiling a neutral reproducer also found that nested scope and argument materialization caused superlinear time and extreme peak memory during schema-aware analysis.

Changes

• pin Polyglot's lightweight validation-scope fix from perf: avoid cloning nested validation scopes polyglot#5
• use targeted native argument access during CTE, filter, and set-operation type analysis
• add a neutral semantic-density guard with 976 models, 958 tests, 1,645 audits, about 33k declared columns, 5.8 MB model SQL, and 13 MB compiled test SQL
• protect a 267-branch, 270-column set-operation shape with cold/warm phase budgets
• retain and document the existing layered, scale, and test-heavy guards separately

Verification

• uv run pytest tests/unit/src/sqlbuild/compiler/compile/_helpers/test_sql_analysis_columns.py -n auto --dist loadfile (73 passed)
• semantic-density performance guard (passed, ~45s test runtime)
• layered performance guard (passed, ~20s test runtime)
• make test (6,760 passed)
• uv sync --all-extras && make check-ci
• public-repository scanner