[kvlonge]

Why

• Compiles with many missing descriptions slowed down quadratically. Each source or seed without a description re-read its whole YAML file to find its line, so a 10k-model project with 13,150 P010 errors spent about 200 s in attachment.
• A comment before UNION ALL in an __expected__ SQL-test fixture raised a false P001 ("branch 2 does not match branch 1"), because the comment stayed attached to the last projection.

Changes

• Index each YAML file's entry lines once, cached by contents, for P010 locations and the warehouse source-table lookup. Output is identical.
• Trim comments around set-operation branches and projections in expected fixtures, on both the native and Python paths. The Python fallback splitter follows the adapter's lexical syntax, and Annotated branches are unwrapped.

Verification

• Same-machine medians, wall (main → fix): 1k 8.3 → 5.4 s, 5k 68 → 20 s, 10k 232 → 41 s, with identical diagnostics JSON.
• Scaling guard (500 vs 2,000 models without descriptions) fails on main and passes here; exact-output CLI test; entry-line unit tests and a randomised comparison with the old lookup.
• DuckDB E2E with commented expected fixtures (line, block, nested), plus negative controls that still report P001.
• One independent review; make test, make check-ci.