[kvlonge]

Why

FAST column lineage still ran in Python and called the Polyglot wheel's parse_one for each model without compact facts, such as models with sql_analysis false. This PR builds the same lineage natively under native-preview. The python and native engines are unchanged.

Changes

• sqlbuild-analysis::lineage: star expansion and the parse fallback, using the crate parser. It ports the reference matching, alias maps, confidence and transform rules, union branches, and the pythonize-shaped to_dict walk.
• The stage runs on the compile catalog's deep-stack analysis pool. Set-operation chains are serialized once and walked iteratively, so deep UNION ALL chains stay linear in cost and depth.
• A model is returned to Python, which builds it exactly as before and records a {"kind","site"} deferral, when:
  • its dialect is not compiled into the crate;
  • the parser panics;
  • the project has no analysis catalog.
• Parse errors are logged with Python's exact debug event.
• build_fast_column_lineage binding and stub, plus lineage failure-corpus cases.

Verification

• Rust unit tests, including a 512-branch chain at the parser guard limit, started from a 256 KiB thread.
• Parity property test, Python vs native, on generated projects with and without compact facts, across 4 dialects and model selections, requiring a minimum number of native answers. Further tests cover parse-error logging and the deferral records.
• DuckDB CLI tests on all three engines comparing compile --json and lineage --mode fast, including a 480-branch UNION ALL beside a second opted-out model. That project segfaulted before this fix.
• Random-SQL fuzz with deep set-operation chains: no mismatches.
• cargo test --workspace --all-features, clippy and fmt 1.99.0, make compiler-differential with CI flags (fast_columns wheel calls 204 to 0, no deferrals), -shipped, make test, make check-ci and the hygiene scan all pass.