[kvlonge]

Why

Star projections over PIVOT and UNPIVOT sources compiled with the wrong columns. SELECT * FROM cte PIVOT (MAX(amount) FOR category IN ('books', 'games')) reported the pre-pivot customer_id, category, amount instead of customer_id, books, games, and its lineage pointed at the wrong inputs. UNPIVOT over a CTE reported the pre-unpivot columns.

Changes

• Upgrade the Polyglot fork to 0.12.3 (Rust crate pinned to the fork release commit, Python package polyglot-sql-chio==0.12.3). The release derives star expansion over PIVOT and UNPIVOT sources from the input relation's columns, and leaves the star unexpanded when those columns are unknown.
• 0.12.3 ships one stable-ABI (cp312-abi3) wheel per platform, so uv.lock lists five wheels instead of one per Python version.
• New integration test: compiles real projects with PIVOT-over-CTE and UNPIVOT-over-CTE models, and checks compiled columns and lineage against DuckDB's actual result columns.

Verification

• The new integration test fails on 0.12.2 and passes on 0.12.3.
• Fork: new table-driven Rust test covering pivot and unpivot shapes in both analysis modes, plus the fork core and query-analysis suites.
• make check-ci passes, as do the focused lineage and analysis-cache tests.
• A private real-project paired comparison showed identical artifacts and semantics in all runs, with no timing regression.