[kvlonge]

Why

DuckDB and PostgreSQL result types were never verified against the real engines, and many inferred types were wrong. For example, DuckDB / returns DOUBLE for every integer and decimal pair, but was inferred as an integer type. Downstream tools that compare inferred types with warehouse columns reported false schema changes as a result. Some DuckDB keyword functions (CURRENT_TIMESTAMP and others) and bare USER were also reported as unknown.

Changes

• Engine ground-truth matrices, from a generator script, stored as test fixtures: DuckDB 1.5.2 (2,769 cases) and PostgreSQL 17 (1,282 cases).
• Per-engine DuckDB/PostgreSQL type rules for arithmetic, literals, aggregates, CASE/COALESCE/NULLIF unification, date/time arithmetic, casts and common scalar functions. Unmodelled functions return no type instead of the first argument's type. Other dialects are unchanged.
• DuckDB // parses as integer division. Bare USER parses as the niladic function where the dialect defines it. DuckDB keyword time functions are registered.
• Validation accepts engine-verified DATE/TIME/TIMESTAMP arithmetic.
• Version 0.13.4.

Verification

• A replay test over both matrices: 0 mismatches. Before the fix, 1,000 DuckDB and 524 PostgreSQL cases were wrong or unknown.
• Local runs of make fmt-check, lint-rust (Rust 1.98), check-consistency, extract-all-fixtures, all four test-rust-ci-* targets and test-python, all green.