[kvlonge]

Why

DuckDB type inference got four expressions wrong. A cast to JSON (CAST(x AS JSON), x::JSON), to_json(...) and x AT TIME ZONE z had no inferred type. TO_TIMESTAMP(<number>) was inferred as TIMESTAMP, but DuckDB returns TIMESTAMPTZ. Callers that compare declared types with inferred ones reported these columns as unproven.

Changes

• A cast to JSON and to_json are typed JSON. A cast to JSON[] is typed JSON[].
• AT TIME ZONE and timezone(z, x) turn TIMESTAMP into TIMESTAMPTZ and TIMESTAMPTZ into TIMESTAMP. Any other or unknown operand stays unknown.
• TO_TIMESTAMP with one argument is TIMESTAMPTZ.
• A cast to TIMESTAMP(p) with p other than 6 stays unknown. DuckDB types it TIMESTAMP_S, _MS or _NS, and AT TIME ZONE over it gives TIMESTAMPTZ_NS.
• Only DuckDB's typing rules change. Other dialects are untouched.

Verification

• New tests fail on main and pass here, under both DuckDB's own parser and the shared parser. Every expected type is DuckDB 2.0.0's typeof.
• The engine-recorded type replay types one more case than before (the JSON cast) and allows one untyped case instead of two.
• Recorded cases: DuckDB is the only target, and 5,253 of 7,961 recorded cases convert. Of those, 584 cases change 786 output types: 774 from unknown to JSON (for example ARRAY_AGG, OBJECT_AGG, MAX_BY and VARIANT reads) and 12 from unknown to TIMESTAMPTZ (TO_TIMESTAMP of a number). No known type changes. DuckDB 2.0.0 DESCRIBE of each converted statement agrees on all 786.
• cargo test -p sqlquack-semantics --all-features, make check-ci
• make verify-recorded: 0 failed, and translation results and SQL sizes are unchanged.