fix(duckdb): Make exp.DateFromParts more lenient

[VaggelisD]

Fixes #6394

Although this is not properly documented, the following query is valid in DuckDB:

D select MAKE_DATE(DATE_PART(['year', 'month', 'day'], TODAY())) AS col;
┌────────────┐
│    col     │
│    date    │
├────────────┤
│ 2025-11-24 │
└────────────┘

This is because:

• DATE_PART returns either an INT64 or a STRUCT of values, depending on the input expression:

D SELECT DATE_PART('day', TODAY()) AS integer, DATE_PART(['year', 'month', 'day'], TODAY()) AS strct;
┌─────────┬─────────────────────────────────────────────────────┐
│ integer │                        strct                        │
│  int64  │ struct("year" bigint, "month" bigint, "day" bigint) │
├─────────┼─────────────────────────────────────────────────────┤
│   24    │ {'year': 2025, 'month': 11, 'day': 24}              │
└─────────┴─────────────────────────────────────────────────────┘

• MAKE_DATE accepts either INT64s or a STRUCT which encodes those kwargs

Note: To properly represent that in the AST we'd have to infer whether the MAKE_DATE expression evaluates to an INT64 or a STRUCT, but since this low ROI we instead store that in the first argument year.

Docs

DuckDB MAKE_DATE | DuckDB DATE_PART

[georgesittas]

Should we just create a new arg named date_part_struct or something?

[VaggelisD]

Should we just create a new arg named date_part_struct or something?

Not sure whether it's straightforward, we could have columns representing either INT/STRUCT, DATE_PART representing either type etc. For instance:

D SELECT MAKE_DATE(col) FROM (SELECT 20416 col);
┌────────────────┐
│ make_date(col) │
│      date      │
├────────────────┤
│ 2025-11-24     │
└────────────────┘


D SELECT MAKE_DATE(col) FROM (SELECT {'year': 2025, 'month': 11, 'day': 24} AS col);
┌────────────────┐
│ make_date(col) │
│      date      │
├────────────────┤
│ 2025-11-24     │
└────────────────┘

[georgesittas]

Ah, I wasn't aware that a single integer value works as well. Looks good then, ship it.