fix(expressions): reject fractional decimal literals on integer columns

[breken-ai]

Rationale for this change

The expression parser reads an unquoted number with a decimal point as a DecimalLiteral. When that literal is bound to an int or long column, DecimalLiteral.to(IntegerType/LongType) rounds it with to_integral_value() (half-even), which changes the predicate and makes a scan return wrong rows:

tbl.append(pa.table({"x": pa.array([1, 2, 3, 4], pa.int32())}))

tbl.scan(row_filter="x > 2.6").to_arrow()["x"]  # [4]     expected [3, 4]  (bound as x > 3)
tbl.scan(row_filter="x < 2.5").to_arrow()["x"]  # [1]     expected [1, 2]  (bound as x < 2)
tbl.scan(row_filter="x = 2.5").to_arrow()["x"]  # [2]     expected []      (bound as x = 2)
tbl.scan(row_filter=GreaterThan("x", Decimal("2.6")))  # same as x > 3

The same rewritten predicate is used for partition and metrics pruning, so this also affects delete()/overwrite() with such a filter.

This PR makes the conversion raise a ValueError (Could not convert 2.6 into a int, value has a fractional part) when the decimal has a fractional part, the same way DecimalLiteral.to(DecimalType) already rejects a mismatched scale, and partition_to_py rejects fractional digits for integer partitions. Integral decimals such as 2.00 still convert, and out-of-range values still become IntAboveMax/IntBelowMin. Java has no decimal-to-integer literal conversion at all, so binding fails there too.

StringLiteral.to(IntegerType) truncates quoted values the same way (x < '2.5' binds as x < 2), but test_string_literal asserts literal("3.141").to(IntegerType()) == literal(3), so I left that path alone. Happy to follow up if you'd like it changed too.

Are these changes tested?

Yes.

• tests/expressions/test_literals.py: test_fractional_decimal_to_integral_type_raises (2.5, 2.6, -2.5, 0.1 for int and long) and test_integral_decimal_to_integral_type.
• tests/catalog/test_catalog_behaviors.py: test_scan_integer_column_with_decimal_literal appends to a real table (memory and SQL catalogs) and checks that x > 2.0 returns [3, 4] and x > 2.6 raises instead of returning [4].

On main the 11 new fractional cases fail with DID NOT RAISE ValueError. With the fix, all 37 selected tests pass. tests/expressions, tests/test_conversions.py, tests/catalog/test_catalog_behaviors.py, tests/catalog/test_sql.py, tests/io/test_pyarrow.py and tests/io/test_pyarrow_visitor.py pass. prek run --files (ruff, ruff-format, mypy, pydocstyle, codespell) passes.

Are there any user-facing changes?

Yes. A filter that compares an integer column with a fractional number now raises a ValueError instead of silently returning wrong rows. Filters with integral numbers are unchanged.

AI disclosure: this bug was found, fixed and tested by an AI coding agent (Claude) running under the breken-ai account; the red/green runs above are its local results.

[Fokko]

Great catch @breken-ai 👍

[Fokko]

@breken-ai Can you fix the merge conflict?

[rambleraptor]

This might be too late, but I had a quick idea for a test. Otherwise, looks great!

[rambleraptor]

Could you also add a test for 2.0? That should also pass

[breken-ai]

Added in c4489d7. 2.0 (and 2, 2.00, -2.0) has no fractional part, so it still converts and filters exactly:

• tests/expressions/test_literals.py::test_integral_decimal_to_integral_type is now parametrized over "2", "2.0", "2.00", "-2.0" for both IntegerType and LongType, and checks each one becomes LongLiteral(int(value)).
• tests/catalog/test_catalog_behaviors.py::test_scan_integer_column_with_decimal_literal checks x > 2.0 returns [3, 4] and now also x = 2.0 returns [2], on the memory, sql and sql_without_rowcount catalogs.

The 2.0 cases pass on both main and this branch, as you expected. The fractional cases (2.5, 2.6, -2.5, 0.1, and the x > 2.6 scan) fail on main and pass here: 11 failed / 8 passed with main's literals.py, and 19 passed with the fix. pytest tests/expressions tests/catalog/test_catalog_behaviors.py: 1307 passed. ruff check and ruff format --check are clean.

[breken-ai]

@Fokko Fixed. I rebased onto main at 068aae5. The only conflict was in tests/catalog/test_catalog_behaviors.py, where test_append_nan_to_identity_partitioned_table landed in the same spot. I kept both tests. The head is now c4489d7. It also adds the 2.0 cases @rambleraptor asked for (see the inline thread). pytest tests/expressions tests/catalog/test_catalog_behaviors.py: 1307 passed. ruff check and format are clean.