[guocn]

Fixes #381.

Problem and change

Python's json.loads() accepts Infinity, -Infinity, and NaN by default. JMESPath classifies these Python floats as numbers, but math.floor() / math.ceil() cannot convert them to integers and raise bare OverflowError / ValueError during search().

Preserve non-finite float arguments unchanged in these two functions, one of the behaviors suggested in the issue. Finite floats still use the standard-library rounding behavior. The check is restricted to floats so arbitrary-size Python integers (e.g. 10 ** 400) do not undergo a float conversion and continue to work. Existing argument type validation is unchanged.

Tests

• Before the implementation change, python -m pytest tests/test_functions.py -q: 1 failed, 7 passed, with the new regression failing on floor(Infinity) with OverflowError.
• After: python -m pytest tests -q: 994 passed, 1 skipped, with 14 existing deprecated-string-literal warnings (Python 3.13.5).
• Regression tests cover both functions with JSON-loaded positive/negative infinity and NaN, finite positive/negative floats, zero, integers including a very large integer, and rejection of booleans/null/strings/arrays/objects.
• Focused Ruff correctness checks (E9,F63,F7,F82) and git diff --check pass.

Tests ran in an independent venv with an empty environment, automatic third-party pytest plugins disabled, and networking disabled through a network namespace. No production services were changed.

Scope / disclosure

This only defines behavior for Python's non-finite float extension; it does not change the JSON/JMESPath numeric specification, to_number(), or other numeric functions. There has not yet been maintainer confirmation of which proposed behavior is preferred; preserving non-finite values is the deliberately narrow choice here, and feedback is welcome.

Prepared with AI assistance. No independent human code review is claimed.