fix: handle infinite relative tolerance for timedelta values

[winklemad]

Summary

pytest.approx() multiplies a timedelta by the relative tolerance while constructing the comparison. When rel=float("inf"), Python raises OverflowError because timedelta has no infinite representation.

Reject infinite relative tolerances alongside the existing negative and NaN validation, producing a clear ValueError before multiplication. Numeric approximate-comparison behavior remains unchanged.

Validation

• tox -e py312 -- testing/python/approx.py::TestApproxDatetime::test_timedelta_rel_must_not_be_infinite
• tox -e py312 -- testing/python/approx.py
• tox -e py312
• tox -e linting
• git diff --check

[Pierre-Sassoulas]

Thank you for working on this ! I'm curious about what others expect from pytest in this case. Silentely succeed, or raise ? And from their linter ? Probably raise a warning as it's a constant condition or an underhanded way to check typing.

[nicoddemus]

I think this should be an improvement rather than a bugfix.

[Pierre-Sassoulas]

LGTM

[winklemad]

Renamed it to 14754.improvement.rst — thanks!

[winklemad]

Heads up: the ubuntu-py315 red is TestPDB::test_pdb_unittest_postmortem timing out in pexpect (1 failed, 4331 passed) — looks like the usual pdb flake rather than anything from this branch, which only renamed the changelog file. Happy to rebase if you'd rather see it green first.