Fix RunGroupBy on numpy 2.x with StringDtype meta columns

[benmsanderson]

Summary

Fixes #318.

On Python 3.12 with pandas 3.x, string-valued meta columns of a DataFrame round-tripped through MultiIndex come back as pandas.StringDtype rather than object. RunGroupBy.__init__ called np.issubdtype(StringDtype, np.number) directly, which numpy 2.x rejects with:

TypeError: Cannot interpret '<StringDtype(storage='python', na_value=nan)>' as a data type

This blocks downstream callers of ScmRun.groupby (and therefore ScmRun.convert_unit and anything else that uses groupby internally) under the newer Python / numpy / pandas stack.

Fix

Routes the numeric-dtype check through a new module-level helper _is_numeric_dtype that wraps np.issubdtype in a try/except. Semantically a dtype numpy cannot classify is not a numpy numeric dtype, so False is the correct fallback.

def _is_numeric_dtype(dtype):
    try:
        return bool(np.issubdtype(dtype, np.number))
    except TypeError:
        return False

No change in behaviour on Python 3.11 / numpy 1.x (where np.issubdtype would have returned False anyway for non-numeric dtypes).

Test

tests/unit/test_groupby.py::test_groupby_with_string_extension_dtype exercises ScmRun.groupby directly with a MultiIndex-built ScmRun whose meta columns end up as StringDtype on pandas 3.x. Pre-fix the test raised TypeError; post-fix it returns a single group as expected. The test is deliberately narrower than the original failing call path (convert_unit), because convert_unit hits a separate pandas 3.x compat issue (ValueError: assignment destination is read-only in apply_units) that is out of scope for this PR.

Environment that reproduces #318

• Python 3.12.12
• scmdata 0.18.0
• numpy 2.4.6
• pandas 3.0.3

[mikapfl]

Thanks, looks good to me.

[mikapfl]

@znicholls @lewisjared I don't have write access here, so I can neither approve workflows nor merge this.

[znicholls]

See #318 (comment)

Happy to merge, not merge, whatever

[znicholls]

As you can see @benmsanderson the repo is relatively broken. Happy for you to go full AI on it if you want and I can just make you a maintainer then you do what you want

[benmsanderson]

Cool - that would be great! Thanks Zeb

[znicholls]

Added you with the permissions I think you'll need now. If something is missing, just ping me and I can either upgrade you to admin or give you what you need