hgrecco / pint-pandas / 37535413640

90%

push

github

Fix several xfailed/skipped extension tests (#315)

* Fix several xfailed/skipped extension tests

- Dispatch np.repeat to PintArray.repeat, as pint does not implement it
- Make PintArray.__array__ follow numpy 2 copy semantics (raise on copy=False)
- Override to_numpy so a read-only PintArray doesn't yield a read-only
  copy (pandas wrongly treats the object conversion as a view)
- Fix broken test_unstack / test_groupby_extension_agg overrides that never
  ran the upstream tests, and narrow their xfails to the cases that fail
- Remove stale skips/xfails for invert, insert_invalid, repeat,
  array_interface_copy and readonly_propagates_to_numpy_array_method

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Remove np.repeat CHANGES entry

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Remove __array__ and to_numpy CHANGES entries

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Fix test fixtures ignoring the pandas subdtype

The data, data_missing and data_for_twos fixtures passed the pandas dtype
as PintArray's `copy` argument, so the subdtype was inferred from the
values: the float64 and object variants actually tested Float64 and Int64.

Testing the real subdtypes exposed these bugs, now fixed:
- PintType stored subdtype as given (str, np.dtype or NumpyEADtype), and
  np.dtype("float64") != NumpyEADtype("float64"); normalise it to the
  dtype pd.array produces
- ufunc results and map re-inferred the subdtype instead of keeping it
- _cast_pointwise_result returned None for Python bool results (object
  subdtype comparisons), breaking Series.combine

Remaining cases are skipped (object magnitudes accept any value) or
xfailed (array arithmetic doesn't promote the result subdtype).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Restore xfail_pandas_dev on test_arith_series_with_array

The merge from master left master's xfail_pandas_dev override directly
above this branch's override, which silently replaced ... (continued)

16 of 19 new or added lines in 1 file covered. (84.21%)

642 of 709 relevant lines covered (90.55%)

1.81 hits per line