fix: preserve fallback dimensions in array finalize

[narutamaaurum]

Summary

• stop PseudoNetCDFVariable.__array_finalize__ from falling back to the dtype char when obj has no dimensions
• keep any existing dimensions tuple already attached to the target view
• add a regression test for the fallback path

Validation

• pytest src/PseudoNetCDF/test/test_core.py -k "testVar or testFromArray or testArrayFinalizeFallsBackToExistingDimensions"

Closes #156

[barronh]

Thank you for this proposed addition. Issue #156 is related to #157, which has been expanded to highlight the underlying problem case that I should have better explained at the time. I will add more description to #156.

When an PseudoNetCDFMaskedVariable ufunc is applied, an array is cast as a PseudoNetCDFMaskedVariable. In that case, there are no dimensions or PseudoNetCDF attributes. So, all the defaults get used. Even in that case, lambda: self.dtype.char is not used. In fact, I have never found a realistic case where that lambda is used... which is why it has persisted for so long.

In the PseudoNetCDFMaskedVariable ufunc case, the default () (i.e, no dimensions) is applied. This results in a zero-dimensional PseudoNetCDF object with a n-dimensional shape. A better approach would be to create (phony_dim_1, ... phony_dim_N) like the NetCDF-C library does with unnamed dimensions in an HDF5 file.

To make this PR useful, the ndimensions should be updated to match the object and the testcase should be updated to be (`ndimensions = tuple([f'phony_dim_{i + 1}' for i in obj.testcase and functionality should be updated.

-         ndimensions = getattr(obj, 'dimensions', lambda: self.dtype.char)
+         ndimensions = tuple([f'phony_dim_{i + 1}' for i in range(obj.ndim)])
+         ndimensions = getattr(self, 'dimensions', ndimensions)
+         ndimensions = getattr(obj, 'dimensions', ndimensions)

Updated test code should look something like this, and it should fail before the updates and pass after the updates.

    def testArrayFinalizeFallsBackToExistingDimensions(self):
        ta = np.array([[1, 2, 3, 4, 5]] * 2)
        var = PseudoNetCDFVariable.from_array(
            'unknown', ta, ('y', 'x'),
            units='unknown', long_name='unknown'
        )
        del var.dimensions
        viewed = var.view(PseudoNetCDFVariable)
        refdims = tuple([f'phony_dim_{i + 1}' for i in range(ta.ndim)])
        assert (viewed.dimensions == refdims)

[narutamaaurum]

Updated to address your feedback:

1. Implementation: Now generates phony_dim_N tuples matching obj.ndim as the fallback instead of an empty tuple, matching the NetCDF-C library behavior for unnamed HDF5 dimensions. Added proper precedence: checks self.dimensions then obj.dimensions before falling back to phony dims.

2. Test: Replaced with your suggested approach — deletes var.dimensions then uses var.view(PseudoNetCDFVariable) to trigger __array_finalize__ with no dimensions on the source. This test fails before the fix (gets a lambda or empty tuple) and passes after (gets ('phony_dim_1', 'phony_dim_2')).

Validation: pytest test_core.py -k testArrayFinalizeFallsBackToPhonyDimensions passes.

[barronh]

Thanks for the update, but I made a mistake in my request. The numbering in netcdf4-c starts at 0. Please just remove the '+ 1' parts and then I will approve this merge. Thanks for your patience. I had put these issues up as notes to myself to address later. I'm grateful that you jumped in.

[barronh]

thanks for the updates