Array API: dead code in cosmicray_median clobbers masked-input handling

[mwcraig]

Problem

In the array branch of cosmicray_median, the masked-input handling is dead code — data = ccd.data is immediately overwritten by the unconditional data = xp.asarray(ccd) on the next statement:

ccdproc/ccdproc/core.py

Lines 2030 to 2045 in 9d25eee

	# Only look at the data array, guessing that if there is a .mask then
	# there is also a .data.
	if hasattr(ccd, "mask"):
	data = ccd.data
	data = xp.asarray(ccd)
	if error_image is None:
	error_image = xp.std(data)
	elif not isinstance(error_image, float):
	if not _is_array(error_image):
	raise TypeError("error_image is not a float or ndarray.")
	# create the median image
	marr = ndimage.median_filter(data, size=(mbox, mbox))

So a masked array passed in is not handled the way the comment describes, and what xp.asarray does with an np.ma.MaskedArray (silently dropping the mask) is backend-dependent.

Notes

• Decide the intended behavior for masked input and implement it deliberately.
• The same function calls scipy.ndimage.median_filter on the (possibly non-numpy) data a few lines later, which is covered by the CPU-only-operations policy issue Array API: decide a policy for inherently CPU-only operations #935.

---

Found during a review of the array API implementation from #885; follow-up to #909 / #912.