subtract_dark(scale=True): device-propagation bug misreported as a unit mismatch under array-api-strict

[mwcraig]

subtract_dark(..., scale=True) fails under CCDPROC_ARRAY_LIBRARY=array-api-strict with a UnitsError, but the units involved are identical ('adu' == 'adu') — the message is misleading. 4/9 parametrized cases in test_subtract_dark fail: exactly the four with scale=True (explicit_times/exposure_keyword don't matter).

FAILED test_subtract_dark[True-True-True]
FAILED test_subtract_dark[True-True-False]
FAILED test_subtract_dark[True-False-True]
FAILED test_subtract_dark[True-False-False]
E   astropy.units.errors.UnitsError: Unit 'adu' of the uncalibrated image
    does not match unit 'adu' of the calibration image
ccdproc/core.py:862: UnitsError

Root cause (note: the #941 table called this a "unit-mismatch/device issue" — the unit-mismatch half is wrong; it is purely a device-propagation bug). In ccdproc/core.py:855 (branch array-api-strict-triage-tooling), the scale-factor array is built without a device argument:

_master_scaled = _master_scaled.multiply(xp.asarray(scale_factor))

The test harness places fixture arrays on non-default device1 (see ccdproc/conftest.py), but xp.asarray(scale_factor) defaults to CPU_DEVICE. The elementwise multiply then raises:

ValueError: Arrays from two different devices (array_api_strict.Device('device1')
and array_api_strict.Device('CPU_DEVICE')) can not be combined.

The broad except (u.UnitsError, u.UnitConversionError, ValueError) at core.py:859 swallows that ValueError and re-raises it as a bogus UnitsError about non-matching units, masking the real problem. This is the same class of bug a CuPy user would hit as a device-to-host error.

Suggested fix: construct the scale-factor array on the same device as the data (e.g. pass device= from the master's array), and narrow the blanket ValueError catch at core.py:859 so genuine non-unit failures aren't relabeled as unit errors.

Part of the array-API triage in #941; found with the array-api-strict tooling from #937/#939. Verified at branch tip d329ac5 (2026-08-01); the true underlying exception was confirmed by reproducing the scale=True path directly, bypassing the catch-all.

🤖 Generated with Claude Code