Normalise combine()'s raw-module array_package and fix correlated add/subtract uncertainty leak

[mwcraig]

Part of #971; fixes #982.

Strict count on top of main 6724c8e: 31 failed → 31 failed (identical failure list; no target from #971 was on the strict failure list). numpy, jax, dask: no regressions; strict/numpy/jax/dask each gain the 8 new regression tests as passes.

combine()'s raw-module array_package (#982). combine() tried to normalise array_package with array_api_compat.array_namespace(), which only accepts arrays, not modules, so for a raw module (e.g. numpy or dask.array) it always raised TypeError and the except branch passed the module through unnormalised instead of converting it. Combiner.__init__ already normalises its own xp argument with array_api_compat.array_namespace(xp.asarray(0)) (fixed for #976); this PR does the same in combine() so the two entry points agree on what a caller may pass, and updates the array_package docstring to say so. While adding the regression test for the from_array() got an unexpected keyword argument 'device' failure described in #982, I found that combine()'s subsequent xp = array_api_compat.array_namespace(ccd.data) (which runs right after the first image's data is converted, whether or not array_package was itself normalised) already re-derives a correct namespace from the resulting concrete array — I could not reproduce the specific crash from #982 through combine() on current main, including with a masked, uncertain, scaled, tiled combination that exercises every device= call in the function. The crash went inert with #995: the conversion is now xp.asarray(_native_numpy(...)) with no device=/dtype=, which a raw dask.array handles, and xp is then re-derived from the converted array. This PR still closes the underlying gap between intent and code (and between Combiner/combine), it just is not observed to fix a live crash on current main.

Arrays as array_package. The old code (and docstring) also accepted an array standing in for its namespace; a third commit keeps that working (is_array_api_obj check before calling .asarray(0)) and covers it in the regression test.

Correlated add/subtract uncertainty leaks to NumPy in the wrapper. _ArrayAPIPropagationMixin._propagate_add/_propagate_subtract delegated correlated-uncertainty math to astropy's _VariancePropagationMixin._propagate_add_sub, whose correlation term (2 * correlation * np.sqrt(this * other)) is hardcoded to NumPy. That is invisible when uncertainty_correlation is 0 (the term is never evaluated), but addition/subtraction with a nonzero correlation on strict fails with TypeError: Expected Array or Python scalar; got numpy.ndarray once the NumPy result is combined with an array-API array. #993 already fixed the same class of leak for _propagate_multiply_divide; this PR mirrors it by adding an array-namespace _propagate_add_sub to the mixin and calling it instead of super()'s, extending test_wrapped_arithmetic_correlated_uncertainty from ["multiply", "divide"] to ["add", "subtract", "multiply", "divide"].

Follow-up note. Combiner(xp=...) accepts an explicit namespace for CCDData input, but combine(array_package=...) is documented as, and remains, ignored when ccd_list is already a list of CCDData objects — array_package only applies when reading from filenames. This asymmetry is unchanged by this PR; flagging it as a possible follow-up.

🤖 Generated with Claude Code

https://claude.ai/code/session_01S36ZzAAVXVm32vuTdtCQME

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 97.76%. Comparing base (f61e08e) to head (46b8c71).
⚠️ Report is 3 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #997      +/-   ##
==========================================
+ Coverage   97.75%   97.76%   +0.01%     
==========================================
  Files           9        9              
  Lines        1780     1788       +8     
==========================================
+ Hits         1740     1748       +8     
  Misses         40       40              

Flag	Coverage Δ
dask	96.91% <100.00%> (+0.01%)	⬆️
jax	96.96% <100.00%> (+0.01%)	⬆️
numpy	97.65% <100.00%> (+0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.