Keep combine result masks in the array namespace

[mwcraig]

Part of #971 (section 1, "combine methods pass mask= to CCDData(...)" — the largest remaining bucket on the strict job).

Combiner.average_combine, median_combine and sum_combine built their result with CCDData(..., mask=mask), and combine() assigned ccd.mask = ... at several points, so astropy's CCDData.mask setter (np.asarray(value, dtype=np.bool_)) ran on every result. On backends whose arrays cannot be converted to NumPy (a non-default array-api-strict device, a GPU) that raises; on jax/dask it silently handed back a NumPy mask on non-NumPy data.

Changes

• The three *_combine methods assign combined_image._mask = mask after construction, the same workaround ccd_process already uses (core.py:381), with the same TODO.
• combine() does the same at its own mask sites, and coerces the template image's mask into the data's namespace and onto its device the way Combiner.__init__ does — a CCDData that received its mask through the setter always carries a NumPy mask, so without this the tile loop fails with "Multiple namespaces for array inputs" once the tiles' masks are no longer NumPy.
• The tile loop's ccd.mask.copy() (NumPy-only; it only worked because the mask was always NumPy) is replaced by xpx.at(...).set(..., copy=True).
• Test bodies that were previously never reached on strict are made array-API clean: positional axis, xp.sqrt(3) → math.sqrt(3), mask.sum() → xp.count_nonzero.

Results

• Strict job locally (tox -e strict): 63 → 54 failed, 418 → 427 passed, no new failures.
• numpy full suite, and dask / jax (JAX_ENABLE_X64=1) test_combiner.py: all green.
• The rest of the original 12-test bucket now stops at the next layer: median_combine uses the default sigma_func → astropy.stats.median_absolute_deviation (Array API: Combiner.sigma_clipping and the default sigma_func densify via astropy.stats #929, section 3), and test_writeable_after_combine writes a device1 array to FITS, which belongs with the section-2 "write a NumPy copy" tests.

🤖 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 97.44%. Comparing base (5399fc2) to head (280f32f).
⚠️ Report is 6 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #992   +/-   ##
=======================================
  Coverage   97.44%   97.44%           
=======================================
  Files           9        9           
  Lines        1760     1761    +1     
=======================================
+ Hits         1715     1716    +1     
  Misses         45       45           

Flag	Coverage Δ
dask	96.58% <100.00%> (+<0.01%)	⬆️
jax	96.63% <100.00%> (+<0.01%)	⬆️
numpy	97.33% <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.