Fix strict array-API test bodies in test_combiner/test_ccdproc/test_image_collection

[mwcraig]

Part of #971 (section 2: strict test-body fixes).

Strict count on top of main 6724c8e: 31 failed → 15 failed (469 passed/36 skipped/43 xfailed/5 xpassed → 483 passed/36 skipped/45 xfailed/5 xpassed). numpy, jax, and dask show no regressions; jax additionally improves (541 passed/7 xfailed → 548 passed/0 xfailed/0 xpassed — see the third commit).

Add numpy_copy/numpy_ccddata test helpers for strict device1 arrays. The strict suite runs on array-api-strict's non-default Device("device1"), where np.asarray() deliberately raises, but several tests need to hand data to astropy.io.fits, which requires NumPy. Adds numpy_copy() (move to the default device, then convert) and numpy_ccddata() (a CCDData with data/mask/uncertainty all converted) to pytest_fixtures.py for the write-path fixes below.

Fix strict-incompatible test bodies in test_combiner.py. _make_mean_scaler called ndarray.mean()/np.ma.average() directly; rewritten to use each argument's own array namespace, since the reference image can be plain NumPy (read from FITS) while the scaled array is a namespace array. Several tests called CCDData.multiply(), iterated over an array, or called xp.asarray() on a list of arrays — all of which either route through astropy's arithmetic wrapper (which calls np.result_type and breaks on strict) or are flatly disallowed by array-api-strict; replaced with in-namespace construction, xp.stack(), and explicit arr[i, ...] indexing. combine() only honors array_package for filename/string input and the Combiner reference built from CCDData.read() defaults to NumPy, so several tests now pass array_package=xp/Combiner(..., xp=xp) (or convert the CCDData list's .data by hand) so both sides of each comparison are namespace arrays. Writing a device1 CCDData via CCDData.write() — which ImageFileCollection round-trips and combine(output_file=...) both do — needs numpy_ccddata(); comparisons against results that land on the namespace's default device use numpy_copy() + assert_allclose() instead of xp.all(...).

Two of this bucket's targets still fail after these fixes, on causes outside test-code's reach: test_combiner_result_dtype's dtype=int case, because combine() passes the bare Python int straight to xp.astype/xp.asarray, which array-api-strict rejects (AttributeError: type object 'int' has no attribute '_np_dtype') — a new gap, not previously reachable because an earlier .multiply() failure masked it; and test_combine_overwrite_output, because combine(output_file=...) writes the still-in-namespace result via ccd.write() at combiner.py:1273 (AttributeError: 'Array' object has no attribute 'astype'), the same class of bug as #935. Both are left failing, unmarked, for the coordinator to track in #971. Three more tests that call default median_combine() (test_combiner_with_scaling, test_combiner_with_scaling_uncertainty[median_combine], test_writeable_after_combine[median_combine]) remain failing on #929's sigma_func/astropy.stats densification, unrelated to and untouched by this PR — test_writeable_after_combine[median_combine] was already known to the plan; the other two are additional #929 instances this PR's fixes exposed by clearing their first-listed cause.

Because these fixes remove every CCDData.multiply() call from the affected tests, the jax-only backend_xfail markers on test_combiner_with_scaling, test_combiner_result_dtype, test_combine_overwrite_output, and test_combiner_with_scaling_uncertainty (added for a jax DeprecationWarning triggered by astropy arithmetic receiving a jax array as dtype=) no longer apply and are removed; jax now XPASSes all of them cleanly.

Fix test_subtract_overscan_model for strict; xfail the astropy.modeling path. The test built its overscan gradient with integer/float division array-api-strict rejects, then unpacked the resulting 3D array by iteration, which array-api-strict also disallows for arrays with more than one dimension; fixed with an explicit float64 dtype/device and [0, ...]/[1, ...] indexing. Also replaced two ndarray.copy() calls with xp.asarray(..., copy=True). With those fixed, both parametrizations reach subtract_overscan(model=...), which goes through astropy.modeling's LinearLSQFitter — NumPy-only and unrelated to this test's array handling (#933); marked backend_xfail("array-api-strict", ...) citing #933. Exact stop: TypeError: object of type 'Array' has no len() from xp.arange(len(oscan)) in ccdproc/core.py's subtract_overscan.

Reset leaked _config_ccd_requires_unit before test_generator_ccds_without_unit. astropy's _arithmetic decorator sets the module global _config_ccd_requires_unit = False around an arithmetic call and restores it only on success, with no try/finally; when an earlier test's arithmetic raises (test_unit_mismatch_behaves_as_expected does this deliberately, on every backend), nothing resets the flag before later tests run (astropy/astropy#20268). This previously only reproduced on jax (and, after this PR's strict fixes, on array-api-strict too) and was marked backend_xfail for both — but removing several successful .multiply() calls in the test_combiner.py commit above also removed incidental resets of that global, which surfaced the same leak as a genuine regression on plain NumPy. Rather than add NumPy to the xfail list, the test now resets astropy.nddata.ccddata._config_ccd_requires_unit = True itself and drops the marker entirely — it passes deterministically on all four backends regardless of what ran before it.

Leftover failures after this PR (for #971 bookkeeping):

• test_combiner_result_dtype — combine(dtype=int) vs. array-api-strict's dtype objects (new gap, undocumented pre-PR).
• test_combine_overwrite_output — combine(output_file=...) writes a namespace array via CCDData.write() (same class as Array API: decide a policy for inherently CPU-only operations #935).
• test_combiner_with_scaling, test_combiner_with_scaling_uncertainty[median_combine], test_writeable_after_combine[median_combine] — Array API: Combiner.sigma_clipping and the default sigma_func densify via astropy.stats #929 (median_combine's default sigma_func densifies via astropy.stats).
• The remaining ~10 failures are the pre-existing, untouched upstream bucket (test_unit_mismatch_behaves_as_expected, test_sigma_func_for_ccddata, test_combiner_dtype, test_combiner_sigmaclip_*, test_combiner_median, test_combine_result_uncertainty_and_mask[median_combine-*], test_rebin_ccddata[True-True]).

🤖 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 (2d34b6e).
⚠️ Report is 16 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #998      +/-   ##
==========================================
+ 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% <ø> (+0.01%)	⬆️
jax	97.13% <ø> (+0.18%)	⬆️
numpy	97.65% <ø> (+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.