Fix numpy-isms in strict-backend test bodies (test hygiene)

[mwcraig]

Summary

Test-only hygiene batch for the array-API migration (#971). Many
array-api-strict test failures were caused by numpy-only idioms in the
test bodies, not by bugs in ccdproc source. This PR fixes only files
under ccdproc/tests/.

Strict suite: 85 failed, 366 passed → 65 failed, 386 passed (36
skipped / 46 xfailed / 5 xpassed unchanged). Numpy backend stays at 0
failed (503 passed); dask backend on the four changed test files also
stays at 0 failed.

No CHANGES.rst entry — this is a test-only change.

Fixed causes (itemized)

1. Test suite: test_subtract_overscan calls numpy-only .copy() on array-API arrays #969 — .copy() is not an array-API method. Replaced
   arr.copy() with xp.asarray(arr, copy=True) in
   test_ccdproc.py::test_subtract_overscan. Closes all 6 parametrized
   failures for that test as numpy-isms; two of the six
   ([True-False-False], [True-False-True], the median=True cases)
   still fail afterward because they hit the separate, already-tracked
   xp.median gap at core.py:629 (another agent is fixing that on a
   different branch) — the .copy() fix itself is still correct and is
   kept per the issue's scope.
2. Test suite: raw np.zeros_like on backend-aware arrays in test_ccdproc_logging.py and test_rebin.py #970 — raw np.zeros_like mixed with backend arrays.
   test_ccdproc_logging.py::test_implicit_logging now builds its bias
   frame with xp.zeros_like (now passes). test_rebin.py::test_rebin_ccddata[True-True]
   was updated the same way, but it still fails afterward: astropy's
   CCDData.mask setter (astropy/nddata/compat.py) forces
   np.asarray(value, dtype=np.bool_), which can't convert an array on
   a non-default strict device. That's an astropy-internals limitation
   outside a test-only fix, so it's kept per the issue's scope and noted
   as a residual failure.
3. test__overscan_schange used xp.allclose, which isn't in the
   array-API standard. Replaced with xp.all(xpx.isclose(...)).
4. test_combiner.py numpy-only method calls (.mean(), .all()) on
   backend arrays, replaced with xp.mean(...) / xp.all(...):
   test_combiner_minmax, test_combiner_minmax_max,
   test_combiner_minmax_min.
5. Broader numpy-isms diagnosed from actual strict tracebacks:
   • Single-axis indexing (arr[i]) on arrays with ndim > 1, which
     the array-API standard requires an explicit ellipsis for
     (arr[i, ...]): test_subtract_overscan_fails,
     test_user_supplied_combine_func_that_relies_on_masks[sum_combine].
   • int/float mixed arithmetic that strict's dtype promotion rejects:
     test_trim_with_wcs_alters_wcs (xp.asarray(shape) / 2 on an int
     array), test_pixelwise_weights, test_combiner_sum_weighted,
     test_combiner_sum_weighted_by_pixel,
     test_combiner_sum_weighted_with_mask,
     test_combiner_uncertainty_average,
     test_combiner_uncertainty_average_mask,
     test_combiner_uncertainty_median_mask,
     test_combiner_uncertainty_sum_mask (int-literal arrays passed to
     xp.std/xp.sqrt, and an unwrapped np.float64 from
     astropy.stats.median_absolute_deviation multiplied against a
     strict Array, following the existing float(...) cast pattern
     already used elsewhere in that file).
   • xp.asarray(list_of_arrays) ("Nested Arrays are not allowed"),
     replaced with xp.stack(...):
     test_user_supplied_combine_func_that_relies_on_masks[sum_combine].

Failures examined and deliberately left (with why)

• test_ccd_process[*], test_ccd_process_gain_corrected,
  test_subtract_overscan[True-False-*] (median path) — the known,
  already-tracked xp.median gap at core.py:629; another agent is
  fixing this on a separate branch, explicitly out of scope here.
• test_flat_correct*, test_gain_correct*, test_gain.py::* — source
  bugs in gain_correct/flat_correct (core.py:934, core.py:1023,
  core.py:1037): xp.asarray(...) called without device=, causing
  "Arrays from two different devices" errors from ccdproc source, not
  the test body.
• test_flat_correct_data_uncertainty — intentionally uses a raw numpy
  array as uncertainty; the test's own comment documents this as a
  regression test for astropy's NDUncertainty explicitly checking for
  numpy.ndarray, to be removed once that's fixed upstream.
• test_flat_correct_deviation, test_sigma_func_for_ccddata,
  test_combiner_median/average/sum/dtype,
  test_average_combine_uncertainty, test_median_combine_uncertainty,
  test_sum_combine_uncertainty, test_combine_result_uncertainty_and_mask[*] —
  fail inside astropy.stats.median_absolute_deviation or
  CCDData.mask/Quantity construction forcing numpy conversion of a
  non-default-device strict array; astropy-internals, not test-body.
• test_combiner_sigmaclip_high/low/single_pix — the test intentionally
  passes astropy.stats.median_absolute_deviation as dev_func, which
  forces a numpy conversion inside combiner.py's sigma_clipping;
  same astropy-internals limitation.
• test_combine_average_fitsimages, test_combine_numpyndarray,
  test_combiner_image_file_collection_input,
  test_combine_image_file_collection_input,
  test_combine_average_ccddata, test_combine_limitedmem_fitsimages,
  test_combine_limitedmem_scale_fitsimages,
  test_combine_ccd_with_uncertainty_and_mask_from_fits[*] — explicitly
  out of scope per the task brief (combine() file-input handling
  returns plain numpy arrays).
• test_clip_extrema_3d/alone/via_combine/with_other_rejection —
  explicitly out of scope per the task brief (clip_extrema fancy
  indexing).
• test_combiner_with_scaling, test_combiner_result_dtype,
  test_combine_overwrite_output, test_combiner_with_scaling_uncertainty[*] —
  astropy ndarithmetic's np.result_type(ref, operand) can't convert
  a strict Array's dtype; astropy-internals ("Could not convert Array"
  pattern named as out of scope in the task brief).
• test_writeable_after_combine[*] — CCDData.write/mask conversion
  forcing numpy on a non-default device; source/astropy-internals.
• test_3d_combiner_with_scaling, test_combiner_with_scaling (helper
  scale_by_mean) — tried fixing the .mean() numpy-ism, but it then
  exposed a genuine ccdproc source bug: Combiner.scaling's setter
  (combiner.py:327) calls xp.asarray(list_of_arrays) on a list of
  strict Arrays ("Nested Arrays are not allowed. Use stack instead."),
  and for test_combiner_with_scaling the test is blocked even earlier
  by the astropy ndarithmetic issue above. Per the task's revert
  policy for non-numbered items, these test-body changes were reverted
  back to the pre-existing (still-numpy) form and left failing.
• test_transform_image[True-True] — the test's user-supplied tran = lambda arr: 10 * arr is applied generically to data, uncertainty, and
  mask by transform_image; ccdproc's array-API wrapper coerces the
  mask to bool before this call, so 10 * bool_array hits strict's
  "Only numeric dtypes are allowed in __rmul__". Making the test's
  transform function dtype-aware would change what the test exercises
  rather than being a simple hygiene fix, so left as-is.
• test_unit_mismatch_behaves_as_expected — fails inside astropy's
  Quantity.__new__ forcing a numpy conversion via
  _arithmetic_data/_prepare_then_do_arithmetic; astropy-internals.
• test_image_collection.py::test_generator_ccds_without_unit — fails
  only when run as part of the full suite (passes in isolation),
  indicating cross-test state leakage unrelated to array-API numpy-isms;
  left as a pre-existing, order-dependent issue outside this batch's
  scope.

Test plan

• CCDPROC_ARRAY_LIBRARY=array-api-strict tox -e strict (from the
  main checkout's tox env): 85 failed, 366 passed → 65 failed, 386
  passed.
• CCDPROC_ARRAY_LIBRARY=numpy full suite: 503 passed, 0 failed (no
  regression).
• CCDPROC_ARRAY_LIBRARY=dask on the four changed test files: 181
  passed, 0 failed.
• ruff check / ruff format --check clean on changed files.

Part of #971

🤖 Generated with Claude Code

https://claude.ai/code/session_01RQMJZUaaxfqGDk41GLSaFK

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 97.42%. Comparing base (3e644d8) to head (288d55a).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #990   +/-   ##
=======================================
  Coverage   97.42%   97.42%           
=======================================
  Files           9        9           
  Lines        1749     1749           
=======================================
  Hits         1704     1704           
  Misses         45       45           

Flag	Coverage Δ
dask	96.55% <ø> (ø)
jax	96.61% <ø> (ø)
numpy	97.31% <ø> (ø)

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.