Write combine(output_file=) from a NumPy copy; map dtype to the array namespace

[mwcraig]

Closes out the last three ccdproc-side items in section 1 of #971.

combine(output_file=...) handed the namespace result straight to CCDData.write (combiner.py:1273), which astropy.io.fits cannot take (AttributeError: 'Array' object has no attribute 'astype' on array-api-strict). The writer now gets a CCDData built from NumPy copies of data, mask and uncertainty.array (same uncertainty class, unit/meta/wcs carried over) and the returned result stays in its namespace; test_combine_overwrite_output now asserts both.

The copy is a new ccdproc.core._to_numpy(arr): the deliberate host-side conversion for NumPy-only consumers. It moves the array to the namespace's default device first (asked for through the standard __array_namespace_info__().default_device(), skipped when that is None as on JAX), because array-api-strict refuses to export from its non-default devices. This is the same body as the numpy_copy test helper from #998, which now just delegates to it so the device logic lives in one place. Two consequences worth knowing:

• it is the first library-side host-copy helper, so it is the natural building block for whatever Array API: decide a policy for inherently CPU-only operations #935 decides for the CPU-only operations;
• in the escape log, test-suite calls through numpy_copy are now attributed to core.py _to_numpy (the innermost non-test ccdproc frame), so that site's count is inflated by the test helpers. It is one baseline entry either way.

Combiner(dtype=) / combine(dtype=) passed the user's dtype straight to xp.asarray/xp.astype (combiner.py:191, :1104); a builtin int or a string such as "float32" is a valid NumPy dtype but array-api-strict rejects it. core._namespace_dtype(dtype, xp) resolves the name with numpy.dtype(...) and looks it up on the namespace; anything NumPy cannot interpret (the namespace's own dtype objects, e.g. array_api_strict.float32) passes through untouched. Verified for builtins, strings, NumPy scalar types/dtypes and namespace dtypes on numpy, jax, dask and strict. New test_combiner_dtype_mapped_to_namespace covers int, float, "float32", np.float32; the dtype docstrings are updated. test_combiner_result_dtype then failed in its own body — comparing the integer result with a float reference, which the standard does not promote — and now compares with a Python int.

subtract_overscan used len(oscan) for the model fit (core.py:698); now oscan.shape[0]. The test stays xfailed on #933 (astropy.modeling), as expected.

Escape baseline (ccdproc/tests/array_escape_baseline.txt): regenerated on dask with CCDPROC_WRITE_ESCAPE_BASELINE=1. _to_numpy is the one new site, tagged BOUNDARY. The rewrite also dropped the average_combine/median_combine/sum_combine entries, which have not fired since #992 replaced the mask= constructor calls with _mask, and updated the header comment (the writer's current text). The dask enforce run passes.

Results

• strict: 13 failed / 502 passed / 36 skipped / 45 xfailed / 0 xpassed (was 15). test_combine_overwrite_output and test_combiner_result_dtype clear. All 13 that remain are upstream-blocked: 11 × Array API: Combiner.sigma_clipping and the default sigma_func densify via astropy.stats #929, Array API: units/Quantity handling with non-numpy arrays #936, Consider marray as a uniform masked-array representation across array backends #983 — section 1 of Tracking: remaining work for the array-API migration #971 is done.
• numpy: 561 passed / 35 skipped. jax and dask: test_combiner.py + test_ccdproc.py 177 passed each. dask + escape ratchet: 554 passed, no new escapes.

🤖 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 97.78%. Comparing base (2d34b6e) to head (57fd985).
⚠️ Report is 9 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #999      +/-   ##
==========================================
+ Coverage   97.76%   97.78%   +0.02%     
==========================================
  Files           9        9              
  Lines        1788     1808      +20     
==========================================
+ Hits         1748     1768      +20     
  Misses         40       40              

Flag	Coverage Δ
dask	96.94% <100.00%> (+0.03%)	⬆️
jax	97.11% <95.65%> (-0.03%)	⬇️
numpy	97.67% <100.00%> (+0.02%)	⬆️

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.

[mwcraig]

Shorten this

[mwcraig]

Yes. Proposed, 4 lines instead of 7 and no longer spelling out every accepted form (the _namespace_dtype docstring does that):

    dtype : dtype-like or None, optional
        The dtype of the stacked data and the results; NumPy-style dtypes
        (e.g. ``int``, ``"float32"``) are mapped to the namespace's dtype
        of the same name. Default is ``None``, i.e. ``float64``.

The combine() one already just points here, so it can stay as is. Can push when you say so.

— Written by Claude at @mwcraig's direction.

[mwcraig]

Done in 57fd985, as proposed above.

— Written by Claude at @mwcraig's direction.

[mwcraig]

Wasn't something like this added earlier?

[mwcraig]

Yes — numpy_copy in ccdproc/tests/pytest_fixtures.py, added in #998 (54f321d, made device-generic in 2d34b6e). Its body is exactly this function.

This PR does not duplicate it, it moves it: the body now lives here as core._to_numpy because the library needs it for the output_file writer, and numpy_copy is reduced to a one-line delegate (return _to_numpy(array)) so the device logic exists in one place. That is the pytest_fixtures.py hunk in the diff. There is no other library-side host copy to reuse — _native_numpy (#995, core.py:102) is the different job of converting a NumPy array to native byte order, and a grep of the package for to_device/default_device/np.asarray finds nothing else outside the test suite and the escape logger.

If you would rather not keep the test-suite alias at all, the two test callers (numpy_copy itself and numpy_ccddata) can import _to_numpy from ccdproc.core directly, the way test_combiner.py already imports _native_numpy; that removes numpy_copy and its docstring entirely. Happy to do either.

— Written by Claude at @mwcraig's direction.

[mwcraig]

Done in 57fd985 — numpy_copy is gone; numpy_ccddata and test_combiner.py import _to_numpy from ccdproc.core directly. Full suite unchanged on numpy (562 passed) and strict (the same 13, all upstream); jax and the dask escape ratchet pass.

— Written by Claude at @mwcraig's direction.