Keep clip_extrema indices in the array namespace

[nomad3]

Summary

• replace NumPy ravel_multi_index with equivalent C-order arithmetic in the selected array namespace
• keep coordinate indices on the input array device and use Array API-compatible reshaping/indexing
• add an asymmetric regression that checks the exact high/low flat indices and non-default-device preservation

Fixes #928

Validation

• python -m pytest ccdproc -q — 380 passed, 29 skipped
• python -m pytest ccdproc/tests/test_combiner.py -q — 87 passed with NumPy and 87 passed with Dask
• JAX Combiner module with 64-bit enabled — 81 passed, 6 established xfails
• focused regression with array-api-strict on a non-default device — 1 passed
• Ruff, Black, git diff --check, and git show --check — passed

array-api-strict deliberately excludes integer-array assignment, so the focused strict regression replaces only the final xpx.at update with a test spy. The production update path is exercised by the NumPy, Dask, and JAX runs; strict verifies that the exact C-order indices are computed in the selected namespace and remain on the data device.

Checklist

• For new contributors: Did you add yourself to the AUTHORS.rst file? (The entry is already present on current main, so it is intentionally not duplicated here.)
• For documentation changes: Does your commit message include a [skip ci]? (Not a documentation-only change.)
• For bugfixes: Did you add an entry to the CHANGES.rst file?
• For bugfixes: Did you add a regression test?
• For bugfixes: Does the commit message include Fixes #928?
• Does this PR add, rename, move or remove any existing functions or parameters?

AI assistance disclosure

OpenAI Codex was used for source inspection, test design, implementation, and automated review. I have personally reviewed the current diff and validation evidence and take responsibility for this contribution.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 95.83%. Comparing base (21041b0) to head (89e90f9).
⚠️ Report is 4 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #954   +/-   ##
=======================================
  Coverage   95.82%   95.83%           
=======================================
  Files           8        8           
  Lines        1629     1631    +2     
=======================================
+ Hits         1561     1563    +2     
  Misses         68       68           

☔ 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]

This PR remains a draft pending the contributor's personal review; the contributor will take responsibility for the final contribution before marking it ready.

Can you please go ahead and make this non-draft when you have a chance to look it over?

[mwcraig]

@nomad3 -- is this one ready for review?

[nomad3]

Yes—thank you for checking. I have personally reviewed the rebased diff and fresh validation, corrected the changelog placement, and marked it ready for review. All completed exact-head CI jobs are green; the remaining Dask+Bottleneck matrix job is still queued.

[mwcraig]

This review was drafted by Claude Code and is being posted on behalf of @mwcraig, who reviewed and approved its contents.

Thanks @nomad3 — this is a careful fix and an unusually well-constructed test.

I verified the replacement arithmetic by hand: the accumulation loop over shape[1:] is exactly C-order ravel_multi_index, and the strict=True zip guards against a coordinate/shape mismatch. Two quiet improvements ride along: argsorted[(i, ...)] removes the old hardcoded assumption of a 3D stack (which also violated array-api-strict indexing rules for 3D-image stacks), and creating mg on the data device removes the implicit host-to-device transfer that #928 was about. I also recomputed the test's hardcoded flat indices and expected mask for the 3×2×3 data and they're correct.

One optional suggestion — happy to merge as-is if you'd rather not:

• The test builds the Combiner via object.__new__ and sets _xp / _data_arr / _data_arr_mask directly, which couples it to private internals for every backend even though only array-api-strict needs the bypass (it can't stack a list of arrays). If you're willing, branching so that non-strict backends construct the Combiner normally (from a list of CCDData) and only strict uses the bypass would make the test more robust against future changes to clip_extrema's internals. Entirely non-blocking.

[mwcraig]

Thanks again @nomad3!