Compute Combiner.clip_extrema mask with rank comparison instead of scatter

[mwcraig]

Part of #971 (the clip_extrema bucket).

array_api_strict strict test run on top of main (711bb26): 41 failed → 37 failed, no new failures on any of the four backends (numpy, jax, dask, strict). The failure list after this change is identical to main's minus the four clip_extrema tests.

Combiner.clip_extrema masked the nlow lowest and nhigh highest per-pixel values by building flat integer indices by hand and scattering into the mask with xpx.at(...)[flat_index].set(True) -- the only integer-array xpx.at(...).set call anywhere in the codebase. array_api_strict rejects fancy-indexed __setitem__, and there is no put_along_axis in the array API standard, so this raised IndexError on strict.

A second argsort along axis 0 inverts the first: ranks[k, ...] is the rank of image k value at each pixel, so comparing ranks against nlow/nhigh produces the identical mask without any scatter. Both argsort calls default to stable=True, so tie-breaking is unchanged from the old code -- verified against a hand-written reference implementation across 8400 random trials (with ties and NaNs, several shapes) run against the new code path, 0 mismatches. This also let the numpy.mgrid-based coordinate construction and the now-unused from numpy import mgrid as np_mgrid import be deleted.

test_clip_extrema_keeps_indices_in_array_namespace, which existed to spy on the removed scatter indices via a monkeypatched xpx.at, no longer has anything to spy on. It is replaced with two tests: one checking the resulting mask namespace/device (reusing the same tied (3, 2, 3) fixture data and expected mask), and one checking the mask content directly against a documented 3-way tie.

The four target strict failures (test_clip_extrema_3d, _alone, _via_combine, _with_other_rejection) all pass; numpy, jax (JAX_ENABLE_X64=True), and dask are unaffected (same pass/skip/xfail counts as main, no failures). Ruff clean.

🤖 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 97.75%. Comparing base (f85764e) to head (f61e08e).
⚠️ Report is 5 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #994      +/-   ##
==========================================
- Coverage   97.75%   97.75%   -0.01%     
==========================================
  Files           9        9              
  Lines        1785     1780       -5     
==========================================
- Hits         1745     1740       -5     
  Misses         40       40              

Flag	Coverage Δ
dask	96.89% <100.00%> (-0.01%)	⬇️
jax	96.95% <100.00%> (-0.01%)	⬇️
numpy	97.64% <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.

[mwcraig]

Why do we still need this guard? I thought the point of this PR was to fix clip_extrema

[mwcraig]

We don't — it's stale. The guard was carried over from the #928 test (a4ffcef, 2026-07-19), when Combiner.__init__ still built the stack with a nested xp.asarray([...]) that array-api-strict rejects. That was fixed by xp.stack in ebfe8de (2026-08-22, the #965 item), so the branch has been dead since; this PR only preserved it because the plan said to reuse the old construction verbatim.

Checked on this branch with the same three images, on both strict's default device and device1:

Combiner([CCDData(xp.asarray(im, device=...), unit=u.adu) for im in data]).clip_extrema(nlow=1, nhigh=1)

__init__ and clip_extrema both succeed and the mask stays in array_api_strict on the input device. test_clip_extrema_3d and friends already construct their Combiner the normal way on strict and pass.

So _clip_extrema_combiner can go entirely; each test just does

c = Combiner([CCDData(xp.asarray(image, device=xp_device), unit=u.adu) for image in data])

(device=xp_device so the device assertion in the first test still exercises the non-default device on strict.) Happy to push that.

— Written by Claude at @mwcraig's direction.