array-api-strict triage: escape-site mapping of the remaining test failures

[mwcraig]

Context

The tooling from #937 and #939 is implemented on branch array-api-strict-triage-tooling (mwcraig fork, commit 94ca9ec):

• CCDPROC_ARRAY_LIBRARY=array-api-strict test backend with fixture arrays placed on the non-default Device("device1"), so any implicit conversion to numpy raises — the same failure signal as CuPy's device-to-host error, with no GPU needed. Includes a py313-strict tox env and CI job.
• Escape-site dedupe report (CCDPROC_TRIAGE_ESCAPES=1): groups test failures by the innermost non-test ccdproc frame.
• backend_xfail / backend_skip markers applied per backend; documented limitations (astroscrappy, scipy.ndimage.median_filter, reproject, block_reduce/block_replicate) are now xfailed under array-api-strict.
• Escape logger (CCDPROC_LOG_ARRAY_ESCAPES=1): warns when np.asarray/np.asanyarray/np.ma.asanyarray receive a non-numpy array-API array, catching backends that convert silently (dask, JAX).

Results

Full suite under CCDPROC_ARRAY_LIBRARY=array-api-strict: 139 failed / 192 passed / 5 skipped. After marking documented limitations: 109 failed / 30 xfailed. As predicted in #939, the dedupe report collapses the failures to about ten distinct escape sites.

Escape-site → root-cause mapping

Escape site	Failures	Root cause	Tracked in
combiner.py:189 Combiner.__init__	44	Stacks per-image arrays by building a nested Array; array-api-strict rejects this ("Use stack instead"). Likely also the biggest CuPy failure source.	no existing issue
_ccddata_wrapper_for_array_api.py:74 _arithmetic_wrapper	11	Device-mismatch / dtype coercion inside the CCDData arithmetic wrapper	possibly related to #927
combiner scaling tests	~9	Quantity leaking into array-API arithmetic (TypeError: Expected Array or Python scalar)	#936
test_ccdproc.py:174 test_subtract_overscan	6	numpy-only .copy() method called on an Array	no existing issue
core.py:862 subtract_dark	4	unit-mismatch/device issue	no existing issue
core.py:611 subtract_overscan (via ccd_process)	4	median fallback doesn't handle namespaces without median/allclose	#906
core.py:1459 rebin	4	numpy-only coordinates.astype("i") method instead of xp.astype(...)	no existing issue
core.py:459 create_deviation	~2	bool-dtype handling in __mul__	no existing issue
core.py:1233 sigma_func + gain/deviation singletons	~4	densification / device mismatch via astropy.stats	#929
test_image_collection.py, test_ccdproc_logging.py:75, misc singletons	~4	test-suite code using raw numpy (np.zeros_like) on foreign-device arrays; an ImageFileCollection gap	no existing issue (partly test-only)

Xfailed as documented limitations (30 tests): astroscrappy (cosmicray_lacosmic, 21), scipy.ndimage.median_filter (cosmicray_median, 7 — overlaps #935), reproject (wcs_project, 4 — see #930/#940), block_reduce/block_replicate (3 — see #940).

Suggested next steps

• Decide whether the rows marked "no existing issue" get their own issues (the Combiner.__init__ nested-array stacking alone accounts for 44 of 109 failures and looks like a one-line xp.stack fix).
• Once sparse (Add sparse test backend as a densification detector (and eventual supported library) #938) is wired up the same way, failures reproducing under both proxies can be treated as confirmed CuPy bugs per the CuPy issues #909 strategy.

---

Follow-up to #909, #937, #939. Counts from a 2026-07-06 run on the branch above.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Qri8rDgkWjCT7Yb7QKJrhf