Tracking: upstream array-API blockers (astropy, reproject, astroscrappy)

[mwcraig]

Purpose

Single place to track the upstream dependencies that currently force ccdproc data back to numpy, with links to upstream issues as they are filed, and the interim decision for each (vendored replacement vs. CPU-only policy vs. wait).

Blockers

• astropy.stats.sigma_clip — uses np.asanyarray + np.ma internally (used by Combiner.sigma_clipping)
• astropy.stats.median_absolute_deviation — numpy-only (default sigma_func for median_combine)
• astropy.nddata block_reduce / block_replicate — force np.asanyarray on inputs
• CCDData / NDData arithmetic — numpy-based (see TODO at

  ccdproc/ccdproc/core.py

  Line 1016 in 9d25eee

  # TODO: Workaround for the fact that CCDData currently uses numpy

  and the _CCDDataWrapperForArrayAPI workaround)
• astropy.modeling fitters — numpy internals (used by subtract_overscan, create_deviation paths)
• astropy.units.Quantity — numpy-backed (Array API: units/Quantity handling with non-numpy arrays #936)
• reproject — CPU-only (used by wcs_project)
• astroscrappy — C extension, CPU by design (used by cosmicray_lacosmic); realistically permanent → document under the CPU-only policy (Array API: decide a policy for inherently CPU-only operations #935)
• __array__(copy=...) numpy-2 protocol warning — origin upstream, see Array API: locate the source of __array__(copy=...) DeprecationWarnings in CuPy runs #934

For each: file or link the upstream issue, and note the interim plan here.

---

Follow-up to #909. Found during a review of the array API implementation from #885.