The 2.3.2-1 wheels here were built using the following procedure:

• Use a Linux or Mac machine.
• Check out this PR.
• cd to the top-level directory of this repository.
• git submodule update --init
• Create and activate a Python 3.13 virtual environment.
• Run pip install git+https://github.com/mhsmith/cibuildwheel@android-numpy
• Run cibuildwheel --only cp313-android_ARCH, where ARCH is either arm64_v8a or x86_64.
• cibuildwheel will attempt to run the test suite on an Android emulator. If you don't want that, set the environment variable CIBW_TEST_SKIP="*".

These wheels pass the entire test suite except for:

• A number of tests involving floating point exceptions which are expected to occur but don't. The discussion in #29724 indicates this may be caused by slight differences between compilers.
• 3 tests called test_legacy_pickle, due to bugs in how the testbed handles .gz files. I have a fix for this in python/cpython#142912, which I'll eventually include in the cibuildwheel PR.
• On Chaquopy 17, this workaround is required.

Thanks: I intend to create a set of similar PRs for numpy, pandas, scipy, and scikit-learn, so I can produce wheels of known versions for some Chaquopy users I'm working with. Once these PRs are fully working, I'll retarget them to the main branch of each project to enable them to release Android wheels on PyPI.

I'd still be happy to merge all the test changes separately.

In case you haven't seen it yet, we now have NEP 57 — NumPy platform support, defining CPython-like support tiers. It may be reasonable for this PR to aim to promote Android to Tier 3 by adding a CI job, similar to the current Emscripten one.