GeezerVison

conda-forge ships two ABIs for CPython 3.13+, and the "Update env" step did not
restate ``python``, so the solver was free to replace the interpreter
setup-miniconda installed. Three of the last sixteen python 3.14 draws on
ubuntu-24.04-arm did exactly that: the log reads ``python
3.14.7-h6bfacdd_100_cp314 --> 3.14.7-hefd156b_0_cp314t`` and pytest then reports
"Python: 3.14.7 free-threading build".

Two of the three passed only because they also drew pandas 3.0. The third drew
pandas 2.3, whose extension raises "The global interpreter lock (GIL) has been
enabled to load module 'pandas._libs.pandas_parser'", and ``filterwarnings =
error`` turns that into a collection error. It took test_formatting.py and
test_recorder.py down with it, the second only because the first had left pandas
half-imported.

The window opened on 2026-08-19, when conda-forge uploaded python 3.14.7 for
linux-aarch64 in both ABIs. Re-solving the failing job's own spec set per
platform: linux-aarch64 and osx-arm64 both pick ``cp314t`` today, so
macos-latest is on the same fuse and has simply not drawn 3.14 with an unlucky
pandas yet. linux-64, osx-64 and win-64 pick the GIL build either way, which is
why this has been an arm-only symptom.

``python-gil`` pins ``python_abi``; restating ``python=`` pins the version,
since ``python-gil`` alone constrains only the ABI and its own version is left
open. setup-miniconda's ``python-version`` is equally ABI-free, so this step is
where both get settled. Re-solved with the pair added: every platform lands on a
``cp314`` build, and no draw becomes unsatisfiable or downgrades numpy, numba,
pandas or psg, including python 3.11 with numpy 1.24 and scipy 1.9. python-gil
has builds for 3.9 through 3.14, so one spec covers every version in the pool.
The pypy guard matches the neighbouring lines: ``pypy python-gil`` is
unsatisfiable, so an unguarded spec would hard-fail a pypy draw if one is ever
added back.

python-suitesparse-graphblas pins the same pair in
``.github/workflows/test.yml``, and asks for ``python-freethreading`` on the
matrix entries that want the other ABI. An entry like that would be worth having
here too, but it is a coverage decision rather than a fix, so it is not in this
commit.