Test on Python 3.15 and lift the numpy cap to <2.6

[brendancol]

Python 3.15.0 came out on 2026-10-09. This puts it in CI and fixes the two things that stopped it from installing at all.

Why it did not install

setup.cfg capped numpy at <2.5. The first numpy with a CPython 3.15 wheel is 2.5.2, so on 3.15 pip went looking for a numpy 2.4.x sdist to compile. numba 0.67 and 0.68 both declare numpy<2.6, so the cap moves to <2.6. The comment block says to re-check the cap on every numba bump; this is that re-check. Rebased on #3754, so the line now reads numpy >=2.0,<2.6.

The two micromamba workflows pinned libjxl=0.11.* (#3639) because conda-forge shipped libjxl 0.12 before rebuilding GDAL against it. That rebuild is done: every current libgdal-core 3.13.3 build requires libjxl >=0.12, and rasterio 1.5.2 / gdal 3.13.3 have py315 builds on linux, macOS and windows. With the pin still there a 3.15 env cannot solve, so it comes out of both files, which is what its own comment asked for.

CI

test.yml: PRs now run 3.15 plus the 3.14t free-threaded probe. Push-to-main and the nightly run 3.12, 3.13, 3.14, 3.15 and 3.14t. There is no 3.15t job. numba 0.68.0 ships cp314t wheels but not cp315t, so a 3.15t job would compile numba from source on every run; a comment in the matrix says so, and it can go in once a numba release carries the wheel.

test-geotiff-corpus.yml: the newest slot moves from 3.14 to 3.15, so PRs run 3.15 and full runs do 3.12 and 3.15. Pin removed.

test-cog-validator.yml: stays on 3.12 since it is a required check and that is the most conservative install. Pin removed, and the comment claiming 3.14 GDAL builds were unavailable is updated.

One thing to be aware of: with the cap raised, every matrix entry resolves numpy 2.5.x now, not just the 3.15 one.

Checks

Nothing in xrspatial/ touches an API removed in 3.15 (the sre_* modules, glob.glob0, load_module, co_lnotab, ByteString, and so on). The only open() calls without encoding= read /proc/meminfo, so the new UTF-8 default is a no-op for us.

Every runtime and test dependency has a cp315 wheel or is pure Python, with two exceptions. zstandard 0.25.0 has no cp315 wheel yet (indygreg/python-zstandard#354) and noise never had wheels for anything. Both compile from sdist, and noise already does that in every existing job. cupy has no cp315 wheels, so GPU is out of scope and gpu-extras-resolve stays on 3.12 until that changes.

Locally, on a conda-forge Python 3.15.0 env with numba 0.68.0 and numpy 2.5.4: pip install -e .[tests] resolves, zstandard and noise build from sdist, the ZSTD codec imports, the setup.cfg metadata guards pass, and the fast lane (pytest -m "not slow") passes: 10992 passed, 3024 skipped (no cupy on 3.15), 1 xfailed, 1 failed. The one failure is a Hypothesis fuzz case that fails identically on 3.14 / numpy 2.3 and is filed as #3752; it is not a 3.15 problem..

[brendancol]

CI evidence, since the PR trigger only runs the newest interpreter.

Dispatched the full matrix from this branch:

• test.yml run 38194687347: all 15 run jobs green, that is 3.12, 3.13, 3.14, 3.15 and 3.14t on ubuntu, macOS and windows, full lane with the slow cells. Every one of them resolved numpy 2.5.4, so the cap change is exercised on every version, not just 3.15. zstandard 0.25.0 and noise built from sdist on all three OSes.
• test-geotiff-corpus.yml run 38194689132: 3.12 and 3.15 on all three OSes green. The 3.15 env solved python 3.15.0, rasterio 1.5.2, gdal 3.13.3 with the libjxl pin gone.
• test-cog-validator.yml on the PR: green on 3.12 without the pin.

min-deps is red on both the PR and the dispatched run. All 30 failures are test_zonal.py dask cases with 'NoneType' object has no attribute 'concat', which is #3751 (pyarrow 26.0.0 needs numpy 2, the job pins numpy 1.26.0). It has been red on main since the 2026-10-10 nightly and is not touched by this change.