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

Python 3.15.0 was released on 2026-10-09. Two things stopped it from
installing: the numpy <2.5 cap in setup.cfg (the first numpy with a
cp315 wheel is 2.5.2; numba 0.67 and 0.68 declare numpy<2.6, so the
cap moves to match) and the libjxl=0.11.* pin in the two micromamba
workflows, which conda-forge's rebuilt GDAL stack (libjxl >=0.12) can
no longer satisfy. The pin's own comment asked for it to be removed
once that rebuild landed.

test.yml: PRs run 3.15 plus the 3.14t probe; main and nightly run
3.12 through 3.15 plus 3.14t. No 3.15t job until numba ships a cp315t
wheel. test-geotiff-corpus.yml moves its newest slot to 3.15.
test-cog-validator.yml stays on 3.12 with the pin removed.

Author: brendancol

.github/workflows/test-cog-validator.yml

@@ -49,12 +49,10 @@ jobs:
       fail-fast: false
       matrix:
         # Pin to a python version where conda-forge ships a stable
-        # rio-cogeo + GDAL combo. 3.12 is the safest target today;
-        # 3.14 wheels for some of the GDAL stack are not yet
-        # universally available, and the gate has to install reliably
-        # to be a useful required check. Add 3.13 / 3.14 here once
-        # conda-forge has caught up with stable rio-cogeo + GDAL
-        # builds for those interpreters.
+        # rio-cogeo + GDAL combo. 3.12 is the most conservative install
+        # for a required check; conda-forge now has GDAL / rasterio
+        # builds for 3.13 through 3.15 as well, so widening this is a
+        # matter of choosing to, not of availability.
         python: ['3.12']
     defaults:
       run:
@@ -70,12 +68,6 @@ jobs:
     steps:
     - uses: actions/checkout@v4
     - name: Set up micromamba env (conda-forge rio-cogeo + GDAL)
-      # libjxl is pinned to 0.11.* because conda-forge published libjxl
-      # 0.12.0 (2026-07-03) before rebuilding libgdal-core / rasterio
-      # against it, so an unpinned solve pairs a GDAL linked to
-      # libjxl.so.0.11 with only 0.12 installed and rasterio fails to
-      # import. Drop the pin once conda-forge ships libgdal-core /
-      # rasterio builds linked against libjxl 0.12.
       uses: mamba-org/setup-micromamba@v1
       with:
         environment-name: xrspatial-cog-validator
@@ -86,7 +78,6 @@ jobs:
           rio-cogeo
           pyyaml
           tifffile
-          libjxl=0.11.*
         condarc: |
           channels:
             - conda-forge

.github/workflows/test-geotiff-corpus.yml

@@ -32,7 +32,7 @@ jobs:
       fail-fast: false
       matrix:
         os: ${{ github.event_name == 'pull_request' && fromJson('["ubuntu-latest"]') || fromJson('["ubuntu-latest", "macos-latest", "windows-latest"]') }}
-        python: ${{ github.event_name == 'pull_request' && fromJson('["3.14"]') || fromJson('["3.12", "3.14"]') }}
+        python: ${{ github.event_name == 'pull_request' && fromJson('["3.15"]') || fromJson('["3.12", "3.15"]') }}
     defaults:
       run:
         # `-el` so micromamba's activation hooks fire and `rasterio`
@@ -41,12 +41,6 @@ jobs:
     steps:
     - uses: actions/checkout@v4
     - name: Set up micromamba env (conda-forge rasterio + GDAL)
-      # libjxl is pinned to 0.11.* because conda-forge published libjxl
-      # 0.12.0 (2026-07-03) before rebuilding libgdal-core / rasterio
-      # against it, so an unpinned solve pairs a GDAL linked to
-      # libjxl.so.0.11 with only 0.12 installed and rasterio fails to
-      # import. Drop the pin once conda-forge ships libgdal-core /
-      # rasterio builds linked against libjxl 0.12.
       uses: mamba-org/setup-micromamba@v1
       with:
         environment-name: xrspatial-geotiff
@@ -56,7 +50,6 @@ jobs:
           gdal
           pyyaml
           tifffile
-          libjxl=0.11.*
         condarc: |
           channels:
             - conda-forge

.github/workflows/test.yml

@@ -34,7 +34,12 @@ jobs:
     strategy:
       matrix:
         os: ['ubuntu-latest', 'macos-latest', 'windows-latest']
-        python: ${{ github.event_name == 'pull_request' && fromJson('["3.14", "3.14t"]') || fromJson('["3.12", "3.13", "3.14", "3.14t"]') }}
+        # PRs run the newest GIL interpreter plus the free-threaded probe;
+        # push-to-main and the nightly run every supported version. There
+        # is no 3.15t entry yet: numba 0.68.0 ships cp314t wheels but no
+        # cp315t, so that job would build numba from source. Add it once a
+        # numba release carries the wheel.
+        python: ${{ github.event_name == 'pull_request' && fromJson('["3.15", "3.14t"]') || fromJson('["3.12", "3.13", "3.14", "3.15", "3.14t"]') }}
     env:
       OS: ${{ matrix.os }}
       PYTHON: ${{ matrix.python }}

setup.cfg

@@ -25,12 +25,14 @@ include_package_data = True
 #   numba, scipy      oldest release shipping a CPython 3.12 wheel, which
 #                     is the floor python_requires allows anyway
 #   numpy >=1.26      first cp312 wheel, and xarray's own numpy floor
-#   numpy <2.5        mirrors numba: 0.64-0.66 declare numpy<2.5 and numba
+#   numpy <2.6        mirrors numba: 0.67-0.68 declare numpy<2.6 and numba
 #                     re-checks it at import in _ensure_critical_deps.
 #                     Without this, pip installs a numpy that makes
 #                     `import xrspatial` raise ImportError. Re-check this
 #                     on every numba bump -- nothing flags a stale cap,
-#                     and leaving it behind blocks a numpy that works.
+#                     and leaving it behind blocks a numpy that works
+#                     (the old <2.5 cap blocked every numpy with a
+#                     CPython 3.15 wheel; the first is 2.5.2).
 #   xarray >=2025.11  earlier releases drop attrs across DataArray binary
 #                     ops, so morph_gradient / morph_white_tophat /
 #                     morph_black_tophat and the mcda and corridor
@@ -44,7 +46,7 @@ install_requires =
     numba >=0.59
     scipy >=1.11.2
     xarray >=2025.11
-    numpy >=1.26,<2.5
+    numpy >=1.26,<2.6
     pandas >=2.2
     urllib3 >=2.7
 packages = find: