commit e4d0c60a812a02d6a840c33673758ecb5afd6f99
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun Aug 16 11:02:00 2026 +0200

dev-python/pycuda: guard the compiler-bindir sed substitution

The src_prepare sed that injects --compiler-bindir into PYCUDA_DEFAULT_NVCC_FLAGS has
no match guard, and sed exits 0 on no-match, so its || die never fires. If the
upstream compiler.py text drifts, the substitution silently no-ops, the nvcc host
compiler falls back to the host default (rejected by CUDA 13), and the failure only
surfaces at runtime JIT. Add a grep -q guard on the substitution target so a drift
dies at src_prepare instead.

commit 1aa419d9b8da8c8945eb2a8d67c575199c9588ad
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Aug 14 17:16:47 2026 +0200

dev-python/pycuda: complete CUDA build requirements

The exact 2026.1 metadata requires wheel and NumPy 1.24 during the build and declares
runtime floors for platformdirs and pytools. Runtime kernel compilation must also use
the CUDA-supported host compiler, and GPU initialization requires disabling sandbox
interception for the privileged test phase.

commit 7e7eff6aded09584cd27a8dce87cf5534969a527
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Apr 29 01:43:10 2026 +0200

dev-python/pycuda: add 2025.1.3

Adds pycuda 2025.1.3 (released 2026-02-18) alongside 2026.1, in
line with the overlay's last-of-each-major-series retention
policy.

The 2025.1.x maintenance branch tracks the 2025-API line in
parallel with 2026.x; install_requires are identical to 2026.1
(mako, numpy, platformdirs, pytools), so the ebuild is a straight
copy of pycuda-2026.1 with only PV/SRC_URI differing.

Tested with ebuild ... clean install against the live cuda
toolkit; no QA notices.

commit 54fbf16529cb77cd87024c992e0ad3582547b34d
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Apr 29 01:42:52 2026 +0200

dev-python/pycuda: set DISTUTILS_EXT=1

The package builds and installs compiled extension modules
(_pvt_struct.so, _driver.so), so distutils-r1.eclass requires
DISTUTILS_EXT=1. Without it, portage emits a QA notice:

QA Notice: Python extension modules (*.so) found installed.
Please set:
DISTUTILS_EXT=1
in the ebuild.

::gentoo's pycuda-2024.1 ebuild has the same omission; this
overlay's ebuild is now ahead.

commit 3510272e319af71171d860684b57e4b96200bd4e
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Apr 23 14:41:20 2026 +0200

dev-python/pycuda: new package, add 2026.1

::gentoo only ships pycuda-2024.1 with PYTHON_COMPAT=python3_.
That leaves dev-python/sasmodels (which cuda?-gates pycuda but claims
python3_) tripping NonsolvableDepsInStable on ~amd64 for the
py3.13/3.14 targets, even though the conditional is never hit for
the cuda-off default. sasmodels can't narrow its own PYTHON_COMPAT
without losing current-default-python coverage, and pkgcheck tests
all USE configurations regardless.

Ship the current upstream pycuda-2026.1 (released 2026-01-15) as an
overlay-local bump. Upstream declares python_requires="~=3.8" and
this keeps landing on py3.13/3.14-capable numpy/boost-python builds
without issue. Changes from ::gentoo's 2024.1 ebuild:

- PYTHON_COMPAT: 3.10..3.12 -> 3.12..3.14 to match current default
python targets and cover sasmodels.
- RDEPEND: upstream install_requires dropped appdirs (replaced by
platformdirs) and decorator. Mirror that here.
- EPYTEST_PLUGINS=() added to quiet pkgcheck's
EPyTestPluginsSuggestion.
- metadata.xml carries personal maintainer in the overlay's
canonical form.

sasmodels' NonsolvableDepsInStable clears as a result; pkgcheck
total 59 -> 56.