commit 81ab2102426e32af91802191d0f866f0dd2383f3
Author: lucascouts <lucascs@protonmail.com>
Date: Sun Aug 16 08:46:09 2026 -0300

fix(sci-ml/{sherpa-onnx-1.13.5, koboldcpp-1.118.1}): guard cuda.eclass src_prepare

sci-ml/sherpa-onnx failed the merge gate in the PREPARE phase, for every user
without the CUDA toolkit installed -- including USE=-cuda:

Could not execute cuda-config
ERROR: failed (prepare phase): cuda-config not found
Called cuda_src_prepare -> cuda_sanitize -> cuda_gccdir -> die

cuda.eclass does `EXPORT_FUNCTIONS src_prepare`, and its cuda_src_prepare calls
cuda_sanitize unconditionally. sherpa-onnx defined no src_prepare of its own,
so it inherited the eclass one with nothing gating it. The package was
uninstallable without CUDA -- the whole default configuration.

Introduced by this branch: the adopted ebuild hardcoded
CUDAHOSTCXX=/usr/bin/g++-15, and replacing that with cuda_gccdir (correct on
its own) meant inheriting cuda.eclass, which brought the exported phase along
unnoticed. pkgcheck reports nothing: it does not run phases.

sherpa-onnx now defines src_prepare calling cmake_src_prepare (not `default` --
the cmake eclass tracks state that cmake_src_configure depends on) and
cuda_src_prepare only under `use cuda`.

koboldcpp is fixed for the mirror-image defect. Its own src_prepare overrode
the exported one, which is why it passed the gate, but it therefore never ran
cuda_sanitize at all -- so a USE=cuda build passed unsanitized NVCCFLAGS
straight to nvcc. Now calls cuda_src_prepare under the same guard. Its
cuda_gccdir/cuda_add_sandbox uses were already correctly inside `use cuda`.

Verified without root via `ebuild ... clean prepare`: both packages report
"Source prepared." with USE=-cuda. koboldcpp's two gate legs must be re-run,
since this touches a phase they exercised.

Gate results so far: llama-swap 2/2 legs merged, koboldcpp 2/2 merged,
sherpa-onnx 0/2 -- this is the defect it found.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

commit 494447b3a9e9c2b9a276dfcfc7f8346b92fa186d
Author: lucascouts <lucascs@protonmail.com>
Date: Sun Aug 16 07:14:48 2026 -0300

add(sci-ml/{koboldcpp-1.118.1, sherpa-onnx-1.13.5, llama-swap-250}), mod(.autoupdate/packages.toml)

Story 003 Tasks 1, 3 and 4. NOT MERGED ANYWHERE: authoring and pkgcheck only.
No emerge has been run against any of the three, which is why this sits on a
branch instead of master. Story 002 recorded what skipping that gate costs --
six defects shipped, two of them invisible to pkgcheck until resolution was
attempted. The gate commands are in each task in .epic/stories/003-*/tasks.md.

koboldcpp -- adapted from istitov/stuff sci-misc/koboldcpp.
LLAMA_PORTABLE=1 keeps the Makefile off -march=native; cpu_flags_x86_* wired
to its three tiers via LLAMA_NOAVX1/LLAMA_NOAVX2. cuda and rocm each get
their own emake run: a single invocation defines ggml-cuda.o twice and the
HIP recipes silently override the CUDA ones, so USE="cuda rocm" would have
produced a "cuBLAS" library built by hipcc. GPU_TARGETS is set explicitly
because upstream's default appends $(shell amdgpu-arch), sizing the package
by the build host's GPU. Five prebuilt blobs beyond the documented glslc
pair are removed. pkgcheck: zero findings.

sherpa-onnx -- adapted from istitov/stuff, links sci-libs/onnxruntime rather
than bundling. 17 distfiles pre-staged so CMake FetchContent resolves
locally; the .zip pins stay .zip because the cmake modules match those
filenames literally and fall through to the network otherwise (verified in
the 1.13.5 sources; pkgcheck's TarballAvailable is a false positive here and
says so inline). Dropped QA_PREBUILT, which claimed to cover prebuilt
libraries that this ebuild in fact compiles. Fixed a doubled Prefix in the
site-packages relocation: python_get_sitedir already carries EPREFIX and was
being passed to dodir and concatenated with ED. No -bin, deliberately.

llama-swap -- SRC_URI is now solely the upstream tarball; the adopted ebuild
took even the main source from a third party's personal account. Adds
-tags embed_ui, without which USE=ui ran the whole npm build and shipped a
binary with no UI. The systemd unit could never have worked: it used %i and
User=%i but installed as llama-swap.service, and systemd only expands %i for
a name containing '@' -- now installed via systemd_newunit as
llama-swap@.service, with the EnvironmentFile it expects shipped as an
example. Keeps ~arm64, unlike the ggml family: this is pure Go with no cgo
and no arch-specific path, and upstream builds linux-arm64 itself.

Uses ego mod download under RESTRICT="network-sandbox", matching
dev-util/trivy and dev-util/act. This is a KNOWING divergence from story 003
R4.1, which asked for a sandbox-clean default: upstream ships no vendor/
directory, so the alternative commits the overlay to hosting a generated
tarball on distfiles.obentoo.org and to hold = true forever after. The
alternative implementation is preserved and documented; the choice between
them should be made once, for this package and sci-ml/localai together.

All three records download-validate against upstream and pass the bentoo
checker. KEYWORDS is ~amd64 for the two C++ packages, consistent with the
decision taken across the ggml family: an unvalidated keyword is worse than
none.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>