commit bd292253458586790bec3e19373709ce4250366c
Author: lucascouts <lucascs@protonmail.com>
Date: Thu Sep 10 21:27:15 2026 -0300

add(dev-util/goose-1.50.0), add(metadata/md5-cache/dev-util/goose-1.50.0), mod(.autoupdate/packages.toml)

goose built from the upstream source tarball, completing the pair with
dev-util/goose-bin. They do not collide: this installs /usr/bin/goose, the
binary package installs under /opt.

1324 crates plus four git crates from two repositories. Two things here are
not boilerplate and should survive future bumps:

* src_prepare rewrites upstream's own [patch.crates-io]. That table points
agent-client-protocol, agent-client-protocol-http and cudaforge at git
URLs, and cargo ADDS it to the [patch] cargo.eclass writes instead of
letting the eclass win -- so cargo still reached for github and died with
"can't checkout ... you are in the offline mode". Rewriting those entries
to the already-unpacked paths is what removes the network access. Every
substitution is asserted afterwards, because sed exits 0 when it matches
nothing: an upstream reshuffle must fail the build loudly rather than
quietly restore the fetch.

* There is no REQUIRED_USE tying cuda and vulkan to local-inference, and
that is deliberate. The obvious "vulkan? ( local-inference )" failed
RequiredUseDefaults on 36 stock profiles, because vulkan and cuda are
global USE flags the desktop profiles enable by default -- it would turn
an ordinary desktop USE setting into a refusal to emerge. Both are folded
in under "if use local-inference" instead, and metadata.xml says they
have no effect without it.

The feature set deviates from upstream's defaults on purpose. local-inference
is off (it pulls candle plus llama-cpp-sys-2, which builds a whole llama.cpp
through cmake) and telemetry is off (that is the user's call, not ours).
"update" is never enabled: it is a self-updater that would overwrite a
Portage-managed binary behind Portage's back, and since it sits in upstream's
default list with no switch to disable it, --no-default-features is what keeps
it out.

llama-cpp-sys-2 links llama.cpp statically into the binary, so unlike the
add_subdirectory(ggml) packages it installs no libggml*.so and no headers --
there is no collision with sci-ml/llama-cpp.

Verified, not assumed: pkgcheck reports 0 findings; the Manifest carries 1327
DIST entries (1324 crates + source tarball + 2 git tarballs), which is the
check that proves SRC_URI+=" $" actually took; and the
package completed compile and install without root with FEATURES=network-
sandbox ON -- the run that proves the CRATES block is complete, since a
missing crate becomes a network error rather than a silent download. The
installed /usr/bin/goose runs and reports 1.50.0, and "goose --help" shows no
update subcommand, confirming the feature exclusion took effect.

Not verified: no merge onto the live system -- this host has no sudo -- so
directory ownership is untested. Only USE="-local-inference" was built; the
cuda and vulkan paths are unexercised.

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