fix: update Gitea token reference in documentation and workflows to use RELEASE_TOKEN

packaging/GITEA-SETUP.md

@@ -65,9 +65,19 @@ that only serves this purpose, worth knowing about regardless.
 Releases are published through Gitea's own API by
 `packaging/publish_gitea.py`, so the job needs a token:
 
-1. **Settings → Applications → Generate New Token**, scope `write:repository`.
+1. **Settings → Applications → Generate New Token.** Give it a name, leave
+   everything at *No Access* except **repository**, set to *Read and Write*.
+   Copy the value shown — Gitea displays it exactly once.
 2. In the repository: **Settings → Actions → Secrets → Add Secret**, named
-   `GITEA_TOKEN`.
+   `RELEASE_TOKEN`, value the token from step 1.
+
+   Not `GITEA_TOKEN`. Gitea rejects any secret name starting with `GITEA_` or
+   `GITHUB_` — both prefixes are reserved for its own built-in variables, and
+   the validation pattern is exactly `^(?!GITEA_|GITHUB_)[a-zA-Z_][a-zA-Z0-9_]*$`.
+   `release.yml` reads `secrets.RELEASE_TOKEN` and maps it to a
+   `GITEA_TOKEN` environment variable inside the job — that env var name is
+   unrelated to the secret name and can be anything, since it's just what
+   `publish_gitea.py` reads from `os.environ`.
 
 `GITEA_URL` and `GITEA_REPOSITORY` come from `github.server_url` and
 `github.repository`, which Gitea populates for compatibility. Nothing else to

packaging/LM-Gambit.sh

@@ -69,15 +69,30 @@ EOF
         exit 1
     fi
 
+    # llama-cpp-python ships only an sdist to PyPI, so a plain install compiles
+    # it from source. On Linux that means a full toolchain and a long wait, and
+    # the upstream index below carries prebuilt manylinux wheels instead.
+    #
+    # Deliberately NOT used on macOS. The PyPI wheel there already carries
+    # Metal support, and adding a CPU-only index risks pip picking it on a tie,
+    # quietly costing every Apple Silicon user their GPU for no visible reason.
+    if [ "$(uname -s)" = "Linux" ]; then
+        PIP_INDEX="--extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu"
+    else
+        PIP_INDEX=""
+    fi
+
     echo "Installing dependencies..."
     "$VENV/bin/python" -m pip install --quiet --upgrade pip
-    if ! "$VENV/bin/python" -m pip install -r "$HERE/requirements.txt"; then
+    # shellcheck disable=SC2086  # intentional word splitting; the URL has no spaces
+    if ! "$VENV/bin/python" -m pip install $PIP_INDEX -r "$HERE/requirements.txt"; then
         cat >&2 <<EOF
 
 Dependency installation failed.
 
-The usual culprit is llama-cpp-python, which compiles from source when no
-prebuilt wheel matches this platform. It needs a C compiler:
+If the error above mentions llama-cpp-python, cmake or a missing compiler,
+then pip found no prebuilt wheel and fell back to building from source. That
+needs a C toolchain:
 
   macOS   xcode-select --install
   Linux   sudo apt install build-essential cmake