fix(docker): bundle llama-mtmd-cli shared libraries and register with ldconfig

The pre-built Ubuntu llama.cpp binaries are dynamically linked against
libllama.so, libmtmd.so, etc. shipped in the same release tarball. We
were extracting only the CLI binary, leaving the shared libs behind, which
caused "missing shared library" failures at runtime.

Fix: also extract all *.so* files into /usr/local/lib/llama/ in the minicpm
stage, COPY them into the runtime image, and register them via ldconfig.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: jkyberneees

docker/Dockerfile

@@ -71,7 +71,9 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
       curl ca-certificates \
  && rm -rf /var/lib/apt/lists/*
 # Map dpkg architecture to the llama.cpp release asset name suffix
-# (amd64 → x64, arm64 → arm64) and extract only llama-mtmd-cli.
+# (amd64 → x64, arm64 → arm64). Extract the CLI binary and all bundled
+# shared libraries (.so*) — the pre-built Ubuntu binaries are dynamically
+# linked against libllama.so, libmtmd.so, etc. shipped in the same tarball.
 RUN set -ex \
  && LLAMA_ARCH=$(dpkg --print-architecture | sed 's/amd64/x64/') \
  && curl -fsSL \
@@ -80,6 +82,8 @@ RUN set -ex \
  && mkdir -p /tmp/llama-bin \
  && tar -xzf /tmp/llama.tar.gz -C /tmp/llama-bin \
  && find /tmp/llama-bin -name "llama-mtmd-cli" -exec install -m 755 {} /usr/local/bin/llama-mtmd-cli \; \
+ && mkdir -p /usr/local/lib/llama \
+ && find /tmp/llama-bin -name "*.so*" -exec cp -P {} /usr/local/lib/llama/ \; \
  && rm -rf /tmp/llama.tar.gz /tmp/llama-bin
 # Fetch the model GGUF and vision projector into a fixed image path (NOT under
 # ~/.odek — bind-mount profiles would hide files baked there). The runtime
@@ -122,8 +126,12 @@ COPY --from=whisper /models/ /usr/local/share/whisper/models/
 
 # Bundle llama-mtmd-cli + MiniCPM-V 4.6 from the minicpm stage so `vision`
 # works with zero setup. The tool auto-detects /usr/local/share/minicpm-v/models.
+# The shared libraries (libllama.so, libmtmd.so, …) are registered via ldconfig
+# so the dynamically-linked binary resolves them at runtime.
 COPY --from=minicpm /usr/local/bin/llama-mtmd-cli /usr/local/bin/llama-mtmd-cli
+COPY --from=minicpm /usr/local/lib/llama/ /usr/local/lib/llama/
 COPY --from=minicpm /usr/local/share/minicpm-v/models/ /usr/local/share/minicpm-v/models/
+RUN echo "/usr/local/lib/llama" > /etc/ld.so.conf.d/llama.conf && ldconfig
 
 # ── Adding extra dependencies the agent can use ──────────────────────────
 # The agent runs shell commands INSIDE this image, so any runtime or CLI it