fix: increase /dev/shm for vLLM CPU test containers

vLLM 0.27.1's multiproc executor allocates its inter-process message
queue in /dev/shm and now requires 160 MiB, but Docker's default
/dev/shm is only 64 MiB. As a result EngineCore fails to start with
"Insufficient space in /dev/shm" and the vLLM server never binds to
its port, causing the inference/embedding readiness checks to time out.

Pass --shm-size=2g to the test container so the engine has enough
shared memory to start.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: nathan-weinberg

.github/actions/setup-vllm/action.yml

@@ -29,9 +29,13 @@ runs:
         fi
 
         # Start vllm container
+        # --shm-size: vLLM's multiproc executor allocates its inter-process
+        # message queue in /dev/shm; the Docker default of 64 MiB is too small
+        # (v0.27.1 requires 160 MiB) and EngineCore fails to start without this.
         docker run -d \
           --name vllm-$VLLM_MODE \
           --net=host \
+          --shm-size=2g \
           $VLLM_IMAGE \
           $VLLM_ARGS