fix(infra): use runtime image for gpu-warmup (Argo emissary compat)

busybox doesn't support Argo's emissary executor protocol, causing the
wait sidecar to hang indefinitely and hold the GPU allocation. Switch to
foxhunt-runtime which has proper /bin/sh and lets the sidecar detect
container exit. Also adds nvidia-smi check during warmup.

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

infra/k8s/argo/gpu-test-pipeline-template.yaml

@@ -61,6 +61,8 @@ spec:
             dependencies: [gpu-warmup]
 
     # ── gpu-warmup: trigger H100 autoscale ──
+    # Requests GPU to force autoscaler to provision node, then releases it.
+    # Uses the runtime image (not busybox) for Argo emissary executor compat.
     - name: gpu-warmup
       nodeSelector:
         k8s.scaleway.com/pool-name: "{{workflow.parameters.gpu-pool}}"
@@ -72,12 +74,13 @@ spec:
           operator: Exists
           effect: NoSchedule
       container:
-        image: busybox:1.37
+        image: gitlab-registry.foxhunt.svc.cluster.local:5000/root/foxhunt/foxhunt-runtime:latest
         command: ["/bin/sh", "-c"]
         args:
           - |
             echo "GPU warmup: triggering node autoscale..."
-            echo "GPU node scheduled, exiting to free resources"
+            nvidia-smi --query-gpu=name,memory.total --format=csv,noheader || true
+            echo "GPU node ready, releasing for compile-and-test"
         resources:
           requests:
             nvidia.com/gpu: "1"