feat(llm-serving): retire verifier-predictor, add grm (qwen2.5:3b)

Frees verifier's GPU from an underused vLLM PRM deployment. qwen2.5:3b-instruct moves off ornith-predictor's shared pod onto its own dedicated GPU (grm.yaml), so verification/judge traffic stops contending with ornith:35b's agent traffic. /v1/qwen/chat/completions now points at grm-predictor; path unchanged.

k8s/apps/llm-serving/ornith.yaml

@@ -38,12 +38,8 @@ spec:
 
         ollama pull ornith:35b
 
-        ollama pull qwen2.5:3b-instruct
-
         ollama run ornith:35b "ok" >/dev/null 2>&1 || true
 
-        ollama run qwen2.5:3b-instruct "ok" >/dev/null 2>&1 || true
-
         wait $SERVE_PID
 
         '
@@ -58,8 +54,11 @@ spec:
         value: '-1'
       - name: OLLAMA_NUM_PARALLEL
         value: '1'
+      # qwen2.5:3b-instruct moved to its own dedicated GPU (llm-serving/grm.yaml)
+      # so verification/judge traffic no longer contends with ornith:35b's
+      # planner/implementer traffic on this one -- single model here now.
       - name: OLLAMA_MAX_LOADED_MODELS
-        value: '2'
+        value: '1'
       image: ollama/ollama:0.32.9@sha256:1685741456770df6e3cceb2a945a5f75e020f658d1701509668d6f4688f1dd3f
       name: kserve-container
       ports:
@@ -70,8 +69,7 @@ spec:
           command:
           - /bin/sh
           - -c
-          - ollama ps 2>/dev/null | grep -q ornith && ollama ps 2>/dev/null |
-            grep -q qwen2.5
+          - ollama ps 2>/dev/null | grep -q ornith
         periodSeconds: 10
       resources:
         limits:
@@ -87,8 +85,7 @@ spec:
           command:
           - /bin/sh
           - -c
-          - ollama ps 2>/dev/null | grep -q ornith && ollama ps 2>/dev/null |
-            grep -q qwen2.5
+          - ollama ps 2>/dev/null | grep -q ornith
         failureThreshold: 120
         periodSeconds: 15
       volumeMounts: