langfuse generation spans with token usage; switch llm to vllm (qwen2.5-coder-7b on mcrndeb)

ctrl/k8s/base/api.yaml

@@ -13,14 +13,20 @@ spec:
       labels:
         app: api
     spec:
-      # lng.local.ar resolves to ::1 via the host's dnsmasq → coredns chain,
-      # which is the pod's own loopback (wrong). Override to the kind bridge
-      # gateway so traffic from this pod reaches the host's system Caddy and
-      # gets reverse-proxied to Langfuse on :3000.
       hostAliases:
+        # lng.local.ar resolves to ::1 via the host's dnsmasq → coredns chain,
+        # which is the pod's own loopback (wrong). Override to the kind bridge
+        # gateway so traffic from this pod reaches the host's system Caddy and
+        # gets reverse-proxied to Langfuse on :3000.
         - ip: "172.19.0.1"
           hostnames:
             - "lng.local.ar"
+        # mcrndeb is the GPU box on the LAN (vLLM lives there). kind-nvi
+        # pods reach it via routing through the host; we just need the name
+        # to resolve since coredns doesn't know about LAN hostnames.
+        - ip: "192.168.1.3"
+          hostnames:
+            - "mcrndeb"
       containers:
         - name: api
           image: nvi-api

ctrl/k8s/base/configmap.yaml

@@ -13,7 +13,13 @@ data:
 
   # LLM. Supported providers: groq, anthropic, openai.
   # Secrets (*_API_KEY) live in the Secret built by kustomize from .env.
-  LLM_PROVIDER: "groq"
+  # Flip LLM_PROVIDER to switch — the other configs stay so we can A/B easily.
+  LLM_PROVIDER: "openai"
+
+  # Local vLLM serving Qwen2.5-Coder-7B-Instruct-AWQ on mcrndeb.
+  OPENAI_BASE_URL: "http://mcrndeb:11000/v1"
+  OPENAI_MODEL: "Qwen/Qwen2.5-Coder-7B-Instruct-AWQ"
+
   GROQ_MODEL: "llama-3.3-70b-versatile"
   ANTHROPIC_MODEL: "claude-sonnet-4-6"