chore(cf-orch): route recipe LLM calls through vllm with model candidates + CF_APP_NAME

Switches recipe generation service type from 'cf-text' to 'vllm' so the
coordinator can route to quantized small models (Qwen2.5-3B, Phi-4-mini)
rather than the full text backend. Passes CF_APP_NAME for per-product
VRAM/request analytics in the coordinator dashboard.

- llm_recipe.py: _SERVICE_TYPE = 'vllm'; _MODEL_CANDIDATES list; passes
  model_candidates and pipeline= to CFOrchClient.allocate()
- compose.cloud.yml: CF_APP_NAME=kiwi env var for coordinator attribution

app/services/recipe/llm_recipe.py

@@ -149,7 +149,8 @@ class LLMRecipeGenerator:
 
         return "\n".join(lines)
 
-    _SERVICE_TYPE = "cf-text"
+    _SERVICE_TYPE = "vllm"
+    _MODEL_CANDIDATES = ["Qwen2.5-3B-Instruct", "Phi-4-mini-instruct"]
     _TTL_S = 300.0
     _CALLER = "kiwi-recipe"
 
@@ -167,8 +168,10 @@ class LLMRecipeGenerator:
                 client = CFOrchClient(cf_orch_url)
                 return client.allocate(
                     service=self._SERVICE_TYPE,
+                    model_candidates=self._MODEL_CANDIDATES,
                     ttl_s=self._TTL_S,
                     caller=self._CALLER,
+                    pipeline=os.environ.get("CF_APP_NAME") or None,
                 )
             except Exception as exc:
                 logger.debug("CFOrchClient init failed, falling back to direct URL: %s", exc)

compose.cloud.yml

@@ -21,6 +21,8 @@ services:
       CLOUD_AUTH_BYPASS_IPS: ${CLOUD_AUTH_BYPASS_IPS:-}
       # cf-orch: route LLM calls through the coordinator for managed GPU inference
       CF_ORCH_URL: http://host.docker.internal:7700
+      # Product identifier for coordinator analytics — per-product VRAM/request breakdown
+      CF_APP_NAME: kiwi
       # Community PostgreSQL — shared across CF products; unset = community features unavailable (fail soft)
       COMMUNITY_DB_URL: ${COMMUNITY_DB_URL:-}
       COMMUNITY_PSEUDONYM_SALT: ${COMMUNITY_PSEUDONYM_SALT:-}