fix(warmup): fire vLLM warmup before planner so RunPod loads during planner+stones

app/fsm.py

@@ -1550,24 +1550,6 @@ def iter_steps(query: str):
     app = build_app(query, step_queue=q)
     final_state_holder: dict[str, Any] = {}
 
-    # Fire a tiny warmup request to the LLM backend immediately so that a
-    # cold RunPod pod starts loading the model while the ~35s specialist
-    # pipeline runs. By the time step_reconcile starts, the GPU should be
-    # warm and the first token arrives quickly.
-    def _warmup_llm():
-        try:
-            from app import llm as _llm
-            _llm.chat(
-                model="granite-8b",
-                messages=[{"role": "user", "content": "hi"}],
-                options={"num_predict": 1, "temperature": 0},
-                stream=False,
-            )
-        except Exception:
-            pass  # warmup failure is silent — reconciler will still try
-
-    _threading.Thread(target=_warmup_llm, daemon=True, name="riprap-warmup").start()
-
     # Threadlocals are per-thread; the request thread (single_address.run
     # / neighborhood.run) sets the strict-mode flag, planner specialist
     # set, and token / Mellea-attempt callbacks, but Burr's app.iterate