fix: Don't pass the model-holding agent as a @spaces.GPU argument

Root cause of "_share_cuda_: only available on CUDA": ZeroGPU's docs
example always accesses the model via closure/module scope from
inside the decorated function -- never as a parameter. We were passing
the whole ControlAIAgent (holding the loaded CUDA model) as an
argument on every call, which appears to force ZeroGPU's argument
marshaling to try to share the model's CUDA tensors across its process
boundary and fail. Both GPU-wrapped functions now call get_agent()
internally instead, matching the documented pattern exactly.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

app.py

@@ -83,13 +83,13 @@ def get_agent() -> ControlAIAgent:
 # releases it afterward). Outside of a ZeroGPU Space this decorator is a
 # harmless no-op, so it's safe to always wrap these.
 @spaces.GPU(duration=300)
-def _run_stream_on_gpu(agent: ControlAIAgent, message: str, history: list[dict[str, str]]):
-    yield from agent.run_stream(message, history=history)
+def _run_stream_on_gpu(message: str, history: list[dict[str, str]]):
+    yield from get_agent().run_stream(message, history=history)
 
 
 @spaces.GPU(duration=300)
-def _run_on_gpu(agent: ControlAIAgent, message: str, history: list[dict[str, str]]):
-    return agent.run(message, history=history, verbose=False)
+def _run_on_gpu(message: str, history: list[dict[str, str]]):
+    return get_agent().run(message, history=history, verbose=False)
 
 
 # ZeroGPU's startup check statically looks for a @spaces.GPU function wired
@@ -202,12 +202,10 @@ async def chat_stream_endpoint(req: ChatRequest):
     if not req.message.strip():
         raise HTTPException(status_code=400, detail="Message cannot be empty")
 
-    agent = get_agent()
-
     def event_generator():
         try:
             with inference_lock:
-                for event in _run_stream_on_gpu(agent, req.message.strip(), req.history):
+                for event in _run_stream_on_gpu(req.message.strip(), req.history):
                     yield f"data: {json.dumps(event, ensure_ascii=False)}\n\n"
         except Exception as exc:
             yield f"data: {json.dumps({'type': 'error', 'error': str(exc)}, ensure_ascii=False)}\n\n"
@@ -230,9 +228,8 @@ async def chat_endpoint(req: ChatRequest) -> ChatResponse:
 
     t0 = time.time()
     try:
-        agent = get_agent()
         with inference_lock:
-            result = _run_on_gpu(agent, req.message.strip(), req.history)
+            result = _run_on_gpu(req.message.strip(), req.history)
         elapsed = time.time() - t0
 
         # Collect tool traces