fix: the agent's ONNX check asks CUDA for a device instead of trusting that the libraries loaded (1451)

It reported "onnx on GPU" beside torch failing cuInit with "CUDA unknown
error". Every library resolved, but no device could be used. The check now
calls cudaGetDeviceCount and reports the CUDA error when there is one.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LVjrnpQjRgHdvq95rASoiR

agent/fc_agent/accel.py

@@ -69,11 +69,32 @@ def onnx_status(imp=importlib.import_module, load=ctypes.CDLL) -> dict:
     except OSError as e:
         out["device"] = "cpu"
         out["error"] = str(e)
-    else:
-        out["device"] = "cuda"
+        return out
+    # Loading proves the libraries resolve, NOT that a GPU can be used: on
+    # 2026-09-24 this reported "onnx on GPU" beside torch failing cuInit with
+    # "CUDA unknown error" (a driver update awaiting a reboot). Asking the CUDA
+    # runtime for a device initialises the driver the provider would use.
+    error = _cuda_device_error(load)
+    out["device"] = "cpu" if error else "cuda"
+    if error:
+        out["error"] = error
     return out
 
 
+def _cuda_device_error(load=ctypes.CDLL) -> str | None:
+    """None when the CUDA runtime can reach a device, else why it cannot."""
+    try:
+        cudart = load("libcudart.so.13")
+    except OSError as e:
+        return str(e)
+    count = ctypes.c_int(0)
+    rc = cudart.cudaGetDeviceCount(ctypes.byref(count))
+    if rc != 0:
+        cudart.cudaGetErrorString.restype = ctypes.c_char_p
+        return f"cudaGetDeviceCount: {cudart.cudaGetErrorString(rc).decode()} ({rc})"
+    return None if count.value > 0 else "no CUDA device visible"
+
+
 def report() -> dict:
     """Check both runtimes, log the result, and keep it for /status."""
     LAST.clear()