Record CUDA benchmark diagnostics

packages/node/meshnet_node/hardware.py

@@ -192,7 +192,7 @@ def detect_hardware() -> dict:
     }
 
 
-def benchmark_throughput(device_str: str = "cpu") -> float:
+def benchmark_throughput_checked(device_str: str = "cpu") -> tuple[float, bool, str | None]:
     """
     Estimate compute throughput via a synthetic transformer GEMM benchmark.
 
@@ -201,7 +201,8 @@ def benchmark_throughput(device_str: str = "cpu") -> float:
     The value is used as benchmark_tokens_per_sec in tracker registration for
     routing tiebreaks; it is not an absolute token rate.
 
-    Falls back to 1.0 if torch is unavailable.
+    Returns (score, ok, error). Score falls back to 1.0 when the requested
+    device cannot run the benchmark.
     """
     try:
         import torch  # type: ignore[import]
@@ -233,6 +234,12 @@ def benchmark_throughput(device_str: str = "cpu") -> float:
         _sync()
         elapsed = time.perf_counter() - t0
 
-        return round(n_iters / max(elapsed, 1e-9), 2)
-    except Exception:
-        return 1.0
+        return round(n_iters / max(elapsed, 1e-9), 2), True, None
+    except Exception as exc:
+        return 1.0, False, f"{type(exc).__name__}: {exc}"
+
+
+def benchmark_throughput(device_str: str = "cpu") -> float:
+    """Return only the numeric throughput index, preserving the legacy API."""
+    score, _ok, _error = benchmark_throughput_checked(device_str)
+    return score