Record CUDA benchmark diagnostics

tests/conftest.py

@@ -14,6 +14,6 @@ def _stub_benchmark_throughput(monkeypatch):
     """
     try:
         import meshnet_node.startup as startup_mod
-        monkeypatch.setattr(startup_mod, "benchmark_throughput", lambda _device: 999.0)
+        monkeypatch.setattr(startup_mod, "benchmark_throughput_checked", lambda _device: (999.0, True, None))
     except ImportError:
         pass

tests/test_node_startup.py

@@ -173,7 +173,7 @@ def test_benchmark_throughput_is_registered_in_payload(monkeypatch, tmp_path):
 
     monkeypatch.setattr(startup_mod, "detect_hardware",
                         lambda: {"device": "cpu", "gpu_name": None, "vram_mb": 0, "ram_mb": 16384})
-    monkeypatch.setattr(startup_mod, "benchmark_throughput", lambda _device: 42.5)
+    monkeypatch.setattr(startup_mod, "benchmark_throughput_checked", lambda _device: (42.5, True, None))
     monkeypatch.setattr(startup_mod, "TorchNodeServer", lambda **_kw: FakeNode())
     monkeypatch.setattr(startup_mod, "_detect_num_layers", lambda _model_id: 24)
     monkeypatch.setattr(startup_mod, "RelayHttpBridge", None)
@@ -193,6 +193,67 @@ def test_benchmark_throughput_is_registered_in_payload(monkeypatch, tmp_path):
     node.stop()
 
     assert captured.get("benchmark_tokens_per_sec") == 42.5
+    assert captured["hardware_profile"]["benchmark_device"] == "cpu"
+    assert captured["hardware_profile"]["benchmark_ok"] is True
+
+
+def test_cuda_benchmark_failure_is_registered_for_inventory_only_gpu(monkeypatch, tmp_path, capsys):
+    import meshnet_node.startup as startup_mod
+
+    captured: dict = {}
+
+    class FakeNode:
+        backend = None
+
+        def start(self):
+            return 7099
+
+        def stop(self):
+            pass
+
+    def fake_benchmark(device):
+        if device == "cuda":
+            return 1.0, False, "AssertionError: Torch not compiled with CUDA enabled"
+        return 55.0, True, None
+
+    monkeypatch.setattr(
+        startup_mod,
+        "detect_hardware",
+        lambda: {
+            "device": "cpu",
+            "gpu_name": "NVIDIA GeForce RTX 4060 Laptop GPU",
+            "vram_mb": 8188,
+            "dedicated_vram_mb": 8188,
+            "shared_vram_mb": 40555,
+            "ram_mb": 81111,
+            "cuda_available": False,
+        },
+    )
+    monkeypatch.setattr(startup_mod, "benchmark_throughput_checked", fake_benchmark)
+    monkeypatch.setattr(startup_mod, "TorchNodeServer", lambda **_kw: FakeNode())
+    monkeypatch.setattr(startup_mod, "_post_json",
+                        lambda _url, payload, timeout=10.0: (captured.update(payload) or {"node_id": "x"}))
+    monkeypatch.setattr(startup_mod, "_start_heartbeat", lambda *a, **kw: None)
+
+    node = run_startup(
+        tracker_url="http://localhost:8080",
+        model_id="Qwen/Qwen2.5-0.5B-Instruct",
+        shard_start=0,
+        shard_end=23,
+        wallet_path=tmp_path / "wallet.json",
+    )
+    node.stop()
+
+    output = capsys.readouterr().out
+    assert "CUDA benchmark unavailable" in output
+    assert "Hardware:     CPU (CUDA inactive)" in output
+    hw = captured["hardware_profile"]
+    assert hw["cuda_benchmark_ok"] is False
+    assert "Torch not compiled with CUDA enabled" in hw["cuda_benchmark_error"]
+    assert hw["benchmark_device"] == "cpu"
+    assert hw["benchmark_ok"] is True
+    assert captured["ram_bytes"] == 81111 * 1024 * 1024
+    assert captured["vram_bytes"] == 0
 
 
 def test_wallet_generates_new_keypair(tmp_path):