feat: a GPU agent on the CPU shows as degraded, in the System view and on its own page (4410)

torch and onnxruntime both fall back to the CPU without raising, so the agent
that ran CPU-bound for weeks after a driver update leased and checked in like
a healthy one.

- The agent sends its startup accel report on every lease and heartbeat.
- The server keeps a bounded copy on the roster row. A running agent with a
  runtime off the GPU becomes `degraded`, with a sentence naming the runtime
  and the reason.
- The top nav shows it amber.
- The agent page carries a banner, and its pill reads "CPU only".

Also: the bandwidth field gets the page's − / + stepper, and both number
fields drop the browser's spin arrows.

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

tests/test_agent_accel.py

@@ -103,3 +103,20 @@ def test_a_missing_runtime_is_reported_not_raised():
 
     assert accel.torch_status(imp)["device"] == "unavailable"
     assert accel.onnx_status(imp)["device"] == "unavailable"
+
+
+def test_summary_is_what_the_server_stores(monkeypatch):
+    """Device plus a bounded reason, per runtime — sent on every lease."""
+    monkeypatch.setattr(accel, "LAST", {
+        "torch": {"version": "2.14.0", "device": "cuda", "gpu": "RTX"},
+        "onnx": {"version": "1.30.0", "device": "cpu", "error": "e" * 500},
+    })
+    assert accel.summary() == {
+        "torch": {"device": "cuda"},
+        "onnx": {"device": "cpu", "error": "e" * 200},
+    }
+
+
+def test_summary_before_the_report_is_none(monkeypatch):
+    monkeypatch.setattr(accel, "LAST", {})
+    assert accel.summary() is None

tests/test_agent_degraded.py

@@ -0,0 +1,59 @@
+"""A GPU agent that fell back to the CPU reads as DEGRADED, not running (#4410).
+
+Both runtimes fall back without raising, so a CPU-bound agent leases, works
+and checks in exactly like a healthy one. On 2026-09-24 one had been doing so
+since a driver update left a stale CDI spec; the only sign was a line in the
+agent's own log. The agent now sends its startup report on every lease and
+heartbeat, and the System view derives the state from it.
+"""
+
+from __future__ import annotations
+
+from backend.app.api.gpu import _accel_detail
+from backend.app.api.system_health import _SEVERITY, _learned_state
+
+GPU = {"torch": {"device": "cuda"}, "onnx": {"device": "cuda"}}
+CPU = {
+    "torch": {"device": "cpu"},
+    "onnx": {"device": "cpu", "error": "cudaGetDeviceCount: unknown error (999)"},
+}
+
+
+def test_a_running_agent_on_the_gpu_is_ok():
+    state, detail = _learned_state("GPU agent", "ok", 5, {"accel": GPU})
+    assert state == "ok"
+    assert detail == "GPU agent is running"
+
+
+def test_a_running_agent_on_the_cpu_is_degraded_and_says_why():
+    state, detail = _learned_state("GPU agent", "ok", 5, {"accel": CPU})
+    assert state == "degraded"
+    assert "onnx (cudaGetDeviceCount: unknown error (999))" in detail
+    assert "torch (cpu)" in detail
+
+
+def test_an_agent_that_never_reported_is_not_called_degraded():
+    """An older agent build sends no `accel`: that is not-yet-known, not slow."""
+    assert _learned_state("GPU agent", "ok", 5, {})[0] == "ok"
+
+
+def test_stopped_outranks_degraded():
+    """Whether a quiet agent is still running is the more urgent question."""
+    state, _ = _learned_state("GPU agent", "down", 900, {"accel": CPU})
+    assert state == "down"
+    assert _SEVERITY["stale"] > _SEVERITY["degraded"] > _SEVERITY["unknown"]
+
+
+def test_the_lease_keeps_only_a_bounded_report():
+    """Written on every lease by a client the server does not control."""
+    body = {"accel": {
+        "torch": {"device": "cpu", "error": "x" * 5000, "extra": "dropped"},
+        "onnx": "not a dict",
+    }}
+    kept = _accel_detail(body)["accel"]
+    assert kept == {"torch": {"device": "cpu", "error": "x" * 200}}
+
+
+def test_a_lease_without_a_report_adds_nothing():
+    assert _accel_detail({}) == {}
+    assert _accel_detail({"accel": None}) == {}