test(local-runtime): trim idle-probe salvage to two invariants

Keep the two tests that were red on origin/main (probe failure keeps the clock and
unloads once telemetry recovers; confirmed busy still resets). Drop the is_idle() bool
contract test — it pins behaviour that did not change — and fold the probe-failure log
call onto two lines.

hermes_cli/local_runtime/supervisor.py

@@ -390,12 +390,8 @@ class LlamaServerSupervisor:
                 continue
             probe = self._probe_idle(model_id)
             if probe is None:
-                idle_for = int(now - self._idle_since.get(model_id, now))
-                logger.info(
-                    "idle probe for %s failed; keeping idle clock (idle %ds)",
-                    model_id,
-                    idle_for,
-                )
+                logger.info("idle probe for %s failed; keeping idle clock (idle %ds)", model_id,
+                            int(now - self._idle_since.get(model_id, now)))
                 continue
             if probe is False:
                 self._idle_since.pop(model_id, None)

tests/hermes_cli/test_local_runtime.py

@@ -645,19 +645,6 @@ def test_idle_sweep_busy_after_probe_failure_still_resets_clock(
     assert handler.unloaded == []
 
 
-def test_is_idle_probe_failure_reports_not_idle(stub_server, tmp_path):
-    """The public bool contract is unchanged: a failed probe reads as "not idle" —
-    callers outside the sweep never unload on a telemetry hiccup."""
-    port, handler = stub_server
-    sup = _make_supervisor(tmp_path, port)
-    handler.models = {"data": [{"id": "m", "status": {"value": "loaded"}}]}
-    handler.slots = []
-    handler.metrics_error = 503
-    assert sup.is_idle() is False
-    handler.metrics_error = 0
-    assert sup.is_idle() is True
-
-
 def test_staged_models_requires_every_split_part(tmp_path, monkeypatch):
     """A split GGUF mid-download must NOT count as staged: the picker, the
     catalog's 'downloaded' flag, and the router's model list all read