tests/test_classifier_modes_dispatch.py

@@ -924,14 +924,16 @@ def test_local_decision_connection_error_opens_circuit_unmetered(monkeypatch):
     assert got1.source == "fallback"
     assert dispatcher._last_classifier_failure > 0.0
     assert dispatcher._classifier_backoff_active() is True
+    first_failure_at = dispatcher._last_classifier_failure
 
     # Second call — should be SKIPPED (post called exactly once total)
     got2 = dispatcher.classify("another task", None)
     assert call_count[0] == 1, f"post called {call_count[0]} times, expected 1"
     assert got2.source == "fallback"
 
-    # _last_classifier_failure unchanged by second call
-    assert dispatcher._last_classifier_failure > 0.0  # still set from first call
+    # A skip must never re-stamp the clock, or the circuit stays open forever
+    # while traffic flows. Equality, not just "still set".
+    assert dispatcher._last_classifier_failure == first_failure_at
 
 
 def test_local_decision_connection_error_opens_circuit_metered(monkeypatch):
@@ -968,16 +970,19 @@ def test_local_decision_connection_error_opens_circuit_metered(monkeypatch):
     assert got1.source == "fallback"
     assert dispatcher._last_classifier_failure > 0.0
     assert dispatcher._classifier_backoff_active() is True
+    first_failure_at = dispatcher._last_classifier_failure
 
     # Second call — should be SKIPPED (post called exactly once total)
     got2 = dispatcher.classify("another task", None)
     assert call_count[0] == 1, f"post called {call_count[0]} times, expected 1"
     assert got2.source == "fallback"
+    assert dispatcher._last_classifier_failure == first_failure_at
 
 
-def test_local_decision_low_confidence_does_not_open_circuit(monkeypatch):
-    """A successful classify with low confidence is a real answer, not a
-    transport failure — the circuit must stay closed."""
+def test_local_decision_below_confidence_min_does_not_open_circuit(monkeypatch):
+    """A verdict below classifier.decision.confidence_min is the dispatcher's
+    own RuntimeError, raised after a healthy answer. The endpoint is fine, so
+    the circuit must stay closed. (The coverage-floor path is the next test.)"""
     monkeypatch.setattr(dispatcher.cfg.classifier, "mode", "local_decision")
     monkeypatch.setattr(dispatcher.cfg.classifier, "fallback_tier", 2)
     monkeypatch.setattr(
@@ -987,8 +992,46 @@ def test_local_decision_low_confidence_does_not_open_circuit(monkeypatch):
     monkeypatch.setattr(dispatcher.cfg.local_compute, "enabled", True)
 
     def fake_post(*a, **k):
-        # Return a response where the winning token has very low logprob
-        # → confidence below 0.5 → below confidence_min → RuntimeError
+        # Four letters with equal mass: coverage is about 0.99 (well above
+        # coverage_min), but the winner holds only a quarter of it, so
+        # confidence is about 0.25, below confidence_min (0.5).
+        return SimpleNamespace(
+            json=lambda: {
+                "logprobs": [{
+                    "top_logprobs": [
+                        {"token": letter, "logprob": -1.4}
+                        for letter in ("A", "B", "C", "D")
+                    ]
+                }],
+            },
+            raise_for_status=lambda: None,
+        )
+
+    monkeypatch.setattr(local_decision.requests, "post", fake_post)
+
+    got = dispatcher.classify("refactor this function", None)
+    assert got.source == "fallback"  # cascaded on the confidence failure
+
+    assert dispatcher._last_classifier_failure == 0.0
+    assert dispatcher._classifier_backoff_active() is False
+
+
+def test_local_decision_coverage_floor_does_not_open_circuit(monkeypatch):
+    """parse_logprobs raising because the model put almost no mass on any
+    option letter (coverage below coverage_min) is also not a transport
+    failure — the circuit must stay closed."""
+    monkeypatch.setattr(dispatcher.cfg.classifier, "mode", "local_decision")
+    monkeypatch.setattr(dispatcher.cfg.classifier, "fallback_tier", 2)
+    monkeypatch.setattr(
+        dispatcher.cfg.classifier, "decision",
+        _decision_conf(confidence_min=0.5, tier_enabled=False),
+    )
+    monkeypatch.setattr(dispatcher.cfg.local_compute, "enabled", True)
+
+    def fake_post(*a, **k):
+        # "Z" is not an option letter and A/B carry almost no mass, so total
+        # option mass (about 0.012) is below coverage_min (0.3) and
+        # parse_logprobs raises RuntimeError before any confidence is computed.
         return SimpleNamespace(
             json=lambda: {
                 "model": "qwen3.5:4b",