fix: preserve classifier parse failures through compliance grading

manifests/context-packs/skill-triggers@1.json

@@ -10,7 +10,7 @@
     "id": "hand-seeded",
     "source": "manual-curation-pending-regeneration"
   },
-  "registryDigest": "5aa40d350c7cfdd74c4c432383b24bbdcba34dde0bf98ba04fe969a0b6a5fdd7",
+  "registryDigest": "b21e33d7de6c1f8c1934a7b11d9cb9c5135abb5fd51c7db155eb393cff5fa959",
   "schemaVersion": 1,
   "triggers": {
     "skill:agent-runtime-gateway-smoke-test": [

skills/skill-comply/scripts/classifier.py

@@ -14,6 +14,10 @@ from scripts.parser import ComplianceSpec, ObservationEvent
 PROMPTS_DIR = Path(__file__).parent.parent / "prompts"
 
 
+class ClassifierParseError(RuntimeError):
+    """Raised when the classifier output cannot be parsed."""
+
+
 def classify_events(
     spec: ComplianceSpec,
     trace: list[ObservationEvent],
@@ -73,13 +77,15 @@ def _parse_classification(text: str) -> dict[str, list[int]]:
     try:
         parsed = json.loads(cleaned)
         if not isinstance(parsed, dict):
-            logger.warning("Classifier returned non-dict JSON: %s", type(parsed).__name__)
-            return {}
+            raise ClassifierParseError(
+                f"classifier returned non-dict JSON: {type(parsed).__name__}"
+            )
         return {
             k: [int(i) for i in v]
             for k, v in parsed.items()
             if isinstance(v, list)
         }
     except (json.JSONDecodeError, ValueError, TypeError) as e:
-        logger.warning("Failed to parse classification output: %s", e)
-        return {}
+        raise ClassifierParseError(
+            f"failed to parse classification output: {e}"
+        ) from e

skills/skill-comply/tests/test_classifier.py

@@ -0,0 +1,55 @@
+"""Tests for scripts.classifier._parse_classification."""
+
+import subprocess
+from unittest.mock import patch
+
+import pytest
+from scripts import classifier
+from scripts.classifier import _parse_classification
+from scripts.grader import grade
+from scripts.parser import ComplianceSpec, Detector, ObservationEvent, Step
+
+
+class TestParseClassification:
+    def test_plain_json(self):
+        assert _parse_classification('{"a": [0, 2]}') == {"a": [0, 2]}
+
+    def test_fenced_json(self):
+        assert _parse_classification('```json\n{"a": [1]}\n```') == {"a": [1]}
+
+    def test_dict_without_list_values_is_empty(self):
+        assert _parse_classification('{"a": "x"}') == {}
+
+    def test_empty_object_is_valid(self):
+        assert _parse_classification("{}") == {}
+
+    def test_garbage_raises(self):
+        with pytest.raises(classifier.ClassifierParseError, match="failed to parse"):
+            _parse_classification("sorry, I cannot classify these")
+
+    def test_non_dict_json_raises(self):
+        with pytest.raises(classifier.ClassifierParseError, match="non-dict"):
+            _parse_classification("[1, 2, 3]")
+
+    def test_non_integer_indices_raise(self):
+        with pytest.raises(classifier.ClassifierParseError, match="failed to parse"):
+            _parse_classification('{"a": ["x"]}')
+
+    def test_is_runtime_error(self):
+        assert issubclass(classifier.ClassifierParseError, RuntimeError)
+
+
+@pytest.mark.parametrize("response", ["not JSON", "[]", '{"a": ["x"]}'])
+def test_grader_propagates_classifier_parse_failure(response):
+    """An invalid successful subprocess response cannot become a zero score."""
+    spec = ComplianceSpec(
+        id="fixture", name="Fixture", source_rule="fixture", version="1",
+        steps=(Step("a", "Read a file", True, Detector("Read a file")),),
+        threshold_promote_to_hook=0.6,
+    )
+    trace = [ObservationEvent("T0000", "tool_complete", "Read", "fixture", "", "")]
+    result = subprocess.CompletedProcess(["claude"], 0, stdout=response, stderr="")
+    with patch("scripts.classifier.subprocess.run", return_value=result) as execute:
+        with pytest.raises(classifier.ClassifierParseError):
+            grade(spec, trace)
+    execute.assert_called_once()