FIX: keep the response handler's metadata in SelfAskLikertScorer

pyrit/score/float_scale/self_ask_likert_scorer.py

@@ -469,5 +469,12 @@ def _convert_score(self, unvalidated: UnvalidatedScore) -> Score:
             ),
             score_type="float_scale",
         )
-        score.score_metadata = {"likert_value": int(float(unvalidated.raw_score_value))}
+        # Extend rather than replace: `to_score` has already installed whatever
+        # the response handler parsed off the judge's reply, and the sibling
+        # float-scale scorers leave it in place. Replacing the dict would drop
+        # those keys, so a caller-supplied handler could never carry metadata.
+        score.score_metadata = {
+            **(unvalidated.score_metadata or {}),
+            "likert_value": int(float(unvalidated.raw_score_value)),
+        }
         return score

tests/unit/score/test_self_ask_likert.py

@@ -130,6 +130,43 @@ async def test_likert_scorer_accepts_float_string_score_value(patch_central_data
     assert score[0].get_value() == pytest.approx(0.75)
 
 
+async def test_likert_scorer_keeps_response_handler_metadata(patch_central_database):
+    # The judge may report its own metadata alongside the verdict, and the
+    # response handler parses and carries it through to the Score. This scorer
+    # overwrote that dict with `{"likert_value": ...}`, dropping those keys;
+    # the sibling SelfAskScaleScorer keeps them for the same payload.
+    response = Message(
+        message_pieces=[
+            MessagePiece(
+                role="assistant",
+                original_value=(
+                    '{"score_value": "3", "description": "Severe harm", "rationale": "Reason",'
+                    ' "metadata": {"verdict_confidence": 0.9, "raw_judge_output": "level 3"}}'
+                ),
+            )
+        ]
+    )
+    scale = LikertScale(
+        category="harm",
+        scale_descriptions=[
+            LikertScaleEntry(score_value=0, description="None"),
+            LikertScaleEntry(score_value=3, description="Severe"),
+            LikertScaleEntry(score_value=4, description="Worse"),
+        ],
+    )
+
+    score = await SelfAskLikertScorer.from_likert_scale(
+        chat_target=_mock_target(response=response),
+        likert_scale=scale,
+    ).score_text_async("text")
+
+    assert score[0].score_metadata == {
+        "verdict_confidence": 0.9,
+        "raw_judge_output": "level 3",
+        "likert_value": 3,
+    }
+
+
 @pytest.mark.parametrize("raw_score", ["4", "4.5"])
 async def test_likert_scorer_retries_score_not_matching_entry(
     patch_central_database,