test: add edge-case tests for training example metadata preservation

- test_metadata_with_future_fields_preserved: unknown fields pass through
- test_metadata_preserved_across_multiple_examples: per-example independence

Verifies fix for #646.

tests/test_tasks_core.py

@@ -406,6 +406,34 @@ class TestNormalizeTrainingExamples:
         assert "category" not in result[0]
 
 
+    def test_metadata_with_future_fields_preserved(self):
+        """Unknown future metadata fields pass through untouched."""
+        examples = [
+            {"prompt": "Q1", "response": "A1", "curriculum_level": 3, "audit_id": "xyz"},
+        ]
+        result = normalize_training_examples(
+            examples, "b001", ["t1"], "fp", "fr"
+        )
+        assert result[0]["curriculum_level"] == 3
+        assert result[0]["audit_id"] == "xyz"
+
+    def test_metadata_preserved_across_multiple_examples(self):
+        """Each example's metadata is independent."""
+        examples = [
+            {"prompt": "Q1", "response": "A1", "category": "crisis", "tags": ["urgent"]},
+            {"prompt": "Q2", "response": "A2", "category": "humor", "source_issue": 42},
+        ]
+        result = normalize_training_examples(
+            examples, "b001", ["t1"], "fp", "fr"
+        )
+        assert result[0]["category"] == "crisis"
+        assert result[0]["tags"] == ["urgent"]
+        assert "source_issue" not in result[0]
+        assert result[1]["category"] == "humor"
+        assert result[1]["source_issue"] == 42
+        assert "tags" not in result[1]
+
+
 class TestNormalizeRubricScores:
     """normalize_rubric_scores() cleans eval rubric output."""