Fix case loading terminology - reference_sql is human-provided ground truth

src/evaluator.py

@@ -109,7 +109,7 @@ class DatasetManager:
             case = CaseConfig(
                 id=case_data['id'],
                 question=case_data['question'],
-                reference_sql=case_data['reference_sql'],
+                reference_sql=case_data['reference_sql'],  # This is already a dict with dialects
                 difficulty=case_data.get('difficulty', 'medium'),
                 description=case_data.get('description', '')
             )