fix: calculate total from criteria scores instead of trusting LLM

LLM sometimes returns inconsistent total values. Now we always
compute total = sum(criteria.score) for accuracy.

.autograde/llm_grade.py

@@ -169,6 +169,12 @@ def main():
                 "confidence": 0.0
             }
     
+    # 重新计算 total（不信任 LLM 返回的 total，使用各项得分之和）
+    criteria = resp.get("criteria", [])
+    if criteria:
+        calculated_total = sum(float(c.get("score", 0)) for c in criteria)
+        resp["total"] = round(calculated_total, 2)
+    
     # 边界带自动送审
     try:
         rubric_data = json.loads(rubric_text)