fix(llm_score): 修复多轮用例 question 提取错位导致普遍打 0 分

## 现象
多轮/动态用例的 llm_score 规则几乎全部返回 0.0/10 失败，即便被测对象
回复质量很高（实测 400+ 字的专业医疗回复也是 0 分）。

## 根因
evaluate() 里 question 提取逻辑错误：
    if len(dialog) >= 2:
        question_text = extract_reply_text(dialog[-2].reply)  # BUG
dialog[-2].reply 是「上一轮智能体的回复」，被误当成「用户问题」。于是评分
LLM 收到的问答对是：
  - "用户问题" = 上一轮 AI 回复
  - "智能体回复" = 当前轮 AI 回复
两段都是 AI 说的话、互不相关，评分 LLM 判定牛头不对马嘴 → 打 0 分。

单轮用例因 len(dialog)<2 走 sent_message 提取（正确），故不受影响；
问题只在多轮/dynamic 用例爆发。

## 修复
- question 始终取当前轮 last_turn.sent_message（用户实际发送的问题），
  移除错误的 dialog[-2].reply 分支
- reason 增加评分 LLM 自己的理由（parsed["reason"]），便于未来诊断
- 新增回归测试：多轮场景验证 question 来自 last sent_message 而非 prev reply

Co-Authored-By: Claude <noreply@anthropic.com>

backend/agenteval/evaluation/rules/llm_score.py

@@ -22,10 +22,11 @@ class LlmScoreRule(EvalRule):
         last_turn = dialog[-1]
         reply_text = extract_reply_text(last_turn.reply)
 
+        # 用户问题取自当前轮发送的消息（sent_message），而非上一轮的智能体回复。
+        # 旧逻辑用 dialog[-2].reply 会把「上一轮 AI 回复」误当成「用户问题」，
+        # 导致多轮/动态用例里评分 LLM 收到牛头不对马嘴的问答对，普遍打 0 分。
         question_text = ""
-        if len(dialog) >= 2:
-            question_text = extract_reply_text(dialog[-2].reply) or ""
-        if not question_text and last_turn.sent_message:
+        if last_turn.sent_message:
             body = last_turn.sent_message.get("msgBody", "")
             if isinstance(body, dict):
                 question_text = body.get("content", "")
@@ -49,10 +50,12 @@ class LlmScoreRule(EvalRule):
             return RuleResult(passed=False, reason=f"LLM 评分失败: {reason}")
 
         passed = score >= min_score
+        verdict = "通过" if passed else "未通过"
+        detail = f"；{reason}" if reason else ""
         return RuleResult(
             passed=passed,
             score=score / 10.0,
-            reason=f"LLM 评分 {score}/10，{'通过' if passed else '未通过'} (阈值 {min_score})",
+            reason=f"LLM 评分 {score}/10，{verdict} (阈值 {min_score}){detail}",
         )
 
     async def _call_llm(

tests/unit/test_llm_score.py

@@ -189,3 +189,50 @@ async def test_llm_score_extracts_question_from_sent_message():
     user_msg = next((m for m in messages if m["role"] == "user"), None)
     assert user_msg is not None
     assert "这是用户的问题" in user_msg["content"]
+
+
+async def test_llm_score_multiturn_uses_last_sent_not_prev_reply():
+    """Regression: for multi-turn dialogs the question must come from the LAST
+    turn's sent_message (the user's question), NOT dialog[-2].reply (the
+    previous AI reply). The old bug fed two AI replies as a Q&A pair, so the
+    judge LLM scored everything 0."""
+    rule = LlmScoreRule({"api_url": "http://mock/v1/chat", "min_score": 5})
+    captured_payload = {}
+
+    async def capture_post(url, **kwargs):
+        captured_payload.update(kwargs.get("json", {}))
+        return _make_llm_response(score=9.0)
+
+    dialog = [
+        _turn("第一轮AI回复", sent_text="第一轮用户问题"),
+        _turn("第二轮AI回复", sent_text="第二轮用户问题"),
+        _turn("第三轮AI回复", sent_text="第三轮用户问题"),
+    ]
+
+    with patch("agenteval.evaluation.rules.llm_score.httpx.AsyncClient") as MockClient:
+        instance = MockClient.return_value.__aenter__.return_value
+        instance.post = AsyncMock(side_effect=capture_post)
+        await rule.evaluate(_case(), dialog)
+
+    user_msg = next(m for m in captured_payload["messages"] if m["role"] == "user")
+    # 问题必须是最后一轮用户发送的问题
+    assert "第三轮用户问题" in user_msg["content"]
+    # 回复必须是最后一轮 AI 回复
+    assert "第三轮AI回复" in user_msg["content"]
+    # 绝不能把上一轮 AI 回复当成"用户问题"
+    assert "第二轮AI回复" not in user_msg["content"]
+
+
+async def test_llm_score_reason_includes_llm_detail():
+    """The rule's reason should surface the judge LLM's own reason for diagnosis."""
+    rule = LlmScoreRule({"api_url": "http://mock/v1/chat", "min_score": 5})
+    mock_resp = _make_llm_response(score=3.0, reason="回复偏离主题")
+
+    with patch("agenteval.evaluation.rules.llm_score.httpx.AsyncClient") as MockClient:
+        instance = MockClient.return_value.__aenter__.return_value
+        instance.post = AsyncMock(return_value=mock_resp)
+        result = await rule.evaluate(_case(), [_turn("answer")])
+
+    assert result.passed is False
+    assert "回复偏离主题" in result.reason
+