feat(api): LLM 整合 Expert System 診斷上下文

長期方案實作: Expert 診斷 + LLM 智能分析

變更:
1. decision_manager._dual_engine_analyze():
   - 測試資源跳過 LLM (省錢)
   - 傳遞 Expert 診斷上下文給 LLM
   - LLM 失敗時根據診斷調整回應

2. openclaw.generate_incident_proposal():
   - 新增 expert_context 參數
   - Prompt 包含 Expert 診斷結果
   - 引導 LLM 基於診斷做決策

流程:
Playbook → Expert診斷 → LLM(with context) → 智能建議

這是「先診斷根因，再決定行動」的正確實作

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

apps/api/src/services/decision_manager.py

@@ -585,32 +585,54 @@ class DecisionManager:
         incident: Incident,
     ) -> dict[str, Any]:
         """
-        三軌決策分析 (Phase 7.5 升級)
+        三軌決策分析 (Phase 7.5 升級 + 2026-03-27 智能診斷重構)
 
         策略:
         1. 先檢查 Playbook 是否有高度匹配 (similarity >= 85%)
         2. Playbook 命中則直接使用 (最快、經驗驗證)
-        3. 否則 LLM + Expert System 雙軌
+        3. Expert System 提供初步診斷 (分類 + 診斷指令)
+        4. LLM 基於診斷上下文提供智能建議
+        5. LLM 失敗時，根據 Expert 診斷決定是否需人工介入
 
-        優先順序: Playbook > LLM > Expert System
+        優先順序: Playbook > LLM(with Expert context) > Expert System
         """
         # Phase 7.5: 先嘗試 Playbook 匹配
         playbook_result = await self._try_playbook_match(incident)
         if playbook_result:
             return playbook_result
 
-        # Expert System 同步執行 (立即可用)
+        # ========== 2026-03-27 重構: 分層智能診斷 ==========
+
+        # Step 1: Expert System 提供初步診斷 (永不失敗)
         expert_result = expert_analyze(incident)
 
-        # LLM 非同步執行
-        try:
-            signals_dict = [s.model_dump() for s in incident.signals]
+        # Step 2: 測試資源直接返回 (不浪費 LLM 呼叫)
+        if expert_result.get("is_test_resource"):
+            logger.info(
+                "dual_engine_test_resource_skip",
+                incident_id=incident.incident_id,
+                target=incident.affected_services[0] if incident.affected_services else "unknown",
+            )
+            return expert_result
 
+        # Step 3: 準備 LLM 上下文 (含 Expert 診斷)
+        signals_dict = [s.model_dump() for s in incident.signals]
+        expert_context = {
+            "initial_diagnosis": expert_result.get("matched_rule"),
+            "diagnosis_description": expert_result.get("description"),
+            "suggested_diagnosis_commands": expert_result.get("diagnosis_commands", []),
+            "expert_confidence": expert_result.get("confidence"),
+            "requires_human_review": expert_result.get("human_review_required", False),
+        }
+
+        # Step 4: LLM 分析 (帶上 Expert 上下文)
+        try:
             llm_result, provider, success = await self._openclaw.generate_incident_proposal(
                 incident_id=incident.incident_id,
                 severity=incident.severity.value,
                 signals=signals_dict,
                 affected_services=incident.affected_services,
+                expert_context=expert_context,  # 傳遞 Expert 診斷上下文
             )
 
             if success and llm_result:
@@ -618,10 +640,12 @@ class DecisionManager:
                     "dual_engine_llm_win",
                     incident_id=incident.incident_id,
                     provider=provider,
+                    expert_rule=expert_result.get("matched_rule"),
                 )
                 return {
                     **llm_result,
                     "source": f"llm_{provider}",
+                    "expert_diagnosis": expert_result.get("matched_rule"),
                 }
 
         except Exception as e:
@@ -629,13 +653,23 @@ class DecisionManager:
                 "dual_engine_llm_failed",
                 incident_id=incident.incident_id,
                 error=str(e),
+                expert_rule=expert_result.get("matched_rule"),
             )
 
-        # LLM 失敗，使用 Expert System
+        # Step 5: LLM 失敗，使用 Expert System 結果
+        # 但根據診斷結果調整回應
         logger.info(
             "dual_engine_expert_fallback",
             incident_id=incident.incident_id,
+            expert_rule=expert_result.get("matched_rule"),
+            human_review=expert_result.get("human_review_required", False),
         )
+
+        # 如果 Expert 標記需人工介入，降低 confidence
+        if expert_result.get("human_review_required"):
+            expert_result["confidence"] = min(expert_result.get("confidence", 0.5), 0.5)
+            expert_result["description"] += " [LLM 分析失敗，建議人工確認]"
+
         return expert_result
 
     async def _try_playbook_match(