feat(knowledge): KB Phase 2 — OpenClaw RAG 整合

_dual_engine_analyze 新增 _query_kb_context():
- Incident 分析前語意搜尋相關 KB 條目 (top-3, threshold=0.4)
- 將 KB context 注入 expert_context.diagnosis_context 傳給 LLM
- 失敗時靜默降級，不影響主分析流程
- dual_engine_llm_win log 新增 kb_rag 欄位，可觀測 RAG 命中率

架構: _query_kb_context 透過 get_knowledge_service() 呼叫 Service 層
符合 leWOOOgo 積木化 — decision_manager 不直接存取 DB/pgvector

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

apps/api/src/services/decision_manager.py

@@ -548,17 +548,56 @@ class DecisionManager:
                 _push_decision_to_telegram(incident, token.proposal_data)
             )
 
+    async def _query_kb_context(self, incident: Incident) -> str:
+        """
+        KB Phase 2: 語意搜尋相關 KB 條目，組裝為 LLM context 字串
+        2026-04-04 Claude Code: KB RAG 整合
+
+        失敗時靜默降級，不影響主分析流程
+        """
+        try:
+            from src.services.knowledge_service import get_knowledge_service
+            query_parts = list(incident.affected_services)
+            if incident.signals:
+                query_parts.insert(0, getattr(incident.signals[0], "alert_name", ""))
+            query = " ".join(filter(None, query_parts))
+
+            svc = get_knowledge_service()
+            results = await svc.semantic_search(query, limit=3, threshold=0.4)
+            if not results:
+                return ""
+
+            lines = ["## 📚 Knowledge Base 相關條目 (KB RAG)"]
+            for entry, score in results:
+                lines.append(
+                    f"\n### [{entry.entry_type}] {entry.title} (similarity={score:.2f})"
+                )
+                lines.append(entry.content[:500])
+                if len(entry.content) > 500:
+                    lines.append("... (truncated)")
+
+            logger.info(
+                "kb_rag_context_injected",
+                incident_id=incident.incident_id,
+                kb_hits=len(results),
+            )
+            return "\n".join(lines)
+
+        except Exception as e:
+            logger.warning("kb_rag_failed", incident_id=incident.incident_id, error=str(e))
+            return ""
+
     async def _dual_engine_analyze(
         self,
         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 + KB Phase 2 RAG 整合)
 
         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. 否則 LLM + Expert System 雙軌 + KB RAG context 注入
 
         優先順序: Playbook > LLM > Expert System
         """
@@ -570,16 +609,29 @@ class DecisionManager:
         # Expert System 同步執行 (立即可用)
         expert_result = expert_analyze(incident)
 
+        # KB Phase 2: 語意搜尋相關知識條目 (失敗時靜默降級)
+        # 2026-04-04 Claude Code: KB RAG 整合，提升 LLM 決策品質
+        kb_context = await self._query_kb_context(incident)
+
         # LLM 非同步執行 (Phase 22: OpenClaw + Nemotron 協作)
         # 2026-03-31 Claude Code: 使用 _with_tools 方法啟用雙軌協作
         try:
             signals_dict = [s.model_dump() for s in incident.signals]
 
+            # 將 KB context 注入 expert_context 傳給 LLM
+            llm_expert_context: dict[str, Any] = {**expert_result} if expert_result else {}
+            if kb_context:
+                existing = str(llm_expert_context.get("diagnosis_context", ""))
+                llm_expert_context["diagnosis_context"] = (
+                    f"{kb_context}\n\n{existing}" if existing else kb_context
+                )
+
             llm_result, provider, success = await self._openclaw.generate_incident_proposal_with_tools(
                 incident_id=incident.incident_id,
                 severity=incident.severity.value,
                 signals=signals_dict,
                 affected_services=incident.affected_services,
+                expert_context=llm_expert_context if llm_expert_context else None,
             )
 
             if success and llm_result:
@@ -587,6 +639,7 @@ class DecisionManager:
                     "dual_engine_llm_win",
                     incident_id=incident.incident_id,
                     provider=provider,
+                    kb_rag=bool(kb_context),
                 )
                 return {
                     **llm_result,