feat(api): Phase 15.1 Langfuse LLMOps 整合 + 模型升級

## 新功能
- Langfuse 自建部署 (192.168.0.110:3100)
- langfuse_client.py - LLM 呼叫追蹤包裝
- OpenClaw 整合 Langfuse trace

## 模型升級 (統帥批准)
- 生產預設: llama3.2:3b → qwen2.5:7b-instruct
- 摘要任務: llama3.2:3b (速度優先)

## 配置更新
- requirements.txt: +langfuse>=2.0.0
- config.py: +LANGFUSE_* 設定
- models.json: 更新 Ollama 模型配置
- K8s: Secret + ConfigMap 更新

## 審查通過
- 模組化檢查 ✅
- 核心測試 31/31 ✅

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

apps/api/src/services/openclaw.py

@@ -33,6 +33,7 @@ from src.core.redis_client import get_redis
 from src.models.ai import (
     OpenClawDecision,
 )
+from src.services.langfuse_client import langfuse_trace
 from src.services.signoz_client import GoldMetrics, get_signoz_client
 from src.utils.timezone import now_taipei_iso
 
@@ -360,7 +361,7 @@ class OpenClawService:
             response = await client.post(
                 f"{settings.OLLAMA_URL}/api/generate",
                 json={
-                    "model": "llama3.2:3b",  # 使用更大的模型提高品質
+                    "model": "qwen2.5:7b-instruct",  # 使用更大的模型提高品質
                     "prompt": prompt,
                     "stream": False,
                     "format": "json",  # 強制 JSON 輸出
@@ -823,34 +824,75 @@ class OpenClawService:
 
         若 MOCK_MODE=True，直接回傳模擬結果。
         若所有 Provider 失敗，fallback 到 Mock。
+
+        Phase 15.1: 整合 Langfuse LLMOps 追蹤
         """
         # Mock Mode: 開發測試用
         if settings.MOCK_MODE:
             logger.info("mock_mode_enabled", using="mock_llm")
             return self._generate_mock_response(alert_context or {}, signoz_metrics), "mock", True
 
-        for provider in settings.AI_FALLBACK_ORDER:
-            logger.info("ai_provider_attempt", provider=provider)
+        # Phase 15.1: Langfuse 追蹤整合
+        with langfuse_trace(
+            "openclaw_fallback_chain",
+            metadata={
+                "prompt_length": len(prompt),
+                "fallback_order": settings.AI_FALLBACK_ORDER,
+                "alert_fingerprint": (alert_context or {}).get("fingerprint", "unknown"),
+            },
+        ) as trace:
+            for provider in settings.AI_FALLBACK_ORDER:
+                logger.info("ai_provider_attempt", provider=provider)
 
-            if provider == "ollama":
-                response, success = await self._call_ollama(prompt)
-            elif provider == "gemini":
-                response, success = await self._call_gemini(prompt)
-            elif provider == "claude":
-                response, success = await self._call_claude(prompt)
-            else:
-                logger.warning("unknown_ai_provider", provider=provider)
-                continue
+                start_time = time.time()
+                model_name = self._get_model_name(provider)
 
-            if success:
-                logger.info("ai_provider_success", provider=provider)
-                return response, provider, True
+                if provider == "ollama":
+                    response, success = await self._call_ollama(prompt)
+                elif provider == "gemini":
+                    response, success = await self._call_gemini(prompt)
+                elif provider == "claude":
+                    response, success = await self._call_claude(prompt)
+                else:
+                    logger.warning("unknown_ai_provider", provider=provider)
+                    continue
 
-            logger.warning("ai_provider_failed_fallback", provider=provider)
+                latency_ms = (time.time() - start_time) * 1000
 
-        # 所有 Provider 失敗時，fallback 到 Mock (優雅降級)
-        logger.warning("all_providers_failed_using_mock", fallback="mock_llm")
-        return self._generate_mock_response(alert_context or {}, signoz_metrics), "mock_fallback", True
+                # Langfuse: 記錄每次 LLM 呼叫
+                trace.generation(
+                    name=f"{provider}_call",
+                    model=model_name,
+                    input=prompt[:500],  # 截斷避免過長
+                    output=response[:500] if success else f"ERROR: {response[:200]}",
+                    metadata={
+                        "success": success,
+                        "latency_ms": round(latency_ms, 2),
+                        "provider": provider,
+                    },
+                )
+
+                if success:
+                    logger.info("ai_provider_success", provider=provider, latency_ms=latency_ms)
+                    # Langfuse: 記錄成功評分
+                    trace.score(name="provider_success", value=1.0, comment=f"Success via {provider}")
+                    return response, provider, True
+
+                logger.warning("ai_provider_failed_fallback", provider=provider, latency_ms=latency_ms)
+
+            # 所有 Provider 失敗時，fallback 到 Mock (優雅降級)
+            logger.warning("all_providers_failed_using_mock", fallback="mock_llm")
+            trace.score(name="provider_success", value=0.0, comment="All providers failed, using mock")
+            return self._generate_mock_response(alert_context or {}, signoz_metrics), "mock_fallback", True
+
+    def _get_model_name(self, provider: str) -> str:
+        """取得 provider 對應的模型名稱"""
+        model_map = {
+            "ollama": "qwen2.5:7b-instruct",
+            "gemini": "gemini-1.5-flash",
+            "claude": "claude-3-haiku-20240307",
+        }
+        return model_map.get(provider, provider)
 
     # =========================================================================
     # Response Parsing (防禦性解析)