feat(api): Phase 13.2 AI Rate Limiter + RAG 基礎設施 (#84)

Rate Limiter (防止 Gemini 用量暴衝):
- ai_rate_limiter.py: RPM/Daily/Token 三層閥值
- openclaw.py: 整合 rate limit 檢查，超限自動降級
- health.py: /health/ai-usage 監控端點

RAG Tool 基礎 (#84 進行中):
- embedding_service.py: Ollama embedding 封裝
- rag_service.py: Redis vector search 服務

閥值設定:
- Gemini: 10 RPM, 500/day, 100K tokens/day
- Claude: 5 RPM, 200/day, 50K tokens/day

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

apps/api/src/services/openclaw.py

@@ -35,6 +35,7 @@ from src.models.ai import (
     OpenClawDecision,
 )
 from src.services.langfuse_client import langfuse_trace
+from src.services.model_registry import get_model_registry
 from src.services.signoz_client import GoldMetrics, get_signoz_client
 from src.utils.k8s_naming import normalize_resource_name
 from src.utils.timezone import now_taipei_iso
@@ -270,17 +271,22 @@ class OpenClawService:
                 prompt_length=len(prompt),
             )
 
+            # 從 ModelRegistry 取得模型配置
+            registry = get_model_registry()
+            model_name = registry.get_model("ollama", "rca")
+            options = registry.get_provider_options("ollama")
+
             response = await client.post(
                 f"{settings.OLLAMA_URL}/api/generate",
                 json={
-                    "model": "qwen2.5:7b-instruct",  # 使用更大的模型提高品質
+                    "model": model_name,
                     "prompt": prompt,
                     "stream": False,
                     "format": "json",  # 強制 JSON 輸出
                     "options": {
-                        "num_predict": 1024,  # 增加輸出長度
-                        "temperature": 0.1,   # 低溫度確保穩定輸出
-                        "top_p": 0.9,
+                        "num_predict": options.get("num_predict", 1024),
+                        "temperature": options.get("temperature", 0.1),
+                        "top_p": options.get("top_p", 0.9),
                     },
                 },
                 timeout=httpx.Timeout(float(settings.OPENCLAW_TIMEOUT), connect=10.0),
@@ -324,9 +330,12 @@ class OpenClawService:
         try:
             client = await self._get_client()
 
-            # Gemini 1.5 Flash 支援 JSON Mode
+            # 從 ModelRegistry 取得模型配置
+            registry = get_model_registry()
+            model_name = registry.get_model("gemini", "rca")
+
             response = await client.post(
-                f"https://generativelanguage.googleapis.com/v1beta/models/gemini-1.5-flash:generateContent?key={settings.GEMINI_API_KEY}",
+                f"https://generativelanguage.googleapis.com/v1beta/models/{model_name}:generateContent?key={settings.GEMINI_API_KEY}",
                 json={
                     "contents": [{"parts": [{"text": prompt}]}],
                     "generationConfig": {
@@ -737,6 +746,24 @@ class OpenClawService:
 
         return response, provider, success, False  # from_cache=False
 
+    # =========================================================================
+    # Public LLM Interface (ILLMProvider Protocol)
+    # =========================================================================
+
+    async def call(self, prompt: str) -> tuple[str, str, bool]:
+        """
+        呼叫 LLM (ILLMProvider Protocol 實作)
+
+        #39 Error Analyzer Agent 使用此方法
+
+        Args:
+            prompt: 完整的 prompt
+
+        Returns:
+            (response, provider, success)
+        """
+        return await self._call_with_fallback(prompt)
+
     # =========================================================================
     # Fallback Chain
     # =========================================================================
@@ -784,7 +811,23 @@ class OpenClawService:
                         DeepLinking.langfuse_trace_url(trace.langfuse_trace_id),
                     )
 
+            # Phase 13.2: Rate Limiter 整合 (2026-03-26)
+            # 防止雲端 API 用量暴衝，超限自動降級
+            from src.services.ai_rate_limiter import get_ai_rate_limiter
+            rate_limiter = get_ai_rate_limiter()
+
             for provider in settings.AI_FALLBACK_ORDER:
+                # Rate Limit 檢查 (gemini/claude 需檢查，ollama 不限)
+                if provider in ("gemini", "claude"):
+                    allowed, reason = await rate_limiter.check_and_increment(provider)
+                    if not allowed:
+                        logger.warning(
+                            "ai_rate_limit_skip",
+                            provider=provider,
+                            reason=reason,
+                        )
+                        continue  # 跳過此 provider，嘗試下一個
+
                 logger.info("ai_provider_attempt", provider=provider)
 
                 start_time = time.time()
@@ -829,13 +872,9 @@ class OpenClawService:
             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)
+        """取得 provider 對應的模型名稱 (從 ModelRegistry)"""
+        registry = get_model_registry()
+        return registry.get_model(provider, "rca")
 
     # =========================================================================
     # Response Parsing (防禦性解析)