fix(api): route direct ollama callers through ordered fallback

apps/api/src/services/intent_classifier.py

@@ -32,7 +32,7 @@ import httpx
 import structlog
 
 from src.services.model_registry import get_model_registry
-from src.services.ollama_endpoint_resolver import resolve_ollama_endpoint
+from src.services.ollama_endpoint_resolver import resolve_ollama_order
 
 logger = structlog.get_logger(__name__)
 
@@ -546,27 +546,56 @@ class IntentClassifier:
             # 取得模型配置
             model_name = self.llm_model  # qwen2.5:1b 或配置值
 
-            # 呼叫 Ollama
+            # 呼叫 Ollama（GCP-A → GCP-B → 111）
+            result_text = ""
+            last_error: Exception | None = None
             async with httpx.AsyncClient() as client:
-                response = await client.post(
-                    f"{resolve_ollama_endpoint('hermes')}/api/generate",
-                    json={
-                        "model": model_name,
-                        "prompt": prompt,
-                        "stream": False,
-                        "format": "json",
-                        "options": {
-                            "num_predict": 128,  # 意圖分類只需短回應
-                            "temperature": 0.0,  # 確定性輸出
-                            "top_p": 0.9,
-                        },
-                    },
-                    timeout=httpx.Timeout(5.0, connect=2.0),  # 嚴格超時
-                )
+                for endpoint in resolve_ollama_order("hermes"):
+                    if not endpoint.url:
+                        continue
+                    try:
+                        response = await client.post(
+                            f"{endpoint.url}/api/generate",
+                            json={
+                                "model": model_name,
+                                "prompt": prompt,
+                                "stream": False,
+                                "format": "json",
+                                "options": {
+                                    "num_predict": 128,  # 意圖分類只需短回應
+                                    "temperature": 0.0,  # 確定性輸出
+                                    "top_p": 0.9,
+                                },
+                            },
+                            timeout=httpx.Timeout(5.0, connect=2.0),  # 嚴格超時
+                        )
 
-                response.raise_for_status()
-                data = response.json()
-                result_text = data.get("response", "")
+                        response.raise_for_status()
+                        data = response.json()
+                        result_text = data.get("response", "")
+                        break
+                    except Exception as e:
+                        last_error = e
+                        logger.warning(
+                            "intent_llm_endpoint_failed",
+                            provider=endpoint.provider_name,
+                            error=str(e),
+                            error_type=type(e).__name__,
+                        )
+
+            if not result_text:
+                _kw_result = self._keyword_classify(text)
+                return IntentResult(
+                    intent=_kw_result.intent,
+                    confidence=_kw_result.confidence,
+                    method="llm_unavailable_keyword",
+                    matched_keywords=_kw_result.matched_keywords,
+                    detected_resources=_kw_result.detected_resources,
+                    reasoning=(
+                        f"LLM 全端點不可用（{type(last_error).__name__ if last_error else 'no_endpoint'}）"
+                        f" → {_kw_result.reasoning}"
+                    ),
+                )
 
             # 解析 JSON 回應
             elapsed_ms = (time.time() - start_time) * 1000