fix(ai): isolate heavy Ollama workloads from GCP alert lane

apps/api/src/services/intent_classifier.py

@@ -31,8 +31,8 @@ from typing import Protocol, runtime_checkable
 import httpx
 import structlog
 
-from src.core.config import settings
 from src.services.model_registry import get_model_registry
+from src.services.ollama_endpoint_resolver import resolve_ollama_endpoint
 
 logger = structlog.get_logger(__name__)
 
@@ -549,7 +549,7 @@ class IntentClassifier:
             # 呼叫 Ollama
             async with httpx.AsyncClient() as client:
                 response = await client.post(
-                    f"{settings.OLLAMA_URL}/api/generate",
+                    f"{resolve_ollama_endpoint('hermes')}/api/generate",
                     json={
                         "model": model_name,
                         "prompt": prompt,