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

apps/api/src/services/embedding_service.py

@@ -21,8 +21,8 @@ from typing import Protocol
 import httpx
 import structlog
 
-from src.core.config import settings
 from src.services.model_registry import get_model as _get_model
+from src.services.ollama_endpoint_resolver import resolve_ollama_endpoint
 
 logger = structlog.get_logger(__name__)
 
@@ -98,7 +98,7 @@ class OllamaEmbeddingService:
         P1 修復 (2026-03-29): 維度配置化，支援更多模型
         """
         self._model = model
-        self._ollama_url = ollama_url or settings.OLLAMA_URL
+        self._ollama_url = ollama_url or resolve_ollama_endpoint("embedding")
         self._timeout = timeout
         self._default_dimension = default_dimension or self.MODEL_DIMENSIONS.get(
             model, self.DEFAULT_DIMENSION