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

apps/api/src/services/knowledge_extractor_service.py

@@ -15,10 +15,11 @@ import structlog
 
 logger = structlog.get_logger(__name__)
 
-# 2026-05-03 ogt: ADR-110 GCP-A Primary — 改從 settings 讀取，不再硬編碼 111
+# 2026-05-05 Codex: KB 萃取走 111 lane，避免污染 GCP alert-fast lane
 def _get_ollama_base() -> str:
-    from src.core.config import get_settings
-    return get_settings().OLLAMA_URL
+    from src.services.ollama_endpoint_resolver import resolve_ollama_endpoint
+
+    return resolve_ollama_endpoint("deep_rca")
 _EXTRACT_MODEL = "llama3.2:3b"
 _EXTRACT_TIMEOUT = 30.0  # 秒，容忍慢速
 
@@ -117,7 +118,11 @@ class KnowledgeExtractorService:
             category = self._infer_category(incident)
 
             # 5. 建立 KB 條目
-            from src.models.knowledge import EntrySource, EntryType, KnowledgeEntryCreate
+            from src.models.knowledge import (
+                EntrySource,
+                EntryType,
+                KnowledgeEntryCreate,
+            )
             from src.services.knowledge_service import get_knowledge_service
 
             entry_data = KnowledgeEntryCreate(