V10.576 修正 GCP-only Ollama retry

services/ollama_service.py

@@ -638,9 +638,18 @@ class OllamaService:
         for attempt in range(max_attempts):
             current_host = _normalize_host(self.host)  # property 每次 lazy resolve
             if not allow_111_fallback and _is_111_fallback_host(current_host):
-                last_error = "111 fallback disabled; no approved GCP Ollama host available"
-                logger.warning("[Ollama] %s", last_error)
-                break
+                next_host = next((host for host in allowed_hosts if host not in attempted_hosts), None)
+                if not next_host:
+                    last_error = "111 fallback disabled; no approved GCP Ollama host available"
+                    logger.warning("[Ollama] %s", last_error)
+                    break
+                logger.info(
+                    "[Ollama] resolver selected 111 but caller disabled fallback; forcing approved GCP host=%s",
+                    next_host,
+                )
+                if self._explicit_host is None:
+                    _clear_resolved_host_cache()
+                current_host = next_host
             if current_host in attempted_hosts:
                 # 已試過同主機時，若是標準三主機鏈且 caller 沒指定 host，
                 # 改走尚未嘗試的下一台。避免 request timeout(60s) 大於