perf: 在 Redis 密集操作前释放 DB 连接，candidate_records 改为异步写入

- 号池排序涉及大量 Redis I/O，在调用前提前释放 DB 连接避免连接池压力
- 新增 create_candidate_records_async，通过 asyncio.to_thread 执行同步 DB 写入
- 同步执行、异步提交、TaskService 三条路径统一改用异步版本

src/services/orchestration/candidate_resolver.py

@@ -4,6 +4,7 @@
 负责获取和排序可用的 Provider/Endpoint/Key 组合
 """
 
+import asyncio
 import uuid
 from datetime import datetime, timezone
 from typing import Any
@@ -378,6 +379,28 @@ class CandidateResolver:
 
         return candidate_record_map
 
+    async def create_candidate_records_async(
+        self,
+        all_candidates: list[ProviderCandidate],
+        request_id: str | None,
+        user_id: str | None,
+        user_api_key: ApiKey | None,
+        required_capabilities: dict[str, bool] | None = None,
+        *,
+        expand_retries: bool = True,
+    ) -> dict[tuple[int, int], str]:
+        """异步版本的 create_candidate_records，将同步 DB 操作放到线程池执行，
+        避免阻塞 asyncio 事件循环。"""
+        return await asyncio.to_thread(
+            self.create_candidate_records,
+            all_candidates,
+            request_id,
+            user_id,
+            user_api_key,
+            required_capabilities,
+            expand_retries=expand_retries,
+        )
+
     def get_active_candidates(
         self,
         all_candidates: list[ProviderCandidate],

src/services/scheduling/aware_scheduler.py

@@ -568,6 +568,9 @@ class CacheAwareScheduler:
             candidates, db, affinity_key, api_format
         )
 
+        # 排序完成后释放 DB 连接，避免后续 Redis 操作期间占用连接
+        release_db_connection_before_await(db)
+
         # 2. 调度模式排序
         if self.scheduling_mode == self.SCHEDULING_MODE_CACHE_AFFINITY:
             if affinity_key and candidates and global_model_id:

src/services/scheduling/candidate_builder.py

@@ -565,6 +565,10 @@ class CandidateBuilder:
                 if pool_cfg is not None:
                     # 号池优化：跳过逐 key 的 _check_key_availability 检查，
                     # 直接收集全部 active key，将检查推迟到 PoolManager 排序后分页执行。
+                    # 在此释放 DB 连接，因为后续的 PoolManager 排序涉及大量 Redis 操作，
+                    # 避免在 Redis I/O 期间长时间占用 DB 连接池。
+                    release_db_connection_before_await(db)
+
                     pool_keys = list(keys_to_check)
 
                     if not pool_keys:

src/services/task/execute/sync_execute.py

@@ -161,12 +161,17 @@ class SyncTaskExecutionService:
             request_body=request_body,
         )
 
+        # 号池排序涉及大量 Redis 操作，提前释放 DB 连接避免连接池压力
+        from src.services.scheduling.utils import release_db_connection_before_await
+
+        release_db_connection_before_await(self.db)
+
         # Account Pool: reorder candidates for claude_code providers.
         all_candidates, pool_traces = await self._pool_ops.apply_pool_reorder(
             all_candidates, request_body=request_body
         )
 
-        candidate_record_map = candidate_resolver.create_candidate_records(
+        candidate_record_map = await candidate_resolver.create_candidate_records_async(
             all_candidates=all_candidates,
             request_id=request_id,
             user_id=user_id,

src/services/task/service.py

@@ -390,11 +390,17 @@ class TaskService:
                 _logger.warning("创建 pending 使用记录失败: {}", str(exc))
 
         all_candidates = list(candidates)
+
+        # 号池排序涉及大量 Redis 操作，提前释放 DB 连接避免连接池压力
+        from src.services.scheduling.utils import release_db_connection_before_await
+
+        release_db_connection_before_await(self.db)
+
         all_candidates, pool_traces = await pool_ops.apply_pool_reorder(
             all_candidates, request_body=request_body
         )
 
-        candidate_record_map = candidate_resolver.create_candidate_records(
+        candidate_record_map = await candidate_resolver.create_candidate_records_async(
             all_candidates=all_candidates,
             request_id=request_id,
             user_id=user_id,

src/services/task/submit/prepare.py

@@ -87,6 +87,11 @@ class AsyncSubmitPreparationService:
                 last_status_code=None,
             )
 
+        # 号池排序涉及大量 Redis 操作，提前释放 DB 连接避免连接池压力
+        from src.services.scheduling.utils import release_db_connection_before_await
+
+        release_db_connection_before_await(self.db)
+
         # Account Pool: keep internal key failover order/skip behavior
         # consistent with the SYNC path.
         candidates, _pool_traces = await apply_pool_reorder(
@@ -103,7 +108,7 @@ class AsyncSubmitPreparationService:
         record_map: dict[tuple[int, int], str] = {}
         if request_id:
             try:
-                record_map = resolver.create_candidate_records(
+                record_map = await resolver.create_candidate_records_async(
                     all_candidates=candidates,
                     request_id=request_id,
                     user_id=str(user_api_key.user_id),