perf(worker): 持续批量填满异步执行槽

异步准入通知在突发提交时会合并，原调度器每次只填充一个 8 任务批次，导致全局容量长期空闲并反向拖慢入口登记。\n\n收到通知或周期唤醒后连续拉取并提交批次，直到队列为空或全局执行容量真正饱和；保留每批 8 个的短事务边界。\n\n验证：env -u AI_GATEWAY_TEST_DATABASE_URL go test ./... -count=1；go vet ./...；gofmt -l 无输出。

apps/api/internal/runner/admission.go

@@ -704,6 +704,7 @@ func (s *Service) dispatchWaitingAsyncAdmissions(ctx context.Context) {
 		case <-s.asyncAdmissionWake:
 		case <-ticker.C:
 		}
+		for {
 			batchLimit := s.cfg.AsyncWorkerHardLimit
 			if batchLimit <= 0 {
 				batchLimit = 64
@@ -714,7 +715,10 @@ func (s *Service) dispatchWaitingAsyncAdmissions(ctx context.Context) {
 			taskIDs, err := s.store.ListWaitingAsyncAdmissionTaskIDs(ctx, batchLimit)
 			if err != nil {
 				s.logger.Warn("list waiting async admissions failed", "error", err)
-			continue
+				break
+			}
+			if len(taskIDs) == 0 {
+				break
 			}
 			tasks := make([]store.GatewayTask, 0, len(taskIDs))
 			for _, taskID := range taskIDs {
@@ -730,16 +734,25 @@ func (s *Service) dispatchWaitingAsyncAdmissions(ctx context.Context) {
 				}
 				tasks = append(tasks, task)
 			}
+			if len(tasks) == 0 {
+				break
+			}
 			saturated, err := s.dispatchWaitingAsyncTasks(ctx, tasks)
 			if err != nil {
 				s.logger.Warn("dispatch waiting async admission batch failed", "tasks", len(tasks), "error", err)
-			continue
+				break
 			}
 			if saturated {
 				// Every asynchronous task shares the global worker-capacity FIFO
 				// scope. Once its current head cannot be admitted, later tasks
 				// cannot make progress until a lease is released.
 				s.observeTaskAdmission("dispatch_saturated")
+				break
+			}
+			// PostgreSQL notifications are wake-up hints and collapse while a
+			// batch is being admitted. Keep filling consecutive batches until
+			// the global execution scope is actually saturated so a large burst
+			// cannot strand most Worker slots behind the one-second fallback.
 		}
 	}
 }