Fix realtime queued task counts

apps/api/internal/httpapi/core_flow_integration_test.go

@@ -783,6 +783,24 @@ WHERE reference_type = 'gateway_task'
 	if len(asyncRateLimitDetail.Attempts) != 0 {
 		t.Fatalf("async rate-limited task should wait in queue without recording a failed attempt: %+v", asyncRateLimitDetail)
 	}
+	var modelRateLimits struct {
+		Items []struct {
+			ModelName   string  `json:"modelName"`
+			ModelAlias  string  `json:"modelAlias"`
+			QueuedTasks float64 `json:"queuedTasks"`
+		} `json:"items"`
+	}
+	doJSON(t, server.URL, http.MethodGet, "/api/admin/runtime/model-rate-limits", loginResponse.AccessToken, nil, http.StatusOK, &modelRateLimits)
+	var queuedTasks float64
+	for _, item := range modelRateLimits.Items {
+		if item.ModelName == rateLimitedModel || item.ModelAlias == rateLimitedModel {
+			queuedTasks = item.QueuedTasks
+			break
+		}
+	}
+	if queuedTasks < 1 {
+		t.Fatalf("realtime load should count async rate-limited task as queued, got %v in %+v", queuedTasks, modelRateLimits.Items)
+	}
 	asyncRateLimitCompleted := waitForTaskStatus(t, server.URL, apiKeyResponse.Secret, asyncRateLimitTask.TaskID, []string{"succeeded"}, time.Duration(rateLimitWindowSeconds+3)*time.Second)
 	if asyncRateLimitCompleted.Status != "succeeded" {
 		t.Fatalf("async rate-limited task should be pulled from queue after the limit window resets, got %+v", asyncRateLimitCompleted)