diff --git a/pkg/executor/util/deployment.go b/pkg/executor/util/deployment.go
index 49cb4f3..7ca6130 100644
--- a/pkg/executor/util/deployment.go
+++ b/pkg/executor/util/deployment.go
@@ -62,9 +62,25 @@ func nextWaitPollDelay(cur time.Duration) time.Duration {
 	return min(cur*3/2, waitPollMaxDelay)
 }
 
-// WaitForDeployment polls the deployment until it has at least replicas
-// AvailableReplicas, or until specializationTimeout seconds elapse. Shared by
-// the newdeploy and container managers. specializationTimeout is floored at
+// deploymentAvailable reports whether depl has at least replicas available,
+// as observed by the deployment controller for the spec currently stored.
+//
+// The generation gate matters on the scale-from-zero path. UpdateScale only
+// writes the spec and bumps metadata.generation; status is written later, by
+// the deployment controller. A Get issued right after ScaleDeployment can
+// therefore still carry the status of the previous generation, which counts
+// the pod the idle reaper is tearing down, so AvailableReplicas alone would
+// report a serving deployment while the Service has no endpoint. Until
+// ObservedGeneration catches up, the status says nothing about this scale
+// request. Same gate as WaitForDeploymentRollout in the integration framework.
+func deploymentAvailable(depl *appsv1.Deployment, replicas int32) bool {
+	return depl.Status.ObservedGeneration >= depl.Generation && depl.Status.AvailableReplicas >= replicas
+}
+
+// WaitForDeployment polls the deployment until the deployment controller has
+// observed the current spec and reports at least replicas AvailableReplicas,
+// or until specializationTimeout seconds elapse. Shared by the newdeploy and
+// container managers. specializationTimeout is floored at
 // fv1.DefaultSpecializationTimeOut. AvailableReplicas is used in preference to
 // ReadyReplicas since the pods may not be able to serve network traffic yet.
 func WaitForDeployment(ctx context.Context, kubeClient kubernetes.Interface, logger logr.Logger, depl *appsv1.Deployment, replicas int32, specializationTimeout int) (latestDepl *appsv1.Deployment, err error) {
@@ -92,7 +108,7 @@ func WaitForDeployment(ctx context.Context, kubeClient kubernetes.Interface, log
 			return nil, err
 		}
 		// TODO check for imagePullerror
-		if latestDepl.Status.AvailableReplicas >= replicas {
+		if deploymentAvailable(latestDepl, replicas) {
 			otelUtils.SpanTrackEvent(ctx, "deploymentAvailable", otelUtils.GetAttributesForDeployment(latestDepl)...)
 			return latestDepl, err
 		}
diff --git a/pkg/executor/util/deployment_test.go b/pkg/executor/util/deployment_test.go
index ccd4f8c..028a3a1 100644
--- a/pkg/executor/util/deployment_test.go
+++ b/pkg/executor/util/deployment_test.go
@@ -194,6 +194,35 @@ func TestWaitForDeployment(t *testing.T) {
 		assert.GreaterOrEqual(t, got.Status.AvailableReplicas, int32(1))
 	})
 
+	t.Run("ignores a status the controller has not caught up with", func(t *testing.T) {
+		t.Parallel()
+		// Scaled in from one replica to zero and back up: the spec is at
+		// generation 2, while the status still describes generation 1 and
+		// counts the pod the idle reaper is tearing down.
+		one := int32(1)
+		scaled := &appsv1.Deployment{
+			Namespace: ns, Name: "scaled-up", Generation: 2,
+			Spec:   appsv1.DeploymentSpec{Replicas: &one},
+			Status: appsv1.DeploymentStatus{ObservedGeneration: 1, AvailableReplicas: 1},
+		}
+		client := fake.NewClientset(scaled)
+		var gets int
+		client.PrependReactor("get", "deployments", func(k8stesting.Action) (bool, runtime.Object, error) {
+			latest := scaled.DeepCopy()
+			gets++
+			if gets > 1 {
+				latest.Status.ObservedGeneration = latest.Generation
+			}
+			return true, latest, nil
+		})
+
+		got, err := WaitForDeployment(t.Context(), client, loggerfactory.GetLogger(), scaled, 1, 1)
+		require.NoError(t, err)
+		require.NotNil(t, got)
+		assert.Equal(t, got.Generation, got.Status.ObservedGeneration,
+			"returned on a status that predates the scale request")
+	})
+
 	t.Run("surfaces a get error", func(t *testing.T) {
 		t.Parallel()
 		client := fake.NewClientset()