fix: use TestProbe instead of TestActor in ClusterShardingDistributedDataSpecs to avoid stale dead ref on Windows (#747)

On Windows CI, TestActor can survive EnsureTestActorAliveAsync's liveness
check but then die in the narrow window before the test body runs, leaving
TestActor as a dead ActorRef. All Identify replies then land in dead letters
and ExpectMsgAsync times out every retry. A TestProbe created in the test
body is unaffected by the startup race since it is created post-startup.

Author: Aaronontheweb

src/Akka.Cluster.Hosting.Tests/ClusterShardingDistributedDataSpecs.cs

@@ -48,11 +48,14 @@ await AwaitAssertAsync(() =>
         var coordinatorName = settings.RestartReplicatorOnFailure ? $"{ReplicatorName}Supervisor" : ReplicatorName;
 
         var actorSelection = Sys.ActorSelection(new RootActorPath(cluster.SelfAddress) / "user" / coordinatorName);
-        
+
+        // Use a fresh TestProbe rather than TestActor: on Windows, TestActor can be a stale dead
+        // reference due to the startup race window between EnsureTestActorAliveAsync and the test body.
+        var probe = CreateTestProbe();
         await AwaitAssertAsync(async () =>
         {
-            actorSelection.Tell(new Identify("coordinator"), TestActor);
-            var identity = await ExpectMsgAsync<ActorIdentity>(TimeSpan.FromSeconds(1));
+            actorSelection.Tell(new Identify("coordinator"), probe.Ref);
+            var identity = await probe.ExpectMsgAsync<ActorIdentity>(TimeSpan.FromSeconds(1));
             Assert.NotNull(identity.Subject);
         }, duration: TimeSpan.FromSeconds(10));
     }