Seeing the same issue with WebLogic 12c on AIX 7.1, with IBM JDK SR3.
The same basic config works correctly on Linux and Windows.
Same as OP, we are using unicast cluster messaging with two managed servers on the same OS instance. We use Apache for the front-end proxy. However, the nodes are dropped and added from the cluster whenever they attempt to send a cluster message of any kind. Although they are quickly re-added, application errors result when performing any operation on a node while it is excluded from the cluster.
Updated cluster to use Multicast, but no difference in behavior. I can now run MulticastMonitor and see that the heartbeat messages are sent OK. And, the monitor sees the session update messages and thinks they are ok. However, any activity in the application causes the non-active peer to be removed with a "due to Peer Gone" message, then re-added, then updated.