In your example, what I found strange is that when C1 needs to access some bean on C3, you need to understand that they have *nothing* in common: they are not the same node (trivial ) *and* they are not part of the same cluster.
Consequently, when you do a lookup, you need to use an initial context that has been initialied with configuration properties that target a node from P2.
Now, one thing for sure, we should probably modify HA-JNDI so you can initialize it by giving it a partition name. Otherwise, if you have two partitions on the same node, you may not tell with which partition you want to work. You can add this as a feature request for clustering on SF (*if it solves a real problem for you only*)