Ok, I found it myself.
We used an own PK class with own write/readObject methods.
Obviously this must not be. Without the writeObject / readObject method it works!
Other things I learnt during testing:
Invalidation works with BMP Entities and local interfaces! Unfortunately the just changed Entity is invalidated as well (although its new data is the reason for invalidation).
Finally: this technique is a great relief for former self-built-cache programmers! Just use local interfaces, commit-option A and cache-invalidation set to true.
BTW: JBoss offers so many container-configurations. Why not one for Standard BMPs with cache invalidation?