Ok, found the root of the problem.
Since the call is from the remote client, an additional thread-pool exists for remote invocations:
<remote connector-ref="remoting-connector" thread-pool-name="default"/>
The default pool is configured with 10. In my tests, the separate SLSB pool was also configured with 10 .
So it could never happen, that the SLSB got exhausted via remote calls, because the default thread pool already blocked and didn't allow the 11th call to be routed to the SLSB pool.
It simply waited for the next thread to be freed. So I increased the thread pool to 20 and I got the expected result.
Sorry for any inconvenience.