RuntimeExceptions result in abnormal behavior from Axis2.
When an AxisFault is thrown, generally the fault is recorded in the message context, then a FaultFlow is invoked. Handlers operating in the FaultFlow context may then discover the fault through examination of the message contexts attached to its operation context.
If a runtime exception is thrown, however, the exception is lost, even though the FaultFlow is properly invoked. FaultFlow Handlers can no longer discover the exception which caused their invocation.
I recognize that best practice would include packaging every thrown exception in an AxisFault or AxisFault derivative. However, given that Axis2 does attempt to respond sensibly to RuntimeExceptions, I do not believe that having a "silent" FaultFlow invocation with no access to the originating exception is proper behavior.
The fix for this issue would involve (minimally) adding a catch clause to various methods in AxisEngine to set the failure reason on the message context when a runtime exception is intercepted, then rethrowing the original exception. Fault handlers would then be able to recover information from the originating exception.