Wildfly recovery testing on the JDK orb blows (see the attached stack trace for details) up in ExtendedResourceRecord when we narrow a resource (using OTSAbstractRecordHelper.narrow (theResource). The reason we don't think it happens with JacORB is that they have an optimization that avoids the remote call for is_a whereas JDK orb does not have the optimisation. I built a version of narayana that replaces the narrow call with an unchecked_narrow and this fixes the recovery issue.
A little bit of implementation detail: JacORB loads the stub for the object being narrowed and calls its _ids method. If any of those ids match the target type then it avoids the remote call. Our guess is that this local search succeeds (but it's a guess).
On the other hand, looking at the code for the JBoss repackaging of the OpenJDK ORB (org.jboss.openjdk-orb): it first checks locally (StubAdapter.getTypeIds) and then falls back to doing the RPC call (getClientRequestDispatcher). The stack trace (see attachment) shows that it goes through the latter path.
After consultation with Mark, our OTS expert, he is of the opinion that the CORBA spec isn't prescriptive about how narrow and is_a are supposed to work. So as long as the right answer is returned, why should we be concerned?
As a parallel task we should investigate why the local check for the OpenJDK ORB doesn't work. Maybe it's an issue for them to fix?