The specific interfaces for the reply side of a protocol can't handle
error replies if the routine has out parameters because then the replies
are too small since the error messages only contain a return value.

This is easy to see by examining the generated skeletons.  Run `make term'
and look at term/device_replyServer.c:_Xdevice_open_reply.  It returns
directly if msgh_size isn't the size of a successful message.

The term translator works around this problem by turning off type checking
by passing -DTypeCheck=0 when compiling term/device_replyServer.c.
However, this is only acceptable if the server sending the replies can
be trusted, which should be the case here since the server is Mach,
or interposed itself between the client and Mach in a privileged fashion.

The root of the problem is that the *_reply.defs interfaces aren't
distinguished by mig from regular interfaces, and the current behavior
is the expected behavior for simple routines.  We could use the
currently ignored retcode tag to distinguish them, which was added for
compatibility with OSF Mig.  Though, we should check that this won't
break the compatibility first.

A more involved but ideal fix would be to add a new flag to mig for
generating the reply side of an interface.  This would also free us from
maintaining the reply versions of interfaces.