p. 3, line 47 is a change from before. We now allow one to make MPI
calls before you initialize threads. I can see the advantages here
but also risks. First, just because an MPI call has completed from
the user perspective does not guarantee it is done from the MPI system
perspective. Second, a library would be hard pressed to know for sure
that there are no pending calls. Thus, I wonder if we are opening a
can of worms we will regret.
p. 4, line 9: I think what we want here is that you can link code with
a call to MPI_THREAD_INIT but the call should never be made. Saying
something about linking might make it clearer.