I think that this may just be a problem with terminology. To me, thread-safe
means safe for threads. Period. If an MPI implementation calls a routine which
is itself not thread-safe, then I would claim this also makes the MPI
implementation non-thread-safe because, exactly as you point out, it is not
possible to protect against any uses outside of MPI.
So if an implementation is truly thread-safe then nothing extra need be said
about it. However, if y'all would like to argue that implementations which
claim to be thread-safe but really aren't should be required to document the
exact nature of their unsafeness, that certainly seems like a reasonable
request.
-- Eric Salo Silicon Graphics Inc. salo@sgi.com