Back to first principles: 1) Can you replicate the *failure* so you'll know whether you've fixed it or made it worse?
2) It's important to understand how the guests detect updates. If it's via a handshake, can you simulate that? I like Harry's almost-atomic update via COPY then RENAME. It wouldn't be hard to write a small assembler program to check the timestamp and issue the ERASE/RENAME if it matches, to minimize the window (though clearly 3 lines of Rexx is pretty fast these days). So the question then becomes whether this is sufficient, and for that you need #2 above.