I don't know the internals of signal handling well enough to know a definitive answer. However given that you are doing a pretty nasty stress test, I would not be surprised if you were able to smoke out very subtle races.

With that in mind, when I look in perlipc I notice that there is a comment in there about SysV which says that on BSD and POSIX systems you don't need to reset the signal handlers, but on SysV you both need to reset it, and you need to reset it after the wait. I don't know Unix well enough to know what Tom Christensen is referring to there, or whether current Linux and Solaris would show the POSIX or the SysV behaviour. So I would test removing the resetting of the handler entirely, or failing that I would try moving that resetting inside of your loop.

If neither of those helps and nobody else comes up with any useful suggestions, I would use the perlbug utility to submit a bug report to p5p. (If you do figure out the issue and see that there is a way to improve the documentation to make it clearer to other people, you can always send a patch to perlipc.pod in...)