I never said the problem was with threading. But threads are required to create the problem, so it's fitting -- no, necessary -- that threads appear in the problem description. All I did was describe the problem more precisely than had been done before (although I failed because I forgot to mention the socket had to be inherited).

I suspect the problem is the fork is waiting for accept to release an exclusive lock on some resource so it can duplicate that resource or otherwise proceed with the fork. I won't even conjecture on whether it's fixable.