This does look like a promising diagnosis.  I'll look at it this weekend.

When using parallel recurse make, several child make processes will  be waiting in pselect.
When some make is writing a token in the pipe, several of these processes will leave the pselect, and trying to read the pipe.
Only 1 will actually get the token.  All others will be blocked in the read.
Eventually tokens will be written to the pipe, unblocking these make processes.

Solution : make the pipe non blocking, in order for all the processes that miss the token to call pselect again.