kqueue uses a userland data structure. The netbsd PR is saying that the userland data structure gets corrupted, so that the event doesn't arrive. Thus the process stops.
That PR says their kernel is not buggy, but the kqueue handling. If they are updated on libuv, maybe they are seeing the same libuv bug? It sounds like a EVFILT_PROC is losing track of a child or exit. I don't know if cmake w/libuv is using EVFILT_PROC. But if the userland data structure is busted, well you can expect it to not notice completion... claudio did some minor work in signals, but I would expect any fallout from his changes to affect lots of stuff, not just 1 program. kqueue isn't like poll or select. It has way more monitoring ability, and if you screw up the userland memory, really weird things happen.
