Turns out we can instigate a core dump safely enough by sending an abort signal instead of a kill signal, and the difference in behavior is likely to still allow for valid data files, since the signal handler is a different thread that doesn't take locks. So now we should send an abort signal instead of a kill.