Ah, it's been a while since I did this and I forgot about the granularity of
the devel branch :-)
There it bisects to 73a146bec7f75da9f78f6d54329c980b75a2318d ("commit
bash-20130215 snapshot"). I'm pretty sure it's related to this change to add
the signal handler inside initialize_shell_signals, which sets a new handler if
SIGTERM isn't SIG_HARD_IGNORE.
If I read the code correctly, in that case we will resume due to SA_RESTART,
but bubbling up readline() will now return 0, which will be passed to
yy_readline_get(), which returns EOF if current_readline_line is 0, so the
shell closes.
I'm not quite sure how best to handle this, maybe setting SIG_IGN as the
default signal handler for interactive shells if there's nothing to inherit?
This is one of those cases where SA_RESTART doesn't quite mimic SIG_IGN...
