I found the problem:

* init uses waitpid(-1,,), thus the __WALL flag is not set
* without __WALL, only processes with exit_signal == SIGCHLD are reaped
* it's impossible for user space processes to die with another
exit_signal, forget_original_parent changes the exit_signal back to
SIGCHLD ("We dont want people slaying init"), and init itself doesn't
use clone.
* kernel threads can die with an arbitrary exit_signal.

Alan, which fix would you prefer:
* init could use wait3 and set __WALL.
* all kernel thread users could set SIGCHLD. Some already do that
(__call_usermodehelper).
* the kernel_thread implementations could force the exit signal to
SIGCHLD.

I'd prefer the last version. 
The attached patch is tested with i386. The alpha, parisc and ppc
assember changes are guessed.

--
	Manfred