With some elaborate redirections, it appears that some file descriptors are not being closed prior to spawning off processes. After running the attached script, one can observe the FDs of the sleep process by typing:

$ ls -l /proc/`pidof sleep`/fd/

This gives:
total 4
lr-x------    1 b        b              64 Nov 18 18:32 0 -> /dev/null
l-wx------    1 b        b              64 Nov 18 18:32 1 -> /dev/null
l-wx------    1 b        b              64 Nov 18 18:32 10 -> pipe:[245971]
l-wx------    1 b        b              64 Nov 18 18:32 2 -> /dev/null

As you can see, there is an extra FD 10 there which is inherited by sleep. The other end of that pipe is attached to the script. Hence the script will hang until the sleep process terminates (or whatever other long-lived process or daemon is spawned by this script).

Looking at bash's source, it appears related to the add_undo_redirect function and its actions. (Should the FD have the close-on-exec flag set?)

The same script operates correctly under zsh and dash.

(The reasons for the crazy redirections is to obtain the exit status of "Stuff" portably. In bash, PIPESTATUS would suffice but is not portable.)

Thanks,
Bernard.