John,

> If we consider a parallel-processing VM, then Andreas' patch won't come 
> close to working, since it won't guarantee that the forked process isn't 
> started too early. Therefore, I wasn't considering a parallel processing 
> scheduler.

  Neither the program that assumes to get #(1 2 3) from your example
code.  So, these are "par", roughly speaking.

  And, the patch makes the behavior easier to reason if we take
current implementation.  So there is some value.

-- Yoshiki