while calling libdiff. It's too ugly and not worth the recursion problems
when there is a malloc failure (which writes to stderr - now diverted via
the buf system, which calls malloc, which causes another error message etc).
We can live with the standard artificial slowdown, but reduce the time a
bit and only delay when we really need to (ie: when running as a server).
The usleep time could probably use some tuning, it basically needs to
replace the time that it used to take to fork a large process, exec gnudiff
and the time that gnudiff took before writing the initial output.
This eliminates a whole mess of other hacks I was considering that changed
use of xmalloc to alloca() etc. It was going too fast in the wrong
direction.
f07a3e4e23