[...]
OK, so the solution will always have a certain level of instability. Just to
put some idea of scale on my previous example graphs, it's possible that
LilyPond will be tossing up between using 5 systems and using 10 systems. 5
systems provides much better spacing but 10 systems has less penalties. The
total badness is _slightly_ less for 5 systems so Lily goes with that.
Then the user inserts an extra bar and the number of systems doubles. I think
that instability shouldn't occur on this scale.
Now, I haven't been playing around with any line penalties, but I have done
experiments with page turn penalties and I have seen scores go from 5 to 8
pages with only small changes.