Changes

Shepperd cleanup

This fixes Solve2 to do the better U3 solve to propagate the
state and then adds a single U5 solve to generate the STM.

There's a comment in Shepperd's paper that you don't want to
generate higher order recursive fractions from lower order, so
do the work to recompute it.

Tests have been fixed to be deterministically random so they
won't randomly break.
 (commit: 31bc132)