The coercion optimizer was optimizing coercions inside of branched
axiom applications, sometimes invalidating the branch choice within
the axiom application. Now, we check to make sure we are not
invalidating this invariant before proceeding with the optimization.