ForAllCo case of
coercionKind does an inefficient one-variable substitution. This could be improved by looking for nested
Furthermore, perhaps we don't need an in-scope set and the full substitution machinery here. The subst is simply propagating the update of a tyvar's kind. No structural changes at all. No need for smart coercion constructors or other processing. So if we're going to optimize this, it might be worth making a specialized version of
subst_co that operate over a
VarEnv Var instead of a full substitution.