If anyone is interested in static, incremental type inference (which
allows type safe redefinitions without #use), I worked out how to do
this for core ML in a paper in POPL '96.  As someone noted earlier, of
course this does not address the semantic issues (of changing
List.map, for example).

-Trevor