Hi, I'm actually the person who wrote the static type checker. Essentially we get around the difficulties of mixing objects and ADTs by having very limited support for subtyping. This way we can still carry out unification-based inference. The subtyping bounds are not propagated during unification so we don't need to worry about calculating closures. This does mean that type inference is incomplete with respect to subtyping, but I haven't found a satisfactory system that would let us do otherwise.
I would also like to note that we do also have a new feature where types can be inferred if you provide tests in a "where:" block attached to a function. This infers a function's type solely from the examples of usage provided.
If you have any other questions feel free to ask!