"Before this release, my primary priority was: prove that FRP is viable and good. If FRP is not the right way, it does not matter how good or bad the type checker is."
This is a lousy justification. If the intent was to prove that FRP was viable and good, why not do it in a dynamic type system? Implementing an extended Hindley Milner inference algorithm can be subtle, so why bother? I can manage a slow type checker, because "eh it's not 1.0", but an unsound type system, no.
I've tried Elm because I really wanted to program functionally a small webapp, but the type checker had so many holes in it that I had to translate the algorithms in Haskell, type them and fix them there before porting them back. I'm now really reluctant to try any new version of Elm.