I was analyzing sources of invalidations (maybe you are already familiar with it, but some context can be found here and here) and it seems that one of the most important is caused by this definition:

Int(::Infinity) = ∞

which is understandable given how it breaks a lot of the assumptions from inference in Base, where Int(n) isa Int, but now it can also be ::Infinity.

I can see how eliminating that line would make supporting the whole array interface really cumbersome, but maybe it is worth trying to see if it can be restructured to improve precompilation and loading times.