Would just subtyping StaticIndexing{I} <: AbstractVector{Int} be OK?

I think so

It also needs size rather than length

I just realised this is still incorrect... now we are wrapping Int indices in a AbstractArray which will change the indexing behavior.

I don't know why StaticArrays is doing this but I think we need to just limit the to_indices here to dispatch on A::StaticArray or this will just be wrong in some cases.

Another option is to not wrap Int or CartesianIndex in StaticIndices at all, so map(StaticIndices, ... needs a conditional on the type.

Well, I'm afraid limiting that to_indices dispatch may break some optimization. If this isn't (close to) good enough then I think I will leave this issue for someone else.

I suspect the rough idea behind StaticIndexing is so that StaticArrays can define implementations like f(A::AbstractArray, Idx::StaticIndexing...) without worry of ambiguities. Making it an StaticIndexing <: AbstractVector won't quite work, as far as I understand it, because it needs to wrap integers, matrices and other-dimensional arrays in addition to vectors.

It really can't dispatch on AbstractArray or we will just keep getting confused Discourse posts when it breaks.

Maybe supporting StaticArray and Array is enough for performance? Or a wider set of Base arrays?

I think it's less wrong to go with the changes that I've just pushed than keep the status quo, so I'm going to merge this (if CI passes and there are no objections). More comprehensive changes can happen in a separate PR (which I'm not planning to work on in the foreseeable future).