Replace Nothing with Missing

[Tokazama]

We've used the convention of returning nothing when something isn't known at compile time (e.g., known(1) == nothing). I think it might actually make more sense to use missing. A minor motivation for this could be that the "known" information is "missing" at compile time, so it has a more clear meaning. But I think the biggest benefit is that you can still operate on return from known. For example, ArrayInterface.known_length(x) = prod(known_size(x)) would work.

This would be a breaking change and could require some extra work down stream, so I want to get opinions on this before getting a bunch of PRs ready.

[JeffreySarnoff]

Would it make more sense to expand known so known(x) where !is_static(x) is x?

[Tokazama]

Do you mean that known(1) == 1 and known(static(1)) == 1? This is what dynamic does. I usually use known on types so that known(Int) === nothing and known(StaticInt{1}) === 1. This is useful for pulling type info down into a useable form, so I'm not sure we'd want to return Int in the scenario.

[JeffreySarnoff]

I understand that, and I use dynamic for the same purpose.
Rereading your initial note, I agree that it makes sense to enable
ArrayInterface.known_length(x) = prod(known_size(x)) and any other otherwise unavailable definitions.

so +1