Typed thunks?

The discussion on FluxML/Zygote.jl#966 - thunks have really evolved into monads now, right? Especially now that we're adding methods for them to lots of linear algebra and other functions. Shouldn't AbstractThunk have a type parameter, then? Kinda like

abstract type AbstractThunk{T} end

struct Thunk{T,F<:Base.Callable} <: AbstractThunk{T}
    body::F
end

unthunk(t::Thunk{T}) where T = t.body()::T

Base.eltype(t::Thunk{T}) where T = T

macro thunk(body)
    func = Expr(:->, Expr(:tuple), Expr(:block, __source__, body))
    return quote
        f = $(esc(func))
        Thunk{Base._return_type(f,()), typeof(f)}(f)
    end
end


thnk = let A = rand(5,5), B = rand(5,5)
    @thunk A * B
end

eltype(thnk) == typeof(unthunk(thnk))

I've seen type inference fail with thunks quite a few times - maybe having typed thunks would help with that, too?

Start by reading the discussion in FluxML/Zygote.jl#966 and inspect the existing AbstractThunk, Thunk, unthunk, and @thunk entry points mentioned here. Determine whether typed thunks would address the reported inference failures; done requires a settled design and a clearly agreed scope for any implementation.