frule invalidation broken

[oxinabox]

We actually have a test for this: https://github.com/JuliaDiff/Diffractor.jl/blob/b37f585d4ba8fa5cb607cd75aaabb5464eef6633/test/rules.jl
but it turns out this file has never been linked up to the test/runtests.jl file.
So it wasn't getting run by CI.

The last test in that file fails.
Both on 1.10 and on nightly

[oxinabox]

We do manually insert an edge to the MethodInstance being AD'ed here

Diffractor.jl/src/codegen/forward.jl

Line 114 in b37f585

ci.edges = MethodInstance[mi]

Do we need to manually insert an edge to every method we call in the generated code?
Surely not, we haven't run type-inference so we can't even know that, and that surely already will insert its own edges, right?

[Keno]

No, but we might need an edge for the frule of the MethodInstance we're currently ADing to cover the case where an frule is added to something that had none before.

[oxinabox]

I don't think so.
That case would get caught from the calling code, no?
Since ∂☆ always calls ∂☆internal first
and that does (basically):

function ∂☆internal(...)
    res = frule(...)
    if nothing===res
        ∂☆recurse...  # this code we are talking about
    else
        shuffle_base(res)
    end
end

So adding a new frule should already invalidate it, since it invalidates the top line of ∂☆internal

[Keno]

Ah yes, probably. Something else going on then.