Custom gradient not getting called, incorrect zero gradients

So I'm trying to specify a custom gradient for an FFI function:

import PyCall
import ReverseDiff

math = PyCall.pyimport("math")
pysin(x) = math.sin(x[1])
ReverseDiff.@grad function pysin(x)
    @error "fwd"
    function pullback(δ)
        @error "bwd"
        (δ * math.cos(x[1]), )
    end
    math.sin(x[1]), pullback
end
ReverseDiff.gradient(pysin, [1.5])

but unfortunately this outputs:

julia> include("difftaichi/zygote_zero_nothing_bug.jl")
1-element Array{Float64,1}:
 0.0

instead of the errors that we would expect. My conclusion from this is that my ReverseDiff.@grad custom gradient is not getting called at all. In addition, the gradient returned is zero, instead of producing an error. So I see two issues I guess:

• the custom gradient is not working
• ReverseDiff gives me incorrect gradients when it doesn't know what to do instead of producing an error (unsound-and-incomplete as opposed to sound-but-incomplete).

cc @ChrisRackauckas

Reproduce the example in difftaichi/zygote_zero_nothing_bug.jl, starting with the ReverseDiff.@grad definition and the ReverseDiff.gradient call around the PyCall math.sin FFI function. Trace whether the custom gradient entry point is reached and how the unsupported operation produces a zero gradient; done means the behavior is covered by a clear regression test and the reported gradient behavior is no longer incorrect.