ChainRules pushforward on Partials sometimes returns Array

julia> x, pf = frule(sin, 0.0)
(0.0, ChainRules.var"#83#sin_pushforward#59"{Float64}(0.0))

julia> pf(Zero(), Partials((1.0,)))
1-element Partials{1,Float64}:
Error showing value of type Partials{1,Float64}:
ERROR: MethodError: no method matching getindex(::Tuple{Float64}, ::Int64, ::Int64)

yeah yeah show fails (trying to do 2d access??) but that's not what I want to debug.

using ModelingToolkit

ForwardDiff.can_dual(::Type{Expression}) = true
ForwardDiff.can_dual(::Type{<:Expression}) = true

@variables x y

julia> a, pf = frule(sin, x)
(sin(x), ChainRules.var"#83#sin_pushforward#59"{Operation}(x))

julia> pf(Zero(), Partials((ModelingToolkit.Constant(1),)))
1-element Array{Operation,1}:
 +(cos(x) * 1)

This is not a Partials array!! And a Dual(x, p) where p is a Partials array is not the same as when p is an Array, so this ruins everything.

I suspect this is a Partials promotion problem and less sure that it's a ChainRules problem, so I'm opening this here and pinging @oxinabox

Start by reproducing the frule examples with Partials and the ModelingToolkit Operation case. Trace the Partials promotion used by the ChainRules pushforward and compare it with the Dual construction; done means the pushforward preserves a Partials tangent rather than returning an Array, without breaking the shown examples.