hessian: the two nested dual layers share a tag, so results depend on the code path and chunk size

[devmotion]

ForwardDiff.hessian evaluates f on nested duals that carry the same tag at both layers, Dual{T,Dual{T,V,N},N} — HessianConfig passes one tag = Tag(f, V) to both of its inner configs (src/config.jl#L221-L227) and Base.eltype reflects that (src/config.jl#L256).

The two layers are therefore type-indistinguishable, so ForwardDiff.value / ForwardDiff.partials called inside f cannot tell them apart, and the tag-ordering machinery (≺, DualMismatchError) has nothing to compare. Which layer gets peeled then depends on how a given code path happens to nest its evaluations.

1. One function, three Hessians and two gradients

The layer is dropped from an intermediate here: ForwardDiff.value(z[1]) is order 1, while f's return value is a full order-2 Dual{T,Dual{T,V,N},N}. Extraction therefore has both layers to read and nothing about the result type looks wrong — the damage is that the one-layer intermediate is re-absorbed at the inner layer on its way back up, the shared tag making the two indistinguishable.

using ForwardDiff, StaticArrays, DiffResults
f(z) = sum(abs2, z) + ForwardDiff.value(z[1]) * z[2]
x  = [1.0, 2.0, 3.0]
sx = SVector(1.0, 2.0, 3.0)

ForwardDiff.hessian(f, x)                                                           # H[1,2]=1.0  H[2,1]=0.0
ForwardDiff.hessian(f, x, ForwardDiff.HessianConfig(f, x, ForwardDiff.Chunk{1}()))  # H[1,2]=0.0  H[2,1]=0.0
ForwardDiff.hessian(f, sx)                                                          # H[1,2]=0.0  H[2,1]=1.0

DiffResults.gradient(ForwardDiff.hessian!(DiffResults.HessianResult(x),  f, x))     # [4.0, 5.0, 6.0]
DiffResults.gradient(ForwardDiff.hessian!(DiffResults.HessianResult(sx), f, sx))    # [2.0, 5.0, 6.0]

Three Hessians and two gradients for the same function, varying with the chunk size, the array type and the result type.

2. A silently zeroed gradient

Here the output itself is order 1, so extraction has no outer layer to read, and the surviving first-order partials are discarded rather than used:

g(z) = ForwardDiff.value(sum(abs2, z))
DiffResults.gradient(ForwardDiff.hessian!(DiffResults.HessianResult(x),  g, x))   # [2.0, 4.0, 6.0]
DiffResults.gradient(ForwardDiff.hessian!(DiffResults.HessianResult(sx), g, sx))  # [0.0, 0.0, 0.0]
ForwardDiff.gradient(z -> sum(abs2, z), x)                                        # [2.0, 4.0, 6.0]

hessian!(::ImmutableDiffResult, f, ::StaticArray) builds dualize(T, dualize(T, x)) and extracts with nested value(T, ·). Because both layers share T, value(T, ·) peels a lone Dual{T,V,N} down to a plain Float64, and the following partials(T, ·, i) hits the zero(x) fallback. The zero Hessian is correct — no second-order information survived — but the gradient should be [2.0, 4.0, 6.0].

Cause and suggested fix

The two layers need distinct tags: an inner T = Tag(f, V) and an outer TO derived from it, so that T ≺ TO. Then value(TO, ::Dual{T}) correctly leaves an inner-layer value alone, and partials(TO, ::Dual{T}) is zero. Both cases above come out right with no special-casing, and the result stops depending on the chunk size.

The paths that already use two tags give the consistent answer today: hessian(f, ::StaticArray) goes through jacobian(Base.Fix1(gradient, f), x), which builds a fresh tag for each level.

Version: ForwardDiff master (v1.4.5), Julia 1.12.7, StaticArrays 1.9.19.