fix(zygote ext): handle NamedTuple cotangents in vofa_u_adjoint (EnsembleSolution end-time grads)

[ChrisRackauckas-Claude]

Please ignore until reviewed by @ChrisRackauckas.

Problem

Reverse-mode (Zygote) differentiation through an EnsembleSolution crashes when the loss reads only a scalar field of each trajectory's solution — e.g. the halting time produced by a terminate! callback:

T_stars = [sol.t[end] for sol in ensemble_sol.u]
mean(T_stars)

with

ERROR: MethodError: no method matching ndims(::@NamedTuple{…})
  [1] RecursiveArrayTools.VectorOfArray(vec::Vector{@NamedTuple{…}})
  [2] vofa_u_adjoint(d::Vector{@NamedTuple{…}}, A::EnsembleSolution{…})

ForwardDiff works; Zygote does not.

Cause

EnsembleSolution <: AbstractVectorOfArray, so accessing .u hits the literal_getproperty(::AbstractVectorOfArray, ::Val{:u}) adjoint → vofa_u_adjoint, which rewraps the per-element cotangents in a VectorOfArray. When only a scalar field of each per-trajectory ODESolution is differentiated, Zygote's cotangent for each element is a structural NamedTuple ((u = nothing, t = […])), not an array. VectorOfArray(::Vector{NamedTuple}) then calls ndims on a NamedTuple, which has no method.

This is the structural-tangent analogue of the ZeroTangent/size(::ZeroTangent) case fixed in #606 (same function, sibling failure mode) — which is likely why the report looked handled but wasn't.

Fix

In both vofa_u_adjoint methods, detect when the mapped cotangents are not array-like and pass them through as a plain vector instead of forcing a VectorOfArray/DiffEqArray reconstruction.

Test

Added to test/AD/adjoints.jl, mirroring the #606 test: feeds NamedTuple cotangents through vofa_u_adjoint for both VectorOfArray and DiffEqArray, and a mixed NamedTuple/nothing/ZeroTangent case. Verified locally that the new test reproduces the exact ndims(::NamedTuple) error on the unmodified source and passes with the fix; the full AD/adjoints.jl suite passes.

Scope / not addressed

This fixes the crash so the cotangent flows through. Whether GaussAdjoint (or other solve adjoints) then propagates the event-time (sol.t[end]) sensitivity correctly through the ContinuousCallback/terminate! is a separate, downstream concern (the substance of SciML/DifferentialEquations.jl#1149) and is not claimed to be solved here.

Fixes SciML/DifferentialEquations.jl#1149

🤖 Generated with Claude Code