Rosenbrock23 drops SLArray labels in out-of-place stage broadcast

[ChrisRackauckas-Claude]

Reproducer

The LabelledArrays.jl/Core job in RecursiveArrayTools' integration workflow fails on clean RecursiveArrayTools master (88fc415) and current OrdinaryDiffEq packages:

using LabelledArrays, OrdinaryDiffEq, SciMLBase

LorenzVector = @SLArray (3,) (:x, :y, :z)
LorenzParameterVector = @SLArray (3,) (:σ, :ρ, :β)

function f(u, p, t)
    x = p.σ * (u.y - u.x)
    y = u.x * (p.ρ - u.z) - u.y
    z = u.x * u.y - p.β * u.z
    return LorenzVector(x, y, z)
end

u0 = LorenzVector(1.0, 0.0, 0.0)
p = LorenzParameterVector(10.0, 28.0, 8 / 3)
solve(ODEProblem(f, u0, (0.0, 10.0), p), Rosenbrock23())

FieldError: type StaticArraysCore.SizedArray has no field `y`, available fields: `data`
  [1] f(u::StaticArraysCore.SizedVector{3, Float64, Vector{Float64}}, ...)
  [2] perform_step!(..., ::Rosenbrock23ConstantCache, ...)

Failing clean-master runs:

• https://github.com/SciML/RecursiveArrayTools.jl/actions/runs/30928196501/job/92055876810
• https://github.com/SciML/RecursiveArrayTools.jl/actions/runs/31269971272/job/93134151575

I reproduced the same error locally from a detached clean RecursiveArrayTools master checkout and a fresh LabelledArrays master checkout.

Regression range

The last green integration run on 2026-07-24 resolved OrdinaryDiffEq 7.1.3 / OrdinaryDiffEqRosenbrock 2.4.2. The first failing run on 2026-08-03 resolved OrdinaryDiffEq 7.2.1 / OrdinaryDiffEqRosenbrock 2.6.3.

Running the old RecursiveArrayTools source (8d8c6a2, used by the last green run) against today's resolver still fails identically, so this is not a RecursiveArrayTools source regression.

OrdinaryDiffEq commit 100fe2369 (#3996) is the relevant source transition: it changed the Rosenbrock constant-cache linear-solve result from a generic reshape to ArrayInterface.restructure(uprev, ...). With an SLArray template today:

r = ArrayInterface.restructure(u0, ones(3))
typeof(r)            # LArray{...}
typeof(r * 2)        # LArray{...}
typeof(u0 .+ r)      # SizedVector{3, ...}
typeof(u0 + r)       # SLArray{...}

The constant cache then computes tmp = @.. uprev + dto2 * k₁; FastBroadcast drops the labels and passes the resulting SizedVector to f.

This is related to wrapper preservation in #4101 and JuliaArrays/ArrayInterface.jl#498, but those changes cover AD tracked arrays, not static labelled arrays. A fix should preserve the state container at every out-of-place Rosenbrock stage and add this LabelledArrays solve as a regression test.