Fix iterate(::AbstractTimeseriesSolution) to respect AbstractArray contract

[ChrisRackauckas-Claude]

Summary

The existing Base.iterate(sol::AbstractTimeseriesSolution, state = 0) returns solution_new_tslocation(sol, state) — a fresh solution of the same concrete type with a bumped tslocation. That violates the AbstractArray contract (an array's first element must not have type identical to the container) and trips Julia 1.12's LinearAlgebra.norm_recursive_check:

ArgumentError: cannot evaluate norm recursively if the type of the initial
element is identical to that of the container

This surfaced in OrdinaryDiffEq's VectorOfArray/StructArray compatibility testsets (OrdinaryDiffEqLowStorageRK, OrdinaryDiffEqSSPRK) where sol_SA ≈ sol_SV compares two ODESolutions whose u is Vector{VectorOfArray{Float64, 2, StructVector{SVector{1, Float64}}}}. The generic Base.isapprox for AbstractArray falls through to LinearAlgebra.norm, which on 1.12 hits the check.

Change

Replace the custom iterate with the natural linear-indexing form:

function Base.iterate(sol::AbstractTimeseriesSolution, state = 1)
    state > length(sol) && return nothing
    return (@inbounds sol[state], state + 1)
end

For VectorOfArray-backed solutions this yields scalars (matching eltype(sol)), which is what LinearAlgebra.norm and isapprox expect. solution_new_tslocation is still exported for the plot recipe to use directly — it was only ever load-bearing inside this method.

Breaking behavior note

Minor behavior change: downstream code that did for s in sol expecting s to be a full solution-like object (with .retcode, .prob, etc.) will need to change. In practice such callers should iterate sol.u / sol.t explicitly, or pair solution_new_tslocation with eachindex(sol.u). One such site lives in OrdinaryDiffEq.jl/test/ad/ad_tests.jl:388 (for s in sol with s.retcode); that's a test-only call and can be fixed in-place to read sol.retcode once.

Bumped to 3.4.0 to flag the iteration semantics change.

Test plan

• Added @testset "iterate does not yield the container" in test/solution_interface.jl with: type check, non-equality check, isfinite(norm(sol)), sol ≈ sol. All pass on Julia 1.12.
• Unblocks the upstream failing VectorOfArray/StructArray compatibility testsets in OrdinaryDiffEqLowStorageRK and OrdinaryDiffEqSSPRK (verified locally against monorepo-dev'd sublibs).
• CI on this PR
• Run OrdinaryDiffEq.jl#3513 against this to confirm the downstream fix can be closed.

Closes-replaces-workaround-for: SciML/OrdinaryDiffEq.jl#3513

🤖 Generated with Claude Code

[ChrisRackauckas-Claude]

Updated per review: deleted the custom iterate entirely rather than replacing it with an explicit linear-indexing form. The AbstractArray fallback (iterate over eachindex(sol)) is the only iteration that's type-consistent with eltype(sol) / size(sol) — any custom implementation risks re-introducing the same class of bug. Same behavior as the previous commit in practice (both delegated to sol[state]), just less code.