Fix for OrdinaryDiffEq 7 / SciMLBase 3 / RecursiveArrayTools 4 (complements #565)

[ChrisRackauckas-Claude]

Fix MethodOfLines for OrdinaryDiffEq 7 / SciMLBase 3 / RecursiveArrayTools 4

Please ignore until reviewed by @ChrisRackauckas.

This fixes a pre-existing master CI failure on the Diffusion, Diffusion_NU,
Nonlinear_Diffusion(_NU), Convection(_WENO), DAE, Wave_Eq_Staggered and related
test groups. master already resolves to OrdinaryDiffEq 7.0.0 / SciMLBase 3.17 /
RecursiveArrayTools 4.3 via the loose OrdinaryDiffEq = "6, 7" / SciMLBase = "2, 3"
compat, so the breakage reproduces on a clean checkout independently of any dependabot
bump. It also unblocks the DifferentialEquations-8 side of #564 and complements #565
(the DomainSets 0.8 fix). No DomainSets changes are bundled here.

Two root causes

1. length(sol) / size(sol) on a PDETimeSeriesSolution (SciMLBase 3 / RAT 4)

PDETimeSeriesSolution <: AbstractTimeseriesSolution <: AbstractVectorOfArray. MoL stores
the solution u as a Dict{Num, Array}. RAT 4's generic
Base.size(::AbstractVectorOfArray) iterates sol.u expecting a vector of arrays and
calls size(u, d) on each element — but iterating the Dict yields Pairs, giving
MethodError: no method matching size(::Pair{Num, Matrix{Float64}}, ::Int64). For
single-variable solutions this silently returned 1 (the number of dvs) instead of
erroring, so for i in 1:length(sol) loops were quietly only checking i = 1.

Fix (src/interface/solution/common.jl): define the array interface for the MOL
PDETimeSeriesSolution in terms of its saved time points —
length(sol) == length(sol.t), size(sol) == (length(sol.t),), ndims(sol) == 1,
plus axes/firstindex/lastindex. This restores the intended time-series semantics
used throughout the tests (for i in 1:length(sol)).

2. OrdinaryDiffEq 7 stopped re-exporting most solver names

using OrdinaryDiffEq (v7) only re-exports a curated subset (Tsit5, Vern6-9,
Rosenbrock23, Rodas5P, FBDF, DefaultODEAlgorithm). Solvers used by the test suite
that are no longer re-exported now live in sub-packages:

Solver	Sub-package
Rodas4, Rodas4P, Rosenbrock32	OrdinaryDiffEqRosenbrock
TRBDF2	OrdinaryDiffEqSDIRK
SSPRK33	OrdinaryDiffEqSSPRK
Euler, SplitEuler	OrdinaryDiffEqLowOrderRK

Fix: add explicit using OrdinaryDiffEq<SubPkg>: <Solver> imports in the affected test
files and add the four sub-packages to the test-only [extras] / [targets] in
Project.toml. src/ itself references no OrdinaryDiffEq solver names, so no source
change is needed for this half.

Note on the Test 03 conservation tolerance

Fixing length(sol) exposed a latent test bug in MOL_1D_Linear_Diffusion.jl Test 03
(homogeneous Neumann diffusion). The assertion sum(u) ≈ 0 atol = 1e-10 was previously
only evaluated at i = 1 (initial condition, exact to ~1e-14) because of the broken
length. With length(sol) corrected, it now runs at every saved time. The sum(u) is a
rectangle-rule proxy for the conserved integral ∫u dx = 0; the edge-aligned grid
places nodes offset by half a step, so its rectangle sum drifts to ~1e-7 at the default
Tsit5 tolerances, while the center-aligned grid stays ~1e-11 at the same tolerances
(verified locally). The drift is therefore a property of the crude quadrature on the
edge-aligned grid, not solver noise and not a discretization regression. The tolerance is
loosened to 1e-6 (still verifying conservation to six digits) with an explanatory
comment.

Local verification (Julia 1.11, env resolves OrdinaryDiffEq 7.0.0 / SciMLBase 3.17 / RAT 4.3 / MTK 11.26)

Baseline on clean master (Diffusion group): 99 passed, 6 errored — exactly
3× UndefVarError: Rodas4 and 3× MethodError: size(::Pair{Num,Matrix}, ::Int).

After this PR, run per CI group (GROUP=… Pkg.test()):

Group	Result
Diffusion	PASS (289/289)
Diffusion_NU	PASS
Nonlinear_Diffusion	PASS
Nonlinear_Diffusion_NU	PASS
Convection	PASS
Convection_WENO	PASS
Wave_Eq_Staggered	PASS
Integrals	PASS

Remaining failure — out of scope here, NOT silenced

DAE (MOL_1D_PDAE.jl), MOL_Interface1 (MOLtest1.jl, "Rearranged Robin") and
Higher_Order (MOL_1D_HigherOrder.jl, KdV soliton) still error with

DAE initialization failed: your u0 did not satisfy the initialization requirements,
normresid = … > abstol = 1.0e-6.

This is a separate SciMLBase-3 behavior change: the default initializealg now
runs a strict CheckInit that rejects MoL's discretized initial condition (the BC /
algebraic-constraint rows are not satisfied to 1e-6). On clean master these same
tests error earlier with UndefVarError: Rodas4 / UndefVarError: SciMLBase, so the
init failure was previously masked. Fixing it properly requires either MoL generating a
consistent IC for the constrained rows or threading an initializealg
(BrownFullBasicInit, which would add an OrdinaryDiffEqNonlinearSolve test dep) —
deliberately left for a separate focused PR rather than silenced here.

🤖 Generated with Claude Code