prepare_user_sparsity: unwrap operator mass matrices before findall (#2929)

[singhharsh1708]

Partially addresses #2929.

Bug

prepare_user_sparsity (lib/OrdinaryDiffEqDifferentiation/src/alg_utils.jl) seeds the sparsity pattern from the mass matrix with:

idxs = findall(!iszero, prob.f.mass_matrix)
for idx in idxs
    sparsity[idx] = prob.f.mass_matrix[idx]
    ...

when the mass matrix isn't a UniformScaling. findall/getindex need a concrete matrix; a SciMLOperator mass matrix (e.g. MatrixOperator) supports neither directly:

using OrdinaryDiffEq, SparseArrays, SciMLOperators
M = sparse([1.0 2.0; 0.0 1.0])
mass_matrix = MatrixOperator(M; update_func = (A, u, p, t) -> M)
jac_prototype = sparse([1.0 0.0; 0.0 1.0])
prob = ODEProblem(ODEFunction((du,u,p,t)->du.=u; mass_matrix, jac_prototype), ones(2), (0.0, 1.0))
solve(prob, ImplicitEuler())
# MethodError: no method matching keys(::MatrixOperator{...})

matching #2929's original report exactly.

Fix

Unwrap the operator to its concrete matrix first: convert(AbstractMatrix, mm). This only reads the operator's currently-cached array (SciMLOperators.matrix.jl's convert(::Type{AbstractMatrix}, L::MatrixOperator) = convert(AbstractMatrix, L.A)), not a live re-evaluation, so no update_coefficients! call is involved.

What this doesn't fix

While tracing #2929's repro I found the baseline case (no jac_prototype at all) has separately regressed since the issue was filed — it now fails with a different error, setfield!: immutable struct of type MatrixOperator cannot be changed. That's entirely inside SciMLOperators.jl: update_coefficients! on a MatrixOperator built with only an out-of-place update_func (no update_func!) does a raw L.A = ... field assignment on what's declared an immutable struct — update_coefficients (no bang) handles the same case correctly via @reset. Root-caused to src/matrix.jl:227 in SciMLOperators v1.24.3, confirmed with a minimal SciMLOperators-only reproducer (no OrdinaryDiffEq involved). That's a separate upstream bug I can't fix here; will report it in SciMLOperators.jl. #2929's repro won't fully pass end-to-end until that's also fixed, but this PR fixes the part that's actually ours.

Tests

New prepare_user_sparsity_tests.jl: MatrixOperator mass matrix (fails on master with the keys MethodError, passes here with the correct merged sparsity pattern), plus regression checks that the UniformScaling and plain-sparse-matrix paths are unchanged.

OrdinaryDiffEqDifferentiation Core suite passes (39/39, including 6 new).

AI Disclosure

Claude assisted with this work.