supports_initialization: use get_jumps to keep Enzyme readonly

[ChrisRackauckas-Claude]

Please ignore until reviewed by @ChrisRackauckas.

Summary

supports_initialization(sys) short-circuits on isempty(get_systems(sys)), so the call to jumps(sys) here only ever reaches the early-return branch where jumps(sys) === get_jumps(sys). However the unreached copy(js) + append!(js, namespace_jumps(subsys)) branch in jumps is enough to prevent Enzyme's static analysis from proving sys read-only at this call site — and that breaks Enzyme.gradient through any closure that reaches _late_binding_update_u0_p_impl, e.g. remake(prob; p = repack(tunables)) on any MTK-generated problem.

Swap to get_jumps(sys), which is a pure field accessor and identical in this short-circuited position.

Verification

The MWE from #4550:

@parameters a b c
@variables x(t) y(t) z(t)
eqs = [D(x) ~ a*x + y + z, 0 ~ y^3 + y - b*x, 0 ~ z^3 + z - c*y]
@mtkbuild sys = ODESystem(eqs, t)
prob = ODEProblem(sys, [x => 1.0], (0.0, 0.1),
                  [a => -0.5, b => 2.0, c => 1.5];
                  guesses = [y => 1.0, z => 0.5], use_scc = false)
iprob = prob.f.initialization_data.initializeprob
itunables, irepack, _ = SS.canonicalize(SS.Tunable(), parameter_values(iprob))
remake_only = let iprob = iprob, irepack = irepack
    p -> sum(remake(iprob, p = irepack(p)).u0)
end
Enzyme.gradient(Enzyme.Reverse, remake_only, itunables)

Before this PR: EnzymeMutabilityException at supports_initialization → jumps(sys).

After this PR: Enzyme.gradient completes (returns a vector). The returned gradient itself is currently [0.0, …] because MTK#4181 (Enzyme rule for SetInitialUnknowns) is still missing — that's a separate semantic-correctness issue that this PR doesn't touch; it just removes the compile-time blocker.

What downstream this unblocks

This is Layer 1 of a 4-layer Enzyme-through-MTK-remake stack documented during the SciMLSensitivity #1323 / #1359 investigation. Layers 2–4 (FunctionWrappersWrappers cache mutation, SciMLBase Const(loss) EnzymeRuntimeActivityError, Enzyme.jl MixedDuplicated{NonlinearSolution} MethodError) still bite once you go past remake to remake + solve. Those need separate upstream work. The @test_broken Enzyme.gradient blocks in SciMLSensitivity's mtk.jl / parameter_initialization.jl / desauty_dae_mwe.jl stay until all four are resolved.

Refs

• Closes Enzyme.gradient through remake fails with EnzymeMutabilityException via supports_initialization → jumps(sys) #4550
• Related: Add Enzyme rule for SetInitialUnknowns to support remake autodiff #4181 (Enzyme rule for SetInitialUnknowns — Layer 4, separate)
• Enzyme: Cannot differentiate through MTK DAE solve (MutabilityException / RuntimeActivityError) SciMLSensitivity.jl#1359, Bump version #1323

Test plan

• Local: Enzyme.gradient on the MWE no longer errors
• CI green

[ChrisRackauckas-Claude]

Layer 2 status update

Confirmed independently that landing this Layer 1 fix is necessary but not sufficient: with this PR applied, bare Enzyme.gradient through remake succeeds, but remake + solve still fails with EnzymeMutabilityException. I reduced Layer 2 to a no-MTK MWE and filed it upstream at EnzymeAD/Enzyme.jl#3117:

Enzyme.gradient over NonlinearSolve.solve works for a NonlinearProblem whose f is a plain anonymous function, but fails as soon as f is a RuntimeGeneratedFunction (or any wrapped-function pattern — eval_expression = true still fails). The failing IR variable is iprob.f1.i.sroa.0.0 — a Union-tag byte store SROA'd through NonlinearFunction.f reported inside SciMLBase's mutable NonlinearProblem constructor.

The full breakdown of layers 2–4 is in the discussion on SciMLSensitivity#1359. This PR remains the right Layer 1 fix on its own; the three @test_broken Enzyme.gradient blocks in SciMLSensitivity (mtk.jl, parameter_initialization.jl, desauty_dae_mwe.jl) flip to passing once both this and EnzymeAD/Enzyme.jl#3117 land.

I also filed SciML/FunctionWrappersWrappers.jl#51 (defensive inactive_type rule for cache storage) — that one doesn't fix Layer 2 by itself but removes one conservative-IPA layer.

[AayushSabharwal]

Test? I feel like this is very easy to accidentally break

[ChrisRackauckas]

The only real answer for whether Enzyme works is if Enzyme works, so we'll get that in the integration tests at the end of the journey.