DiffEqBaseEnzymeExt: accept any sensealg Annotation (unblocks SciMLSensitivity #1456)

[ChrisRackauckas-Claude]

[DRAFT — please ignore until reviewed by @ChrisRackauckas]

Summary

DiffEqBaseEnzymeExt.augmented_primal / reverse rules on DiffEqBase.solve_up were signed for sensealg::Union{Const{Nothing}, Const{<:AbstractSensitivityAlgorithm}}. That signature was too narrow: under Enzyme.set_runtime_activity(Reverse), Enzyme can promote a module-level const _MTK_SENSEALG = GaussAdjoint(...) reference — or any sensealg captured in an ODEProblem field — to Duplicated{<:AbstractSensitivityAlgorithm} before the rule is matched. The rule then failed to dispatch and Enzyme fell back to native reverse, producing a MethodError.

This loosens both rules to accept any Enzyme.Annotation on sensealg (Const / Duplicated / MixedDuplicated / Active) and reads only sensealg.val inside; the shadow on sensealg has no semantic meaning because an AbstractSensitivityAlgorithm is a configuration token, not a derivative-carrying value.

Why this is safe

SciMLBase already declares EnzymeRules.inactive_type(::Type{<:AbstractSensitivityAlgorithm}) = true, expressing the same semantic. The narrow Const-only signature was an unnecessary additional restriction. The body of both rules now skips sensealg's slot when accumulating cotangents, regardless of which annotation it arrived under.

Reproducer

The @test_broken Enzyme block in SciML/SciMLSensitivity.jl test/mtk.jl calls solve(prob, Rodas5P(); sensealg = _MTK_SENSEALG, ...) under Enzyme.autodiff(set_runtime_activity(Reverse), ..., Duplicated(prob, diprob)). Before this PR, it failed with MethodError on the solve_up rule's sensealg signature.

After this PR, the rule dispatches and execution proceeds into _solve_adjoint. It then hits a separate downstream MethodError inside Enzyme's runtime_generic_augfwd trying to construct MixedDuplicated(::ODESolution, ...) without the required Base.RefValue{T1} second argument — the MTK runtime-activity wrapping issue tracked in SciMLSensitivity.jl#1359 / EnzymeAD/Enzyme.jl#3117. That is out of scope here, so the @test_broken in #1456 stays broken pending the upstream Enzyme fix.

Bumps

DiffEqBase 7.5.1 → 7.5.2.

Test plan

• CI green
• Confirms solve_up Enzyme rule still dispatches under plain Reverse (sanity, all existing tests)
• Manual: focused MWE shows MethodError moves from rule dispatch to downstream MTK wrapping (documented above)

Linked: SciML/SciMLSensitivity.jl#1456