DiffEqBaseEnzymeExt: port NS#936/937 fixes (caches walk + alias preserve)

[ChrisRackauckas-Claude]

Please ignore this PR until reviewed by @ChrisRackauckas.

Summary

The Enzyme reverse rule for DiffEqBase.solve_up in lib/DiffEqBase/ext/DiffEqBaseEnzymeExt.jl had the same two latent bugs that NonlinearSolveBaseEnzymeExt accumulated over the past year on the steady-state / NLLS adjoint path. Both are silent gradient drops, and both fire on the ODE / DDE adjoint as soon as prob.p is an MTKParameters carrying non-empty caches (anything wired to MTK's explicitfuns! SCC coupling or paraminit-driven init).

This PR is a direct port of the two NonlinearSolveBase fixes:

• NS#936 (caches walk): replace the bare ptr.dval .+= darg accumulation in the reverse rule with _accum_tangent!(ptr.dval, darg; diff_tunables). The new helper:

  • Goes through SciMLStructures.canonicalize(Tunable(), …) / replace! when the shadow is a SciMLStructure, so the Tunable slice is accumulated correctly instead of falling back to an undefined element-wise iteration.
  • When diff_tunables = false, also walks each non-Tunable field (caches, initials, discrete, constant, nonnumeric, …) of a structured darg. This is exactly what SciMLSensitivity.adjointbackpass returns whenever automatic_sensealg_choice picks diff_tunables = Val(false) (i.e. whenever prob.p is a SciMLStructure with non-empty caches).
  • The diff_tunables value is derived inside the reverse rule the same way NS#936 derives it: honor sensealg.diff_tunables when the user passed an explicit sensealg, else mirror automatic_sensealg_choice's predicate on prob.p.

• NS#937 (alias preserve): replace Enzyme.make_zero(res[1])::RT in the augmented primal with _make_solution_zero(res[1])::RT. The new helper pre-seeds the make_zero IdDict so sol.prob.p and sol.prob.u0 keep aliasing the outer Enzyme shadows. Without this, recursive make_zero allocates fresh buffers for those fields and any cotangent written into the returned sol.prob.p by a downstream consumer goes into a dangling buffer instead of the buffer the outer tape is tracking.

The helpers are lifted verbatim from NonlinearSolveBaseEnzymeExt (same kwarg name diff_tunables, same predicate, same _accum_nested! dispatch tree, same _preseed_alias! guard). Future fixes to one extension should transplant cleanly to the other.

Reference fixes:

• NonlinearSolve.jl/lib/NonlinearSolveBase/ext/NonlinearSolveBaseEnzymeExt.jl:25-89 (_accum_tangent!)
• NonlinearSolve.jl/lib/NonlinearSolveBase/ext/NonlinearSolveBaseEnzymeExt.jl:140-153 (_make_solution_zero)

automatic_sensealg_choice predicate the kwarg derivation mirrors lives in SciMLSensitivity/src/concrete_solve.jl:136-336.

Test plan

New regression tests in lib/DiffEqBase/test/:

• enzyme_accum_tangent.jl — MockMTKParams regression. With the old .+= path the caches stay zero; with the new accumulator under diff_tunables = false the caches accumulate and += repeated-call semantics hold. With diff_tunables = true (default), only Tunable is touched. 6/6 pass.
• enzyme_make_solution_zero.jl — builds an ODESolution and verifies dsol.prob.p === sol.prob.p, dsol.prob.u0 === sol.prob.u0, while dsol.u is still a fresh zero buffer. Naive Enzyme.make_zero fails the aliasing check (regression guard). Also covers the NullParameters / nothing path of the pre-seed helper. 9/9 pass.
• Both testsets are wired into the Core group of lib/DiffEqBase/test/runtests.jl (gated on isempty(VERSION.prerelease), matching the extension's own gate).
• Full Pkg.test("DiffEqBase") on Julia 1.10 — all pre-existing tests still pass alongside the new ones.

Notes

• End-to-end Enzyme.gradient on a real ODE with MTKParameters + non-empty caches is best verified against the desauty-equivalent reproducer post-merge — that requires a much heavier MTK / SciMLSensitivity setup than is appropriate for this PR's unit-test surface.

🤖 Generated with Claude Code

[ChrisRackauckas-Claude]

Closing as not required. Ablation study with the desauty SCC init test rewritten to use the proper Enzyme API (Enzyme.autodiff + Duplicated(iprob), SciML/SciMLSensitivity.jl#1454) confirms this port doesn't change the outcome — desauty still passes with this PR reverted (DiffEqBase pinned to registry v7.5.1). The bug pattern this PR mirrors (NS#936/#937) is real, but neither half is load-bearing on the ODE adjoint path when downstream callers follow the documented Enzyme contract. Can be reopened if a concrete test case demonstrates the need.