NonlinearSolveBase: unwrap FunctionWrappersWrapper on the EnzymeOriginator adjoint path (MTK DAE init)

[ChrisRackauckas-Claude]

Please ignore until reviewed by @ChrisRackauckas.

Summary

Follow-up to #944 (which fixed the solve_up Enzyme rule plumbing). With that landed, Enzyme reverse-mode through an MTK DAE solve gets into the initialization NonlinearProblem solve and then fails with:

EnzymeRuntimeException: Enzyme execution failed.
Expected return type of primal to be SciMLBase.NonlinearSolution{...} but did not find a value of that type
  get_initial_values  (the OverrideInit init solve, under Enzyme reverse)

Root cause — a FunctionWrappersWrapper type-instability:

• The custom solve_up rule re-runs the solve via _solve_adjoint → get_concrete_problem, which maybe_wrap_nonlinear_f-wraps the IIP function in AutoSpecializeCallable{FunctionWrappersWrapper} (confirmed by instrumenting typeof(res[1])).
• That wrapping is type-unstable. Enzyme's traced forward unwraps it via maybe_unwrap_prob_for_enzyme (Skip FunctionWrappersWrapper in maybe_wrap_nonlinear_f under Enzyme AD #940), so the return type Enzyme infers carries the bare GeneratedFunctionWrapper; the adjoint re-wraps it, so the rule's primal type no longer matches → Enzyme aborts.
• maybe_unwrap_prob_for_enzyme keys off the solver's own autodiff, which for an MTK DAE init is ForwardDiff even when the outer differentiation is Enzyme — so it doesn't fire here.

Fix: key off the originator instead — unwrap _prob.f.f via get_raw_f when originator isa EnzymeOriginator, so the adjoint primal matches the (unwrapped) type Enzyme's traced forward produces. No-op for non-Enzyme originators and for already-unwrapped functions.

Verification status (please read)

• The fix is a 13-line addition; the edit is confirmed sound (the stack loads and forward-solves with retcode = Success).
• Strong evidence it clears the error, but not yet a completed end-to-end gradient check. Pre-fix, the MTK-init Enzyme reverse threw the EnzymeRuntimeException ~21 min into compilation; with the fix, the run reached ~26 min in the full reverse compile with no such error before the process was OOM-killed.
• The completed correct-gradient comparison (vs ForwardDiff) is pending: the test machine is under heavy multi-agent load and repeatedly system-OOM-kills this ~26-min Enzyme compile (oom_kill 77 on the session cgroup). I'll post the gradient comparison once it completes in a less-contended window / via CI.

Companion to SciML/SciMLSensitivity#1463; built on #944 (merged) and DiffEqBase 7.5.4 / OrdinaryDiffEq#3700 (merged).

🤖 Generated with Claude Code