Box Enzyme solve_up tape closure

[ChrisRackauckas-Claude]

This PR should be ignored until reviewed by @ChrisRackauckas.

Summary

• Box the solve_up Enzyme custom-rule tape in Any[...] so the SciMLSensitivity pullback closure does not become part of the concrete Enzyme tape type.
• Keep the same primal, shadow, and reverse pullback behavior; this only reduces compiler/tape type pressure on the macOS Julia 1.10 Enzyme reverse-mode path.

Context

The latest master root CI failed intermittently in tests / Adjoint (julia lts, macos-latest) with a Julia signal 11 in test/Adjoint/adjoint_tests__item3.jl. Recent runs show the same job/file/line can pass with identical Enzyme/SciMLSensitivity versions, so this targets robustness rather than changing the tested behavior.

Validation

• timeout 3600 env NONLINEARSOLVE_TEST_GROUP=Adjoint ~/.juliaup/bin/julia +1.10 --project=. -e 'using Pkg; Pkg.test(; julia_args=["--check-bounds=auto"])'\n- timeout 3600 ~/.juliaup/bin/julia +1 --project=. -e 'using Pkg; Pkg.test(; julia_args=["--check-bounds=auto"])' from lib/NonlinearSolveBase\n- git diff --check\n- Runic.format_file(...; inplace=true) on the changed file from a temporary Julia environment with Runic v1.7.0\n

[ChrisRackauckas-Claude]

Scratchpad update: the NonlinearSolve root matrix and NonlinearSolveBase sublibrary checks are green for this Enzyme tape fix. The remaining red checks are ModelingToolkit downstream jobs; local/log investigation traced these to SciML/.github#110 (downstream.yml@v1 not exporting the requested group input as GROUP), not to this branch.

[ChrisRackauckas-Claude]

CI update: SciML/.github v1 now points at v1.21.0 / 12371ec79ba3, which includes the downstream GROUP export fix. I pushed empty commit 65f5c53d5 to trigger fresh CI on this PR because direct rerun failed with HTTP 403 for this token. Fresh downstream run started as https://github.com/SciML/NonlinearSolve.jl/actions/runs/29000408648.

[ChrisRackauckas-Claude]

Rebased this branch on current upstream/master and force-pushed with lease; new head is 73f1a10675eec10ba6d3d09deab306dfe5c506ba.

Local verification on the rebased branch:

• timeout 3600 env GROUP=Adjoint /home/crackauc/.juliaup/bin/julia +1 --project=. -e 'using Pkg; Pkg.test(; coverage=false)' passed.
  • Adjoint Tests: 3/3 pass
  • maybe_wrap_nonlinear_f skips wrapping inside Enzyme.autodiff (#939): 3/3 pass
  • Enzyme reverse-mode over IIP NonlinearProblem (#939): 1/1 pass
  • log ended with Testing NonlinearSolve tests passed
• git diff --check upstream/master...HEAD passed.
• Runic check on lib/NonlinearSolveBase/ext/NonlinearSolveBaseEnzymeExt.jl passed.

The stale OrdinaryDiffEq.jl/Interface downstream failure was from SciMLBase.constructorof in OrdinaryDiffEq. Current OrdinaryDiffEq master includes SciML/OrdinaryDiffEq.jl#3846, and I locally ran GROUP=InterfaceI there; it passed, including Interpolation and Cache Stripping Tests 16/16.

[ChrisRackauckas-Claude]

Rebased onto current master (c22bc9482); new head is 6ea947c37.

Local verification on the rebased head:

• GROUP=Adjoint julia +1 --project=. -e 'using Pkg; Pkg.test(; coverage=false)': passed (Adjoint Tests 3/3, Enzyme wrapping 3/3, Enzyme reverse-mode IIP 1/1).
• Runic check on lib/NonlinearSolveBase/ext/NonlinearSolveBaseEnzymeExt.jl: passed.
• git diff --check upstream/master...HEAD: passed.

The prior ModelingToolkit InterfaceII failure is reproduced on unmodified NonlinearSolve master and is unrelated to this diff: registered OrdinaryDiffEqFIRK v2.3.2 with OrdinaryDiffEqNonlinearSolve v2.1.0 raises UndefVarError: FastConvergence. OrdinaryDiffEq source FIRK v2.4.0 fixes it, and ModelingToolkit's full fractional_to_ordinary.jl test passes locally with that source. The remaining ecosystem blocker is stale General registration PR JuliaRegistries/General#160479, which points at an unreachable pre-squash commit and needs re-registration from a reachable OrdinaryDiffEq commit.

[ChrisRackauckas-Claude]

The inherited solf spellcheck regression is addressed separately in draft PR #1031. It bisects to master commit 611cdf895, and #1031 passes exact typos v1.48.0, full Runic, git diff --check, and the full GROUP=Core test (Testing NonlinearSolve tests passed). This PR remains scoped to the Enzyme tape fix.