fix: let the despecialization barrier accept unwrapped parameters

[ChrisRackauckas-Claude]

Please ignore this PR until reviewed by @ChrisRackauckas.

What changed and why

Allow the parameter-despecialization barrier to receive zero DespecializedParameters wrappers and pass that call through unchanged. DAE initialization wraps its residual in a NonlinearFunction, which unwraps parameters before forwarding to the outer AutoDespecialize function; the barrier therefore re-enters with zero wrappers and previously threw. More than one wrapper remains an error.

The branch is rebuilt as one commit on current master while preserving the original author and the required Chris Rackauckas co-author trailer. DiffEqBase is bumped from 7.18.0 to 7.18.1.

Failing before / passing after

With the checked-in regression test present and the source fix reverted, the full DiffEqBase Core group fails:

the barrier accepts already-unwrapped parameters: Error During Test
Expression: DiffEqBase._invoke_parameter_despecialization(collect_args, (1, p, 2)) == (1, p, 2)
a parameter-despecialization barrier requires exactly one parameter wrapper
Despecialized-p Hook | Pass 45 Fail 1 Error 1 Total 47
ERROR: Package DiffEqBase errored during testing

With the fix restored, the identical full Core command passes:

Despecialized-p Hook | Pass 47 Total 47
Testing DiffEqBase tests passed

The public DAE reproducer also reaches a successful solve and satisfies its algebraic constraint:

retcode=Success, constraint=1.0

Local verification

• GROUP=Core julia +1.12 --startup-file=no --project=lib/DiffEqBase -e 'using Pkg; Pkg.test()': passed on current master; the discriminating hook is 47/47.
• GROUP=QA julia +1.12 --startup-file=no --project=lib/DiffEqBase -e 'using Pkg; Pkg.test()': Aqua 9/9; package tests passed.
• Standalone public DAE solve using DFBDF, AutoDespecialize, and BrownFullBasicInit: retcode=Success, constraint sum 1.0.
• Runic --check --diff, typos on all changed files, and git diff --check: passed.

The complete Downstream/Downstream2, GPU, and prerelease-Julia groups were not run locally. The public DAE reproducer covers the observed behavior, while the full Core and QA groups cover the changed DiffEqBase package.

[ChrisRackauckas-Claude]

The requested DiffEqBase source change triggered the full matrix. The completed run produced 268 completed CPU checks (236 success, 28 failure, 3 skipped, 1 cancelled); the remaining GPU jobs were queued/running when the branch was rebased. Every red CPU job is now assigned:

• AD: Mooncake tuple tangent — Normalize Mooncake promote_f tuple tangents #4242
• Integrators II: DiffCache resize assertion — Use DiffCache accessor in resize test #4230
• Interface I/downgrade: tstop round-off expectation — Correct derivative discontinuity tstop expectation #4261
• Interface V: JLArrays 0.3.2 host-LAPACK conversion — JLArrays 0.3.2 breaks InterfaceV implicit solve via host LAPACK Ptr conversion #4179 and Define LU factorization for JLArray JuliaGPU/GPUArrays.jl#757
• Regression I: Newton–Krylov tolerance — Floor the Newton direction's linear tolerance at sqrt(eps) #4301
• BDF Core/downgrade — Restore the scalar DAE Jacobian test helper #4309
• Default downgrade — merged Raise OrdinaryDiffEqDefault DiffEqBase floor to 7.15 #4274
• NonlinearSolve downgrade — Raise the NonlinearSolve Core compat floor #4324
• NonlinearSolve matrix-free Core failure — merged Handle matrix-free instability diagnostics #4310
• NonlinearSolve ModelingToolkit conversion — Wrap NLStep parameters for despecialized caches #4325; its separately bisected fixed-step convergence follow-up is still in final local Core validation
• DiffEqBase Downstream invalid positional solver — Preserve invalid positional solver errors after default preparation #4257
• StochasticDiffEqROCK JET — Fix definite assignment in ROCK stages #4321
• DiffEqDevTools rendered docs — merged Render ConvergenceTrajectory in package docs #4312; the remaining Aqua wrapper-precompile race has a validated upstream Aqua patch but needs external-repository permission
• TaylorSeries Aqua wrapper exit — tracked at TaylorSeries QA: Aqua persistent-tasks check fails on CI since 2026-08-03 (wrapper precompile exits, not reproducible locally) #4277; the exact failing tree/environment passes locally, so no unsupported Taylor source change was made
• SciMLSensitivity Julia 1.10: independently reduced to Mooncake array growth plus a separate historical inline-initdt verifier issue; the Ordinary initdt fix is in final AD/QA validation
• ProbNum Julia 1.10 resolution — Use the shared downstream workflow on Julia 1 #4323; the Julia 1.12 matrix reaches the tests and exposes one conditioning-dependent unseeded ProbNum test flake
• PositiveIntegrators — Bump the all-julia-dependencies group across 3 directories with 1 update NumericalMathematics/PositiveIntegrators.jl#216

Full triggered CPU/GPU workflow: https://github.com/SciML/OrdinaryDiffEq.jl/actions/runs/32391249144
Sublibrary workflow: https://github.com/SciML/OrdinaryDiffEq.jl/actions/runs/32391250140
Downgrade-sublibrary workflow: https://github.com/SciML/OrdinaryDiffEq.jl/actions/runs/32391248972
Downstream workflow: https://github.com/SciML/OrdinaryDiffEq.jl/actions/runs/32391247975

[ChrisRackauckas-Claude]

Final OrdinaryDiffEq follow-ups from the matrix are now published: the Julia 1.10 initdt forward-mode verifier fix is #4327 (QA 92/92; Interface I green with #4261; AD 28,302/28,302 with #4242), and the fixed-step multistep convergence fix is the stacked #4328 (Core 891 pass/40 pre-existing broken; ModelingToolkit and QA green). No CPU failure from the triggered matrix remains unassigned.