Handle despecialized parameter cotangents in solution AD

[ChrisRackauckas-Claude]

Ignore this PR until it has been reviewed by @ChrisRackauckas.

What changed and why

Unwrap a ChainRulesCore.Tangent{DespecializedParameters} before the observable ODESolution pullback remakes its problem. The tangent's params component is the cotangent for the original parameter object; passing the outer structural tangent into remake instead reaches ModelingToolkit initialization as if it were a primal parameter value.

This is a three-line extension fix. It adds no public API, dependency, documentation, or version change.

Root cause

ModelingToolkit commit SciML/ModelingToolkit.jl@5585668 from SciML/ModelingToolkit.jl#4919 changed generated problems from AutoSpecialize to AutoDespecialize. The exact registered boundary is ModelingToolkitBase 1.65.0 to 1.66.0, as independently established in #1545.

The observable pullback consequently receives a structural tangent around DespecializedParameters. SciMLBase's rrule passed that wrapper directly to remake, leading to:

MethodError: no method matching promote_type_with_nothing(
    ::Type{Float64},
    ::ChainRulesCore.Tangent{SciMLBase.DespecializedParameters,
        @NamedTuple{params::Vector{Float64}}})

The ModelingToolkit error is downstream fallout: the invalid primal value is created at the SciMLBase rrule/remake boundary, so the fix belongs here.

Failing before

On clean SciMLBase master 3d95609f7dc1c86a8fded41b9d357bcf25b83ad3, Julia 1.10.12:

GROUP=DownstreamAD julia +1.10.12 --threads=11 --project=. -e 'using Pkg; Pkg.test()'

Test Summary:   | Pass  Error  Broken  Total      Time
Autodiff Remake |    4      1       1      6  32m02.4s
ERROR: Some tests did not pass: 4 passed, 0 failed, 1 errored, 1 broken.

The hosted Julia LTS lane has the same method, argument types, and stack through ODESolution_getindex_pullback: https://github.com/SciML/SciMLBase.jl/actions/runs/33447000990/job/99668297921.

Passing after

The same owning testset passes with the fix:

Test Summary:   | Pass  Broken  Total     Time
Autodiff Remake |    7       1      8  7m14.0s

QA passes with the system Python selected to avoid a broken generated CondaPkg environment on this host:

JULIA_CONDAPKG_BACKEND=Null JULIA_PYTHONCALL_EXE=/usr/bin/python3 \
  GROUP=QA julia +1.10.12 --threads=11 --project=. -e 'using Pkg; Pkg.test()'

Test Summary: | Pass  Total     Time
QA            |  121    121  1m34.4s
Testing SciMLBase tests passed

Runic, typos, and git diff --check pass for the changed file.

Independent failures / not verified

The full local DownstreamAD group proceeds past Autodiff Remake, then fails in the pre-existing observables_autodiff.jl failures: three Mooncake _copy_output NamedTuple type assertions, plus two AutoZygote assertions that still assume the pre-AutoDespecialize parameter shape. Those failures do not execute the changed rrule path. The independent AutoZygote fixture assumptions are being handled separately.

The hosted Julia 1 job https://github.com/SciML/SciMLBase.jl/actions/runs/33447000990/job/99668297994 has only the three Mooncake errors; it does not contain this PR's promote_type_with_nothing error. This PR does not claim to fix that lane.

Docs were not built because this PR changes neither public API nor documentation. GPU paths were not run.

🤖 Generated with Claude Code
https://chatgpt.com/codex/tasks/01a03a17-ad6f-7131-82fc-d0fd57ea6512