Add MixedDuplicated(::T, ::T) shim for Enzyme#3126 workaround

[ChrisRackauckas-Claude]

Please ignore until reviewed by @ChrisRackauckas.

Summary

Enzyme's runtime_generic_augfwd → create_activity_wrapper (Enzyme/src/rules/jitrules.jl line 14) constructs MixedDuplicated(primarg, shadowarg) with shadowarg::T rather than Base.RefValue{T}. EnzymeCore.MixedDuplicated only defines MixedDuplicated(::T, ::Base.RefValue{T}), so the bare-T call raises a MethodError. This blocks reverse-mode AD chains that capture ODESolution (e.g. the MTK parameter-init path) — exactly the residual blocker tracked in EnzymeAD/Enzyme.jl#3126.

Until #3126 lands upstream, this PR adds a workaround:

• A type-piracy shim in src/enzyme_rules.jl that intercepts MixedDuplicated(x::T, dx::T, check::Bool=true) and delegates to the existing constructor by wrapping the shadow in Base.RefValue. The constructor's check::Bool flag is currently unused by EnzymeCore, so the shim is invariant-preserving.
• Flips two @test_broken blocks blocked specifically on this MethodError back to @test:
  • test/parameter_initialization.jl: "Adjoint through Prob (Enzyme)"
  • test/mtk.jl: MTK Forward Mode Enzyme block (line 204)
• Mooncake @test_broken in test/mtk.jl (line 230) is unrelated and left alone.

Risk acknowledgment (type piracy)

This is type piracy on EnzymeCore. Concerns:

• The shim adds a method to EnzymeCore.MixedDuplicated that fires for any (::T, ::T) call site. If Enzyme (or downstream) ever calls MixedDuplicated(::T, ::T) with a legitimate intent (e.g. a T = Base.RefValue{S} primal whose shadow really is a bare RefValue{S}), this shim silently double-wraps and the result is wrong. I have not found such a call site in the current Enzyme tree — every other MixedDuplicated construction either explicitly wraps in Base.RefValue or goes through the broken create_activity_wrapper path — but the surface is wide.
• The remedy if it bites: remove this shim and the two @test_broken flips, return to @test_broken, and wait for Enzyme#3126 to land upstream.

CI will surface any breakage of other Enzyme code paths via the full test matrix.

Verification

I attempted to run Pkg.test() on Julia 1.12 with GROUP=Core8 locally. SciMLSensitivity precompiles cleanly with the shim. The actual parameter_initialization.jl test entered the Enzyme reverse-mode compilation phase and was still grinding through code generation past 25 minutes elapsed (with 4+ GB RSS) when I had to push for CI verification. I did NOT see a final pass/fail locally. CI here is the authoritative verification — please review CI results before merging.

Removal plan

When EnzymeAD/Enzyme.jl#3126 lands upstream, delete:

• The MixedDuplicated(x::T, dx::T, ...) method in src/enzyme_rules.jl
• The import Enzyme.EnzymeCore: MixedDuplicated line

Tests should continue to pass on the fixed Enzyme version.

🤖 Generated with Claude Code

[wsmoses]

This is absolutely wrong and type piracy and should not be done

[ChrisRackauckas]

I know it's just testing CI to see if it would make the test pass.