Unwrap FunctionWrappersWrapper in DAE init closures

[ChrisRackauckas-Claude]

Summary

• Unwrap FunctionWrappersWrapper from f at the top of each ODEProblem _initialize_dae! method via SciMLBase.unwrapped_f(f)

Problem

DiffEqBase #1284 (v6.208.0) enabled AutoSpecialize for mass matrix ODEProblems by removing the f.mass_matrix isa UniformScaling guard from promote_f. This wraps f.f in a FunctionWrappersWrapper with 4 pre-compiled variants coordinated with the ODE solver's ForwardDiff (tag = OrdinaryDiffEqTag, chunk size = 1).

However, _initialize_dae! (both BrownFullBasicInit and ShampineCollocationInit) passes f into closures that NonlinearSolve differentiates through with its own ForwardDiff. NonlinearSolve's duals have a different tag and chunk size, so none of the 4 pre-compiled variants match, causing NoFunctionWrapperFoundError.

This breaks the OrdinaryDiffEqRosenbrock DAE AD tests (dae_rosenbrock_ad_tests.jl) on all Julia versions.

Fix

Call SciMLBase.unwrapped_f(f) after destructuring (; p, t, f) = integrator in the 4 ODEProblem methods of _initialize_dae!. This replaces the FunctionWrappersWrapper with the raw function for DAE initialization only. The solver's hot path in perform_step! still uses the wrapped version from integrator.f.

This is a no-op when AutoSpecialize is not active (i.e., when f.f is not a FunctionWrappersWrapper).

Test plan

• OrdinaryDiffEqRosenbrock tests pass (specifically dae_rosenbrock_ad_tests.jl with AutoForwardDiff)
• No regression in other subpackage tests

🤖 Generated with Claude Code

[ChrisRackauckas-Claude]

Update: Added forwarddiff_chunksize trait implementation

The original PR fixed the DAE initialization issue (NonlinearSolve tag mismatch). This update adds the second piece of the fix: implementing the forwarddiff_chunksize trait to resolve the FunctionWrapper chunk size mismatch.

New commit: Implement forwarddiff_chunksize trait

Changes:

• lib/OrdinaryDiffEqCore/src/alg_utils.jl: Implement SciMLBase.forwarddiff_chunksize for all algorithm types that use ForwardDiff (implicit, DAE, exponential, composite)
• lib/OrdinaryDiffEqCore/Project.toml: Bump SciMLBase compat to 2.143, DiffEqBase compat to 6.209
• Project.toml: Bump SciMLBase and DiffEqBase compat
• lib/OrdinaryDiffEqRosenbrock/test/dae_rosenbrock_ad_tests.jl: Fix @test_broken → plain assignment (now inferable)

How it works

1. SciMLBase (Add forwarddiff_chunksize algorithm trait SciMLBase.jl#1244): New forwarddiff_chunksize(alg) trait (defaults to 0)
2. DiffEqBase (Thread ForwardDiff chunk size through FunctionWrapper compilation DiffEqBase.jl#1287): Threads chunk size through promote_f → wrapfun_iip, compiles FunctionWrapper variants with chunk=CS for Jacobian and chunk=1 for time derivative
3. OrdinaryDiffEq (this PR): Implements the trait, returning the algorithm's ForwardDiff chunk size

Test results

All tests pass locally:

• DAE Rosenbrock AD Tests: 8/8 pass
• Rosenbrock Convergence Tests: 853/853 pass

[ChrisRackauckas-Claude]

Pushed two additional commits:

1. e0b3bbb - Implement forwarddiff_chunksize trait for all OrdinaryDiffEq algorithm types (was committed locally but not pushed previously)
2. d4e7dc5 - Simplify to use existing _get_fwd_chunksize(AD) which already returns Val{CS}(), matching the updated trait contract from SciMLBase#1244. Consolidates three methods into one Union dispatch, consistent with the existing get_chunksize pattern.

Depends on SciMLBase#1244 and DiffEqBase#1287.