Replace OrdinaryDiffEq weakdep with OrdinaryDiffEqSymplecticRK

[ChrisRackauckas-Claude]

Note

This PR should be ignored until reviewed by @ChrisRackauckas.

Summary

Replaces the OrdinaryDiffEq weakdep with the smaller OrdinaryDiffEqSymplecticRK sub-package. The DiffEqIntegrator extension only ever used the symplectic solvers (VerletLeapfrog and friends) plus the standard init / step! / DynamicalODEProblem interface — none of which requires the full meta-package.

Why

• Lighter dep graph. Users no longer pull in BDF, Rosenbrock, Tsit5, Verner, Default, OrdinaryDiffEqCore, NonlinearSolve, LinearSolve, etc. just to use a leapfrog inside HMC.
• Honest deps. The extension actually only uses symplectic RK methods.
• Sidesteps the OrdinaryDiffEq v6 ↔ v7 compat ratchet entirely. OrdinaryDiffEqSymplecticRK v1 (built against OrdinaryDiffEq v6) and v2 (built against v7) both expose the same surface, so compat is just "1, 2".
• Aligns with the v7 ecosystem story. In OrdinaryDiffEq v7, VerletLeapfrog no longer re-exports out of OrdinaryDiffEq itself (it lives in the symplectic sublib). Depending on the sublib directly is the forward-compatible shape.
• Unblocks downstream Turing v0.45 + SciML stack composition (e.g. SciMLExpectations.jl/docs), where Turing 0.45 → AdvancedHMC ≥ 0.8.3 and OrdinaryDiffEq 7 cannot currently coexist due to the OrdinaryDiffEq = "6" weakdep compat.

Changes

• Project.toml: weakdep OrdinaryDiffEq → OrdinaryDiffEqSymplecticRK; extension entry, compat ("1, 2"), and [extras] updated.
• ext/AdvancedHMCOrdinaryDiffEqExt.jl → ext/AdvancedHMCOrdinaryDiffEqSymplecticRKExt.jl (rename + module name + import).
• test/runtests.jl, test/integrator.jl: switch using OrdinaryDiffEq to using OrdinaryDiffEqSymplecticRK.
• src/AdvancedHMC.jl: register_error_hint message updated ("Did you forget to load OrdinaryDiffEqSymplecticRK?").
• src/integrator.jl: inline comment updated.

Verification

Tested locally on Julia 1.10:

• AHMC_TEST_GROUP=AdvancedHMC test suite: 5052 tests, exit 0, no errors.
• Integrator testset: 35/35 pass.
• Numerical equivalence check (independent script): DiffEqIntegrator(VerletLeapfrog()) produces bit-identical trajectories to native Leapfrog over 1000 steps; Hamiltonian drift ~1.7e-5 over the same trajectory.

Relation to existing CompatHelper PRs

CompatHelper opened #494 and #495 with the simpler compat-only bump on the OrdinaryDiffEq weakdep ("6" → "6, 7"). That approach works but keeps the whole meta-package as a dep, and breaks the existing test/integrator.jl test which references OrdinaryDiffEq.VerletLeapfrog (in v7 that name moved to OrdinaryDiffEqSymplecticRK). This PR avoids both problems by depending on the sublib directly. Closing #494 and #495 in favor of this would be reasonable.

Test plan

• Extension precompiles against both OrdinaryDiffEqSymplecticRK v1 and v2
• Full AdvancedHMC test group passes locally on Julia 1.10
• DiffEqIntegrator(VerletLeapfrog()) matches native Leapfrog numerically
• CI: GitHub Actions (Julia 1, min, pre × ubuntu/macOS/windows)
• CI: buildkite CUDA tests

🤖 Generated with Claude Code

[ChrisRackauckas]

This version I think is better, just moving to the smaller dep that you actually use. I think this is good to go.