Bump compat for OrdinaryDiffEq v7 / SciMLBase v3 ecosystem

[ChrisRackauckas-Claude]

Summary

Bumps compat bounds to support the OrdinaryDiffEq v7 / SciMLBase v3 ecosystem (refs SciML/OrdinaryDiffEq.jl#3562 and SciML/OrdinaryDiffEq.jl#3565, see NEWS.md).

Compat changes (Project.toml)

Package	Before	After
SciMLBase	"1.8.3, 2"	"1.8.3, 2, 3"
OrdinaryDiffEqTsit5	(unset)	"1, 2"
Catalyst	"15"	"15, 16"
ModelingToolkit	"9"	"9, 10, 11"
SymbolicUtils	"3"	"3, 4"
Symbolics	"6"	"6, 7"

Source migration (src/utils.jl)

RecursiveArrayTools v4 (bundled with SciMLBase v3) makes AbstractVectorOfArray <: AbstractArray, changing length and [] indexing semantics:

• length(sol.u[1]) → length(sol.u[1].u) in get_raw_moments, get_central_moments, get_cumulants — length of an ODESolution now returns total elements (n_vars * n_timesteps), not number of timesteps.
• sol[t_pt] → sol.u[t_pt] in get_moments_FSP — integer indexing into an ODESolution now returns a scalar element, not the t_pt-th timestep.

Both fixes use the sol.u[i] form that works on both RAT v3 and v4 (documented in NEWS.md as the forward-compatible migration path).

Test plan

• CI passes on the standard test suite
• sample_moment_tests.jl exercises get_raw_moments, get_central_moments, get_cumulants
• fsp_tests.jl exercises get_moments_FSP

🤖 Generated with Claude Code

[ChrisRackauckas-Claude]

Changes overview

Project.toml compat bumps:

• SciMLBase: "1.8.3, 2" → "1.8.3, 2, 3" — SciMLBase v3 was released along with OrdinaryDiffEq v7 (Cascade major version bump across all sublibs OrdinaryDiffEq.jl#3562 / Revert OrdinaryDiffEqCore and DiffEqDevTools major bumps from #3562 OrdinaryDiffEq.jl#3565)
• OrdinaryDiffEqTsit5: added "1, 2" — OrdinaryDiffEqTsit5 v2 (test dependency) requires SciMLBase v3
• Catalyst: "15" → "15, 16" — v16 released
• ModelingToolkit: "9" → "9, 10, 11" — v11 supports SciMLBase v3
• SymbolicUtils: "3" → "3, 4" — v4 released
• Symbolics: "6" → "6, 7" — v7 released

src/utils.jl code migration:

RecursiveArrayTools v4 (ships with SciMLBase v3) makes AbstractVectorOfArray <: AbstractArray. This changes length and integer indexing semantics:

1. length(sol.u[1]) → length(sol.u[1].u) in get_raw_moments, get_central_moments, get_cumulants:
   sol.u[1] is an ODESolution; length() now returns total elements (n_vars × n_timesteps) not just n_timesteps.

2. sol[t_pt] → sol.u[t_pt] in get_moments_FSP:
   Integer indexing now returns a scalar element, not the t_pt-th timestep.

Note: the runic CI check failure is a pre-existing infrastructure issue with the self-hosted runner (deepsea3-1) not having Julia installed. Local Runic.format_string confirms utils.jl passes formatting.

[ChrisRackauckas-Claude]

Iteration update: The first CI cycle revealed that bumping ModelingToolkit = "9, 10, 11" allowed Pkg to resolve to MTK v11, which removed get_noiseeqs (renamed to get_noise_eqs) and refactored SDESystem. Migrating to that API is out of scope for an OrdinaryDiffEq v7 / SciMLBase v3 ecosystem compat bump, so I reverted the MTK / Catalyst / SymbolicUtils / Symbolics bounds back to their original values (commit ca8a802).

The remaining changes — SciMLBase = "1.8.3, 2, 3", OrdinaryDiffEqTsit5 = "1, 2" in [compat], and the RecursiveArrayTools v4 sol.u indexing migration in src/utils.jl — are kept. These are forward-compatible with the existing dependency graph (Catalyst v15 still constrains SciMLBase to v2 transitively), and ready to allow SciMLBase v3 once Catalyst gets its own bump.

The Documentation build is now passing.

[ChrisRackauckas-Claude]

CI status (cycle 2):

• ✅ Julia LTS (1.10): pass (8m22s)
• ✅ Documentation build: pass (4m44s)
• ✅ runic format-check: pass
• ❌ Julia 1 (1.12): fail — unrelated to this PR
• ⏳ Julia pre: in progress

The Julia 1 failure is in JumpProcessesOrdinaryDiffEqCoreExt precompilation:

ERROR: LoadError: UndefVarError: `StochasticDiffEqAlgorithm` not defined in `JumpProcessesOrdinaryDiffEqCoreExt`

This is an upstream incompatibility between JumpProcesses' OrdinaryDiffEqCore extension and a newer OrdinaryDiffEqCore that no longer exports StochasticDiffEqAlgorithm/StochasticDiffEqRODEAlgorithm. It happens during precompilation of test deps before any MomentClosure code runs, and the same issue would reproduce on main against the current dependency graph. It is not caused by the compat or src/utils.jl changes in this PR (which only touch SciMLBase + RecursiveArrayTools API surface).