Fix NonStiffODE under v7 stack: manifest refresh + script bugfixes

[ChrisRackauckas-Claude]

Root cause

The v7 stack run on amdci1-1 (CI job 75199969438, run 25374549824) failed in EnrightPryce_wpd.jmd because the pinned Manifest.toml had SymbolicUtils v4.26.0, which throws Method overwriting is not permitted during Module precompilation in apply_optimization_rules. That killed the ParameterizedFunctions precompile (the very first using line in EnrightPryce_wpd.jmd). The cascade of "no such directory" errors against the next six .jmd files in the same log was just the runner cleanup tripping over the killed job — not real per-file failures.

While reproducing locally on a fresh manifest, three additional bugs surfaced that the older pinned stack had been masking:

1. enright_pryce.jl SF1 typo. Equation 2 of the SF1 system was 1880 * [y[4] - y[2] * (1 + k)] (vector literal [...]) instead of (...). New SymbolicUtils.AddWorkerBuffer correctly throws throw_unequal_shape_error when adding a scalar symbolic to a 1-vector symbolic.

2. enright_pryce.jl stray ε parameter. The ND Kepler family declared @parameters ε and threaded [ε => e] through make_ds(...), but ε doesn't appear in the equations themselves (the eccentricity enters via initial conditions only). MTK now errors: Expected an Initial parameter to exist for variable ε, but did not find one. Removed.

3. IRKGL16 (IRKGaussLegendre v1) treats verbose as a Bool, but the v7 stack hands it a DEVerbosity struct. Every WorkPrecisionSet chunk containing IRKGL16(...) raised TypeError: non-boolean ... DEVerbosity ... used in boolean context. Added verbose = SciMLLogging.None() to those chunks in LotkaVolterra / Pleiades / RigidBody / ThreeBody.

4. OrdinaryDiffEqSIMDRK (MER5v2/MER6v2/RK6v4) StackOverflows on NaNMath.sqrt(::VectorizationBase.Vec{2,Float64}) whenever the RHS calls sqrt — which is exactly the ND Kepler family. Split out a setups_nd without SIMD-RK for the ND section of EnrightPryce_wpd.jmd. NA / NB / NC either don't use sqrt or already skip the affected problem, so SIMD-RK stays in setups for them.

Also: EnrightPryce_wpd.jmd referenced NC_PROBLEMS[5], but nc5prob is intentionally commented out in enright_pryce.jl (TODO: nc5sys — complex N-body system needs special handling for MTK v10). Dropped the NC5 chunk in the .jmd until that's restored.

Fix

• Pkg.update() on benchmarks/NonStiffODE/ (no Project.toml / [compat] changes). Notable bumps:
  • SymbolicUtils 4.26.0 -> 4.30.1 (root cause)
  • ModelingToolkit 11.24.1 -> 11.26.3
  • SciMLBase 3.7.1 -> 3.13.0
  • LinearSolve 3.75 -> 3.80
  • OrdinaryDiffEqCore / BDF / Default / Rosenbrock / SDIRK / Verner / FIRK / LowOrderRK / Extrapolation: 2.0.0 -> 2.x.y
  • Symbolics 7.21 -> 7.24, DiffEqBase 7.1 -> 7.5, Sundials 6.1 -> 6.2, plus the usual smaller transitives
• enright_pryce.jl: fix SF1 [...] -> (...) typo; drop stray @parameters ε.
• EnrightPryce_wpd.jmd: drop the NC5 chunk; introduce setups_nd (no SIMD-RK) for the ND section.
• LotkaVolterra_wpd.jmd / Pleiades_wpd.jmd / RigidBody_wpd.jmd / ThreeBody_wpd.jmd: add verbose = SciMLLogging.None() to every WorkPrecisionSet chunk that includes IRKGL16(...).

Local verification

Julia 1.11.9, --project=benchmarks/NonStiffODE:

• Pkg.update() clean; all packages precompile.
• SciMLBenchmarks.weave_file(..., (:script,)) tangles cleanly for all 7 .jmd files (EnrightPryce, FitzhughNagumo, linear, LotkaVolterra, Pleiades, RigidBody, ThreeBody).
• EnrightPryce smoke run (numruns=3, abstols 1e-6..1e-8, reltols 1e-3..1e-5):
  • NA1, NA2, NA4, NA5 — pass with full setups (incl. SIMD-RK)
  • NB1, NB2, NB3, NB5 — pass with full setups
  • NC1, NC2, NC3, NC4 — pass with full setups
  • ND1, ND2, ND3, ND4, ND5 — pass with setups_nd (Tsit5/Vern6/Vern7/Vern9)
• LotkaVolterra full (:github,) weave: previously raised TypeError in the IRKGL16 chunk. After the verbose patch, a direct IRKGL16 + verbose=SciMLLogging.None() WPSet call returns clean.

Test plan

• Wait for the benchmarks/NonStiffODE job on amdci1-1 (or equivalent) to go green under the v7 stack.

Please ignore until reviewed by @ChrisRackauckas.

[ChrisRackauckas]

CI failure diagnosis and fix (run 26699704480, job 80471466039)

The previous run got through EnrightPryce_wpd.jmd cleanly, then failed 6 of the 7 files — FitzhughNagumo, LotkaVolterra, Pleiades, RigidBody, ThreeBody, linear — all with the same error class, always in the chunk whose WorkPrecisionSet contains lsoda():

ERROR: TypeError: non-boolean (DiffEqBase.DEVerbosity{...}) used in boolean context
ERROR: TypeError: non-boolean (SciMLLogging.None) used in boolean context
  ...
  [1] __solve(prob::..., alg::LSODA.lsoda, ...)
    @ LSODA ~/.julia/packages/LSODA/dEJVs/src/common.jl:41

Root cause

LSODA.jl v1.0.0 (latest release, master identical) does a plain Bool test if verbose in its __solve. Under the v7 stack the verbose kwarg that reaches it is never a Bool:

• in chunks without a verbose kwarg, DiffEqBase 7.5.0's default DEVerbosity() struct flows through (FitzhughNagumo, RigidBody, linear);
• in the chunks where this PR added verbose = SciMLLogging.None() for IRKGL16, that None() flows through to LSODA too (LotkaVolterra, Pleiades, ThreeBody).

A chunk-level verbose = false is not an option either — DiffEqBase 7.5.0 throws ArgumentError: Passing a Bool for verbose is no longer supported in OrdinaryDiffEq v7 for every OrdinaryDiffEq/ODEInterface/IRKGL16 solver in the same set. LSODA wants a Bool; everything else refuses one.

Fix (commit a21a93d)

Use DiffEqDevTools' per-setup kwarg mechanism: setup-dict keys in DiffEqBase.allowedkeywords are splatted after the chunk-level kwargs and override them per solver. In all 6 files:

Dict(:alg=>lsoda(), :verbose=>false) # LSODA.jl requires a Bool verbose

Verified locally against the pinned manifest (Julia 1.11):

• solve(prob, lsoda()) and lsoda + verbose=None() reproduce the exact CI TypeError; solve(prob, lsoda(); verbose=false) succeeds.
• WorkPrecisionSet with Dict(:alg=>lsoda(), :verbose=>false) passes both with chunk-level verbose=SciMLLogging.None() and with the default verbosity.
• ddeabm/odex/dop853/ARKODE/CVODE_Adams (the solvers ordered after lsoda in these chunks, some never reached in CI) all pass with the default DEVerbosity.

Long-term, LSODA.jl should accept the v7 verbosity types upstream; this keeps lsoda in the benchmark in the meantime.

Merge with master (commit b94b815)

While this was in flight, master moved (#1583/#1584/#1586/#1599) and the PR became CONFLICTING, which is why the first push (a21a93d) produced no CI run — GitHub skips pull_request runs when no merge ref exists. Merged upstream/master in; the only conflict was enright_pryce.jl (this PR's removal of the stray ε parameter vs. master's Runic reformat). Resolved keeping both: Runic style, no ε, SF1 (...) typo fix intact. The merged enright_pryce.jl loads cleanly under the pinned manifest (all *_PROBLEMS arrays build).

Status

Benchmarks workflow retriggered on b94b815: https://github.com/SciML/SciMLBenchmarks.jl/actions/runs/27268572622

🤖 Generated with Claude Code