Fix mass-matrix Rosenbrock test and raise downgrade compat floors (#462)

The CPU/JLArrays/OpenCL/QA test groups all errored on the first test,
test/gpu_kernel_de/stiff_ode/gpu_ode_mass_matrix.jl, with a
`Cannot convert StaticWOperator{Float64} to StaticWOperator{Float32}`
MethodError. The reference CPU bench solve used `dt = 0.1` (Float64) on
an otherwise all-Float32 problem (u0, p, tspan, tolerances are Float32,
and the GPU solve already uses `dt = 0.1f0`). OrdinaryDiffEqRosenbrock's
JacReuseState types its `cached_W` slot from the state eltype (Float32),
but with a Float64 dt `calc_W` builds a Float64 StaticWOperator that
cannot be stored, erroring at OrdinaryDiffEqDifferentiation
derivative_utils.jl:922 (`jac_reuse.cached_W = W`). Use a type-consistent
`dt = 0.1f0` for the bench solve; the GPU-vs-reference tolerance
assertions are unchanged. Verified locally on Julia 1.12 with GROUP=CPU:
the mass-matrix safetestset passes 2/2.

Also raise three compat floors so the Downgrade job can get past the
ModelingToolkit unsatisfiability it currently hits (StaticArrays was
pinned to its 1.9 floor at downgrade, which forced MTK below v11 while
the test pins MTK >= 11.17):
  - StaticArrays 1.9 -> 1.9.14 (MTK 10.18-11 needs StaticArrays >= 1.9.14)
  - SciMLBase 3 -> 3.1 (lowest MTK compatible with SciMLBase 3 is 11.22;
    JumpProcesses 9.25.1+ needs SciMLBase >= 3.1.0, excluding 3.0.0)
  - ModelingToolkit 11.17.0 -> 11.23.0 (test; lowest MTK consistent with
    the SciMLBase 3.1 and OrdinaryDiffEqRosenbrock 2 floors)
Verified the floor set resolves on Julia 1.10 (SciMLBase 3.1.0 / MTK
11.23.0 / StaticArrays 1.9.14 / OrdinaryDiffEqRosenbrock 2.0.0) and that
normal latest-version resolution is unaffected (SciMLBase 3.21, MTK
11.28, StaticArrays 1.9.18). The Downgrade job may still hit the
pre-existing, maintainer-noted ForwardDiff/LogExpFunctions strict-floor
wall, which is an upstream issue tracked separately.

Co-authored-by: ChrisRackauckas-Claude <accounts@chrisrackauckas.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: 3 people

Project.toml

@@ -59,11 +59,11 @@ MuladdMacro = "0.2"
 OpenCL = "0.10"
 Parameters = "0.12, 0.13"
 RecursiveArrayTools = "4"
-SciMLBase = "3"
+SciMLBase = "3.1"
 Setfield = "1"
 SimpleDiffEq = "1.11"
 SimpleNonlinearSolve = "2"
-StaticArrays = "1.9"
+StaticArrays = "1.9.14"
 TOML = "1"
 ZygoteRules = "0.2.7"
 julia = "1.10"

test/Project.toml

@@ -41,7 +41,7 @@ ForwardDiff = "1"
 GPUArraysCore = "0.2"
 JLArrays = "0.3"
 KernelAbstractions = "0.9"
-ModelingToolkit = "11.17.0"
+ModelingToolkit = "11.23.0"
 OpenCL = "0.10"
 Optimization = "5"
 OptimizationOptimisers = "0.3"
@@ -50,8 +50,8 @@ OrdinaryDiffEqRosenbrock = "2"
 OrdinaryDiffEqSDIRK = "2"
 OrdinaryDiffEqStabilizedRK = "2"
 SafeTestsets = "0.1"
-SciMLBase = "3"
-StaticArrays = "1.9"
+SciMLBase = "3.1"
+StaticArrays = "1.9.14"
 Statistics = "1"
 StochasticDiffEq = "7"
 TOML = "1"

test/gpu_kernel_de/stiff_ode/gpu_ode_mass_matrix.jl

@@ -35,8 +35,11 @@ monteprob = EnsembleProblem(prob, safetycopy = false)
 
 alg = GPURosenbrock23()
 
+# dt must match the Float32 eltype of u0/p. A Float64 dt makes calc_W build a
+# Float64 StaticWOperator that OrdinaryDiffEqRosenbrock's JacReuseState (typed
+# from the Float32 state) cannot store, erroring in the cached_W assignment.
 bench_sol = solve(
-    prob, Rosenbrock23(), dt = 0.1, abstol = 1.0f-5, reltol = 1.0f-5,
+    prob, Rosenbrock23(), dt = 0.1f0, abstol = 1.0f-5, reltol = 1.0f-5,
     initializealg = BrownFullBasicInit()
 )