NonlinearSolve downstream IntegrationTest intermittently fails in dynamic optimization on Julia 1.12

[ChrisRackauckas-Claude]

NonlinearSolve.jl master downstream IntegrationTest has repeated ModelingToolkit failures on Julia 1.12 in the Optimization/dynamic optimization path. This was first noticed while checking SciML/NonlinearSolve.jl#1012, but it also occurs on the immediately preceding NonlinearSolve master commit, so it does not appear to be caused by that PR.

CI evidence:

• SciML/NonlinearSolve.jl IntegrationTest run 28916278487 on master commit 978a718a45673ede2950f59631cead4c0fe6eac8, created 2026-07-08T03:55:06Z, failed in ModelingToolkit jobs: https://github.com/SciML/NonlinearSolve.jl/actions/runs/28916278487
• PR run 28928416097 for NonlinearSolve.jl#1012 showed the same InterfaceII failure symptoms: https://github.com/SciML/NonlinearSolve.jl/actions/runs/28928416097

Representative failures from the pre-#1012 master run logs:

• InfiniteInterpolations extension precompile/load error:
  • ArgumentError: Package Interpolations [a98d9a8b-a2ab-59e6-89dd-64a1c18fca59] is required but does not seem to be installed
  • source path: ~/.julia/packages/InfiniteOpt/aM7Wg/ext/InfiniteInterpolations.jl:5
  • seen in InterfaceI, InterfaceII, and Initialization logs while running the ModelingToolkit downstream job.
• Dynamic optimization CasADi error:
  • TypeError: non-boolean (CasADi.MX) used in boolean context
  • stack includes Base.isequal(x::CasADi.MX, y::Int64), SymbolicUtils ... hashconsing.jl, WeakCacheSets ... WeakCacheSets.jl, and ModelingToolkitBase/src/systems/optimal_control_interface.jl:471 in add_cost_function!.
  • failing test locations varied by shard, e.g. lib/ModelingToolkitBase/test/optimization/dynamic_optimization.jl:434, :607.

Resolved versions in the failing CI logs included:

• Julia 1.12.6
• ModelingToolkit v11.31.0 from checkout, ModelingToolkitBase v1.51.0 from lib/ModelingToolkitBase
• SymbolicUtils v4.38.1
• WeakCacheSets v0.1.0
• CasADi v1.2.0
• InfiniteOpt v0.6.3
• DataInterpolations v8.10.0
• Interpolations v0.16.3 was installed/precompiled earlier in the job, but the InfiniteOpt extension later reported it as missing.

Local reproduction attempt that did not reproduce:

# separate clean checkout, NonlinearSolve pre-#1012 master commit 978a718a
# ModelingToolkit checkout at 3239ad1184c38d26e62ce8f0e80b1c623e22574d
env JULIA_DEPOT_PATH=/home/crackauc/sandbox/tmp_20260708_073226_7642/julia-depot-repro \
    JULIA_PKG_SERVER_REGISTRY_PREFERENCE=eager \
    GROUP=Optimization \
    timeout 3600 \
    /home/crackauc/.julia/juliaup/julia-1.12.6+0.x64.linux.gnu/bin/julia \
    --color=yes --project=downstream \
    -e 'using Pkg; Pkg.develop(PackageSpec(path=".")); Pkg.update(); Pkg.test(coverage=false)'

Local result:

Test Summary: | Pass Broken Total Time
Optimization  | 156      1   157 19m31.8s
Testing ModelingToolkit tests passed

So this may be CI-environment/order/cache sensitive rather than a straightforward deterministic failure from the NonlinearSolve source tree. A plain NonlinearSolve commit bisect is not valid because the same NonlinearSolve commit can resolve/run against changing downstream package state, and the targeted local reproduction did not fail even with the versions listed above.

What would help next:

• Check whether the downstream CI job is running a different group/order than GROUP=Optimization despite the dynamic optimization failure appearing under InterfaceI/InterfaceII labels.
• Capture or preserve the exact temp test environment manifest from a failing CI job.
• Investigate why InfiniteOpt's InfiniteInterpolations extension can be requested when Interpolations is not visible in the active environment.
• Investigate whether SymbolicUtils/hashconsing should guard isequal on CasADi.MX values whose equality returns a symbolic MX rather than Bool.