QA: track pre-existing Aqua/JET/ExplicitImports findings (root + ModelingToolkitBase)

[ChrisRackauckas-Claude]

The QA lanes for the ModelingToolkit root package and the ModelingToolkitBase sublibrary are currently red, and the conversion to the run_qa v1.6 form (PR forthcoming) surfaces every individual finding. To get the lanes green-or-broken without silencing anything, the conversion marks each finding as @test_broken (via run_qa's aqua_broken / jet_broken / ei_broken). This issue tracks the underlying findings so the broken-markers can be dropped as they are fixed (each auto-flags an Unexpected Pass once resolved).

Root ModelingToolkit

Aqua (aqua_broken):

• undefined_exports — ModelingToolkit.Variable is reexported but not defined in the namespace.
• piracies — deliberate definitions of SciMLBase problem constructors (SCCNonlinearProblem(::System, op), ...) and a few MTKBase/Symbolics methods on upstream types.

JET (jet_broken):

• report_package re-parses the toplevel; the @recompile_invalidations begin using StaticArrays; using DiffEqBase; ... end blocks make it require transitive deps that are not directly resolvable, surfacing as toplevel errors.

ExplicitImports (ei_broken):

• no_implicit_imports / no_stale_explicit_imports — the StructuralTransformations submodule relies on a large block of implicit imports (whole packages: BipartiteGraphs, Graphs, ...) and carries a large block of stale explicit imports; the top-level module is also unanalyzable (dynamic include in src/precompile.jl).
• all_qualified_accesses_are_public / all_explicit_imports_are_public — dominated by names not (yet) declared public in their defining packages (Symbolics, SymbolicUtils, SciMLBase, StateSelection, ModelingToolkitBase, ModelingToolkitTearing, ...). Fix is upstream public declarations.

lib/ModelingToolkitBase

Aqua (aqua_broken):

• ambiguities — method ambiguities.
• unbound_args — unbound type parameters (e.g. _remake_buffer with an unbound P).
• undefined_exports — ModelingToolkitBase.Variable reexported but undefined.
• stale_deps — SimpleNonlinearSolve declared but unused.
• deps_compat — the Optimization test extra lacks a [compat] entry.
• piracies — deliberate search_variables! / toexpr methods on upstream types.

ExplicitImports (ei_broken):

• no_implicit_imports / no_stale_explicit_imports — large implicit-import and stale-import blocks.
• all_qualified_accesses_are_public / all_explicit_imports_are_public — ~300 names not (yet) public in their defining packages.

stale_deps (SimpleNonlinearSolve) and deps_compat (Optimization) are likely quick Project.toml fixes; they were left as broken-markers to keep the QA-conversion PR focused.

[ChrisRackauckas-Claude]

Broken-markers added (and to be removed as findings are fixed) in #4671.

[ChrisRackauckas-Claude]

Clean-master reproduction and boundary audit (2026-07-12):

Current master

I reproduced the root-package Aqua failure on an isolated checkout of exact upstream master commit ee7dd34d8b7cfe8a085e0fa04cec28a607053f66, using Julia 1.12.6, Aqua 0.8.16, and SciMLTesting 1.8.0:

timeout 7200 julia +1.12 --startup-file=no --project=test/qa \
  -e 'using Pkg; Pkg.instantiate(); using ModelingToolkit, Aqua, Test; Aqua.test_piracies(ModelingToolkit)'

Result: exit 1 with exactly 31 methods. The main groups are:

• 4 SCCNonlinearProblem(::System, ...) constructor/keyword methods;
• Sample() and ShiftIndex() constructors;
• 2 __mtkcompile(::System, ...) methods;
• complete(::StateSelection.DiffGraph);
• 8 generate_ODENLStepData(::System, ...) positional/keyword methods;
• 2 get_initialization_problem_type(::System, ::System, ...) methods;
• 8 input_timedomain/output_timedomain methods for Sample and Hold;
• 2 simplify_sde_system(::System, ...) methods;
• singular_check(::TearingState) and torn_system_jacobian_sparsity(::System).

These are cross-package ownership definitions: the methods are defined by ModelingToolkit, while their functions and all dispatched argument types are owned by SciMLBase, ModelingToolkitBase, StateSelection, or ModelingToolkitTearing.

Historical boundaries

There are two distinct boundaries:

1. Code ownership: fb2450ea82af06595598f61ea1bb4ef96c35edd9 (refactor: split out ModelingToolkitBase) moved System, ShiftIndex, and interface functions into ModelingToolkitBase while leaving downstream methods in ModelingToolkit. Its parent ed9acb0734dbb081441f40f3b4724d79da43139e, tested with the same Aqua 0.8.16, reports 14 older and entirely different piracies (search_variables!/toexpr). The split commit itself is not directly loadable (default_missing_guess_value is referenced before it exists); its immediate child 92a5505ac0ce413bda718a026349249d0814f61b is the first loadable post-split revision and reports 15 new cross-package methods centered on System, SCCNonlinearProblem, ShiftIndex, and MTKBase hooks.

2. Failing QA lane: d647d61ac46f3a5696021f4a0e52769aacb2194c added GROUP=QA, test/qa, and the unqualified Aqua check. Its sole parent c33e4efc9d57e319059a6c71e3425af2624f2d0c has no QA lane. Running the direct Aqua scan at d647d61ac4 with Julia 1.12.6/Aqua 0.8.16 reports exactly the same 31-method set as current master. Thus the QA lane was red from its first revision; no later ModelingToolkit commit regressed a previously green QA lane.

A proper fix needs the methods moved into their owning packages or integration hooks redesigned so a ModelingToolkit-owned argument participates in dispatch. I did not add treat_as_own, @test_broken, skips, or other test bypasses.

[ChrisRackauckas-Claude]

JET invalid cache_argtypes follow-up (2026-07-12): this is a Julia 1.12 compiler regression, not a ModelingToolkit source regression.

Clean package-level reproduction

On an isolated checkout of exact ModelingToolkit upstream master commit ee7dd34d8b7cfe8a085e0fa04cec28a607053f66, this faithful QA invocation resolved Julia 1.12.6, JET 0.11.5, and SciMLTesting 1.8.0:

timeout 7200 env GROUP=QA julia +1.12 --startup-file=no --project=. \
  -e 'using Pkg; Pkg.test(; coverage=false)'

JET logged signature-analysis errors for generate_semiquadratic_functions(::System, Any, Any, Any), generate_semiquadratic_jacobian(::System, Vararg{Any,4}), and SemilinearODEFunction(::System), then threw:

AssertionError: invalid `cache_argtypes` for `mi`
  [1] cache_lookup(...) @ Compiler .../inferenceresult.jl:189
  [2] cache_lookup(...) @ JET .../abstractinterpret/typeinfer.jl:301
  [3] const_prop_call(...) @ Compiler .../abstractinterpretation.jl

SciMLTesting reports Quality Assurance: JET-test failed for JET.report_package(ModelingToolkit; toplevel_logger=nothing, target_defined_modules=true).

Independent compiler reproducer and version boundary

The regression test added by JuliaLang/julia#61502 reproduces without ModelingToolkit or JET:

julia +1.12 --startup-file=no -e '
Base.@nospecializeinfer function jetls618(a, @nospecialize(rest...))
    z = a > 0 ? a + length(rest) : 0
    println(z)
    return z
end
println(Base.infer_return_type() do
    jetls618(1, 2, 3), jetls618(1, 2, 3, 4)
end)'

Observed locally:

• Julia 1.12.6: exit 1 with the identical assertion at Compiler/src/inferenceresult.jl:189.
• Julia 1.10.11: equivalent Core.Compiler.return_type invocation succeeds with Tuple{Int64, Int64}.
• Julia 1.13.0-rc1: succeeds with Tuple{Int64, Int64}.

This was previously reported as aviatesk/JETLS.jl#618. JuliaLang/julia#61502 fixes the cache-shape mismatch by respecting @nospecializeinfer during constant propagation. JuliaLang/julia#62003 (Backports for 1.12.7) merged on 2026-07-11 and explicitly backports #61502; current release-1.12 contains the fix, while installed Julia 1.12.6 does not.

Expected resolution: rerun this QA lane on Julia 1.12.7 once released. No ModelingToolkit test bypass or source workaround is appropriate for this assertion.

[ChrisRackauckas-Claude]

Clean current-master ModelingToolkitBase QA audit (2026-07-12)

I reproduced the official sublibrary lane from exact upstream master ea1c2d00727ec7af3bd4521ea7981eb6bc2782d9 with Julia 1.12.6 and Aqua 0.8.16:

timeout 3600 env GROUP=QA julia +1 --project=lib/ModelingToolkitBase \
  -e 'import Pkg; Pkg.test(; coverage=true, julia_args=["--check-bounds=auto", "--compiled-modules=yes", "--depwarn=yes"], force_latest_compatible_version=false, allow_reresolve=true)'

Baseline result: JET 54/54; Aqua 6 pass / 5 fail (exit 1). No broken markers, skips, ignores, or treat_as_own settings were used.

Metadata findings fixed

Draft PR #4737 removes the genuinely unused SimpleNonlinearSolve dependency and adds Optimization = "5.5" for the test extra. Both metadata defects entered with the ModelingToolkitBase split, fb2450ea82af06595598f61ea1bb4ef96c35edd9.

With #4737: focused Aqua stale-dependency and compat checks pass; official GROUP=QA reports JET 54/54 and Aqua 8 pass / 3 fail. The remaining failures are only the code-level classes below.

45 method ambiguities

• 34 are intersections between SciMLBase ODEProblem{iip, FunctionWrapperSpecialize} and ModelingToolkitBase deprecated ODEProblem{iip,spec}(::System, ...) constructors. SciMLBase has had its specialized constructor since 37715146e3ce2d56e50d3967bccb15e3c4dec9bc; the conflicting MTK deprecation family entered in 018e2d631d24ca0d418bceef6b215ee927f160d1 / PR refactor: add depwarns for old system and problem constructors #3682. These need real intersection methods in MTK or removal of the deprecated API under an appropriate SemVer-breaking release.
• 2 are local _split_helper intersections (Tuple/BlockedArray with Val{N} versus the Val{0} fallback), introduced by bfc73a3085bb81bd4377c4d7aa68926eb1f4a986 and the later BlockedArray specialization 41e801ee6ffd02dd408a280cf66a0e4f09dfec8a.
• 9 involve renamespace(::AbstractVector, x) versus more-specific second-argument methods. The vector fold method dates to 12bb1ed1882; counterpart boundaries include Equation a5c9440eb47702d7d653ed8392bb952dd7fc0d90, the type-stable System/Symbol/Num/Arr/CallAndWrap/Symbolic family de835be869d10e72423d0d0c76bfdb1ed1f60bf2, AnalysisPoint bc2c1287c24fd80ce12aa5068c6924c3395665c7, and symbolic arrays f5761ee69510d3b76d1dc5b8aa6ba5c4eadd76e2.

Five unbound-method reports from three definitions

• with_updated_parameter_timeseries_values(..., args::Pair{A,B}...) where {A,B<:NestedGetIndex}: the vararg can be empty, leaving both parameters unbound. The family originates in bfc73a3085bb81bd4377c4d7aa68926eb1f4a986 and was updated in 41e801ee6ffd02dd408a280cf66a0e4f09dfec8a.
• _remake_buffer(... idxs::Union{Tuple{...}, AbstractArray{<:ParameterIndex{P}}}, ...) where P: choosing the tuple branch leaves P unbound; Aqua reports the generated method plus its keyword wrappers. Introduced in 96694de0e47208da1e2d1e0f557a9c7f271e6fa2.
• rm_union(::Type{Union{T,Nothing}}) where T: union simplification can erase T. Introduced in a785adc22819ef641cfb97f918d2cc469ae33ae9.

These are local signature defects and should be fixed directly, not ignored.

14 piracy-reported methods

• 4 methods (including keyword wrappers) extend SymbolicUtils.search_variables! for JumpProcesses jump types. The current search_variables! form entered in b4f709ed43ca3d2d0cf8a015998efbc2e8960f66. The legal owner is JumpProcesses (which owns the dispatched types), or the integration must use an MTK-owned function/type.
• 10 methods extend SymbolicUtils.Code.toexpr for Symbolics Equation/Inequality/array/numeric types. The family dates to eb6cb6d8249254367b5f0ea6df08871eb165d666 and moved into ModelingToolkitBase at the split. These methods belong in Symbolics (owner of the dispatched symbolic types) or behind an MTK-owned interface.

No duplicate issue was opened. Draft #4671s broken-marker approach was not used.

[ChrisRackauckas-Claude]

The focused unbound-argument fix from this audit is now draft PR #4738. It removes the three unbound static-parameter definitions without Aqua ignores and adds numeric-union regression coverage.

Local verification: Aqua.test_unbound_args(ModelingToolkitBase) exits 0; GROUP=SymbolicIndexingInterface passes 64/64, 1804/1804, and 118/118; Runic passes repository-wide. On this independent branch the official QA result is JET 54/54 and Aqua 7 pass / 4 fail, with only the separately tracked ambiguity, stale dependency, missing compat, and piracy classes remaining. The metadata pair is handled independently in #4737.