Require Dual-safe BoundaryValueDiffEqMIRK

[ChrisRackauckas-Claude]

Please ignore this PR until it has been reviewed by @ChrisRackauckas.

Summary

• raise ModelingToolkitBase's BoundaryValueDiffEqMIRK compatibility floor from 1.7.0 to 1.17.3
• keep downgrade-compatible dependency graphs on the first MIRK release with Dual-safe evaluation buffers

Root cause

The ModelingToolkitBase InterfaceII downgrade job in #4852 admitted a MIRK tree older than 1.17.3. That implementation reused Float64 solution-evaluation storage when the nonlinear solve evaluated a ForwardDiff.Dual iterate, producing MethodError: no method matching Float64(::ForwardDiff.Dual) in both the Lotka–Volterra BVP and BVP cost-function tests.

SciML/BoundaryValueDiffEq.jl#568 changed MIRK evaluation storage to select the iterate element type and shipped that fix in 1.17.3. ModelingToolkit.jl#4843 restored the real BVP assertions that expose the old implementation, so the ModelingToolkitBase compatibility bound needs to express that runtime requirement.

Latest-compatible environments are unaffected; this changes the old-version boundary selected by downgrade testing and by users whose environments would otherwise admit an affected MIRK release.

Validation

• Reproduced the original downgrade failure from clean d39174937fead779b29fa5baa50ba975adbde8c9, Julia 1.10.11, and the exact julia-actions/julia-downgrade-compat@2c401f8b47edbcafc2e1edbff8e71f33dce71822 alldeps workflow. The official GROUP=InterfaceII run exited 1 with the same two Dual-conversion errors: 13 passed, 2 errored, and 5 existing broken checks in Optimal Control + Constraints Tests.
• Repeated the exact downgrade workflow with this compatibility floor. It selected BoundaryValueDiffEqMIRK 1.17.3, and the official GROUP=InterfaceII run exited 0 with 22 passed and 6 existing broken checks in Optimal Control + Constraints Tests, followed by Testing ModelingToolkitBase tests passed.
• Ran bvproblem.jl directly in the exact downgraded test sandbox: exit 0; Lotka–Volterra 4/4, Pendulum 4/4, boundary compilation 2/2, constraints 6 passed plus 3 existing broken, cost compilation 6 passed plus 2 existing broken, and parameter estimation 1 existing broken.
• Ran whole-repository Runic 1.7.0: exit 0.
• Ran GROUP=QA on the feature branch and exact clean master with Julia 1.10.11 and the same depot/registry snapshot. Both passed JET 54/54 and reported the identical pre-existing Aqua baseline (46 ambiguity signatures and 14 piracy signatures); the sorted signature sets match exactly. Those independent QA fixes remain separate in Fix ModelingToolkitBase method ambiguities #4849 and QA: track pre-existing Aqua/JET/ExplicitImports findings (root + ModelingToolkitBase) #4670.

The existing BVP tests are the regression coverage for this compatibility correction; no duplicate test or workaround is added.

[ChrisRackauckas-Claude]

CI classification: the independent ModelOrderReduction/Core failure is addressed by SciML/ModelOrderReduction.jl#169. Its exact clean-MTK-master downstream Core reproduction passes 30/30 at commit 80e1f75688eb8cc3827ab2f48ae25b0388a1c19a; it is unrelated to the BoundaryValueDiffEqMIRK compatibility-floor change here.

[ChrisRackauckas-Claude]

Terminal hosted-CI update

The matrix is terminal at 62 success / 28 failure / 3 skipped. The last pending check, downgrade-sublibraries / test (lib/ModelingToolkitBase), did not report a test assertion failure at its endpoint:

• the target Optimal Control + Constraints Tests that fails on clean master completed 22 pass / 6 broken on this branch;
• the job continued through later suites and had entered the QA environment precompile;
• at 2026-07-30 11:28:43 UTC the self-hosted runner reported The runner has received a shutdown signal, Julia received SIGTERM, and the check annotation became The operation was canceled.

The log also exposes independent LinearSolveKrylovKitExt / LinearSolveArnoldiMethodExt precompile errors involving EigenvalueProblem in later isolated test environments. Their surrounding completed testsets continued to pass, so they are not evidence against this one-line MIRK floor correction, but they need a separate dependency/depot-boundary audit rather than being hidden here.

The downstream NNPDE1 failure is likewise independent and now has an exact clean-master, coverage-sensitive reproducer at SciML/NeuralPDE.jl#1095. Clean master and this PR's merge tree produce byte-identical covered trajectories and the same failure; both pass without coverage.

No branch change or test suppression was made in response to either independent failure.

[ChrisRackauckas-Claude]

The NeuralPDE.jl/NNPDE1/1 failure is independent of this PR.

I reproduced the hosted failure on clean ModelingToolkit master
(61cc6bca692a208124eec413c95c2ab73fd8388f) using Julia 1.12.6, the hosted
outer testset RNG, and coverage. Clean master and this PR's merge produced
byte-for-byte identical callback values, terminal loss, prediction vector, and
failure:

• initial loss 3.3244116222205697
• terminal loss 4.502195204451791e-8
• prediction L2 error 1.2247332813759246e-4
• unchanged atol = 1e-4 array assertion fails

Without coverage, the same RNG passes on both clean master and the PR merge,
again with identical output. The PR's only code-tree difference from its base
is the BoundaryValueDiffEqMIRK test compat in ModelingToolkitBase, and that
package is absent from the NeuralPDE graph.

The untouched official clean-master GROUP=NNPDE1 run with
Pkg.test(coverage=true) also passed 22/22 across all eight files in 59m10.5s;
it selected a different random initialization. The recorded hosted RNG still
fails deterministically under coverage.

Tracked separately in SciML/NeuralPDE.jl#1095.