Isolate ModelingToolkit aggregate test environments

[ChrisRackauckas-Claude]

This PR should be ignored until reviewed by @ChrisRackauckas.

Summary

• Opts ModelingToolkit aggregate test runs into fresh-process isolation for environment-backed groups.
• Requires SciMLTesting 2.3 in the root test target and each standalone group environment.
• Declares the packages and stdlibs imported directly by the Downstream, FMI, ModelingToolkitBase Extensions, and ModelingToolkitBase Optimization test bodies in their own Project.toml files.

Dependency

This is stacked on SciML/SciMLTesting.jl#23 and must remain draft until that change is merged and SciMLTesting 2.3 is available.

Motivation

The aggregate runner previously loaded LinearSolve 3.87 from one group, activated a later group environment resolving LinearSolve 4.3, and then combined the already-loaded module with the newer extension source. The extension referenced LinearSolve.ArnoldiMethodJL, which does not exist in 3.87. A fresh Julia process per environment-backed aggregate member prevents that mixed-module state. Direct GROUP selections intentionally retain their existing in-process behavior for fast iteration and CI lanes.

Local validation

Using Julia 1.10.11, the SciMLTesting #23 checkout, an isolated local depot, and a local wrapper that re-entered this repository test harness through the new fresh-process path:

• Downstream: 37 passed, exit 0, 8m08.9s. The validation also asserted a new child PID, LinearSolve 4.3, the presence of ArnoldiMethodJL, and local ModelingToolkit/SciMLTesting sources.
• FMI: 60 passed, exit 0, 32m12.5s.
• ModelingToolkitBase Extensions: 96 passed, exit 0, 27m29.5s.
• ModelingToolkitBase Optimization: 131 passed and 1 pre-existing broken assertion, exit 0, 20m49.6s.
• Whole-repository Runic inplace and check/diff passes exited 0. A post-rebase whole-repository Runic check/diff also exited 0.

No QA pass is claimed. An isolated aggregate QA run reported 10 passes and 3 failures: the clean-base Aqua piracy set, existing JET top-level analysis errors, and the SciMLTesting 2.3 public-docstring audit. Direct GROUP=QA behavior is unchanged. The 31-method clean-base piracy reproduction and historical boundary are tracked in #4670. The exact 28-name SciMLTesting 2.3 documentation exposure is recorded on #4715. No skip, broken marker, allow-list, relaxed assertion, or JET-mode weakening is included here.

A full aggregate pass is also not claimed: a clean-base cross-group state failure is under a separate audit rather than being hidden by this environment fix.

Process notes

I first reproduced the mixed LinearSolve versions in one Julia process, then confirmed that a fresh process resolves and loads one coherent version. The SciMLTesting prerequisite adds synthetic same-UUID v1/v2 regression coverage on Julia 1.10 and 1.12; its complete local suites passed 299/299 and 314/314 respectively. After validating each real environment-backed ModelingToolkit group, I audited direct imports, added only the dependencies actually used by each standalone test body, ran repository-wide formatting, synchronized the fork master without overwriting divergent commits, rebased this branch onto current upstream master, and repeated the formatting/diff checks.

[ChrisRackauckas-Claude]

Scratchpad / commit d6200ad:

• Rebased the single focused consumer commit onto upstream master db56a6a before publication.
• Set isolate_group_environments=true only at the aggregate run_tests call. Direct named groups remain in process.
• Updated SciMLTesting compat to 2.3 in the root and five group environments.
• Added only imports that the four functional group bodies use directly; no internal dependency APIs were introduced.
• Verified the four environment-backed functional groups locally, then ran whole-repository Runic and git diff checks after the rebase.
• Kept QA findings visible and linked QA: track pre-existing Aqua/JET/ExplicitImports findings (root + ModelingToolkitBase) #4670 / Document public API coverage gaps #4715 instead of adding suppression.

Prerequisite implementation and its synthetic regression tests: SciML/SciMLTesting.jl#23.

[ChrisRackauckas-Claude]

CI gate classification (initial run):

• Inspected all 26 completed failure logs: 25 root test-matrix jobs plus the downgrade job.
• 26/26 fail at dependency resolution because the registry offers SciMLTesting through 2.2.0 while this stacked PR correctly requires 2.3.
• Unexpected failure classes: 0.
• Runtime tests started: 0. The logs contain no test/runtests.jl execution and no Test Summary; each matrix job stops while Pkg constructs the test sandbox.

This is the expected prerelease gate until SciML/SciMLTesting.jl#23 is merged and SciMLTesting 2.3 is registered. No source change, retry, or test bypass is appropriate for this run.