Separate full_equations timeout from subprocess setup

[ChrisRackauckas-Claude]

Depends on #4844. This branch is based directly on its head commit, 9d7b22a35bc107a880baa10b54b90cb41f561b3e; the only commit added here is 10949fa3552a3cd13aeb7f303d879b12f1a497ee.

Ignore this PR until it has been reviewed by @ChrisRackauckas.

Problem

The full_equations recursion regression test slept for 180 seconds before checking its child process. That combined package loading, cache-lock waits, precompilation, model setup, and the call being tested into one timeout.

On clean master, the child needed 265.37 seconds to finish after compiling ModelingToolkitBase. A contention-focused run spent 293 seconds precompiling before the tested call and then passed. The old assertion could therefore fail without full_equations hanging.

The subprocess itself is necessary: running the model against the parent of the original fix entered full_equations and was still running when an external 3600-second timeout terminated it.

Change

• Touch a readiness marker after package loading, model construction, and mtkcompile.
• Allow setup/cache work up to 3600 seconds.
• Preserve the 180-second timeout specifically around full_equations.
• Reap the child and require a successful exit, so a crash no longer satisfies the test.

Investigation

• f215ec083a439be63603b83bc42178db24ce685e introduced the subprocess regression test with a 120-second combined timeout.
• 9ed9aa758c759c8d49155679a915e8db0a88fcf1 increased that timeout to 180 seconds.
• Clean-master GROUP=InterfaceI passed 1475 tests with 3 pre-existing broken tests in 49m03s, confirming that the reported failure is contention-dependent rather than deterministic.

Validation

• Focused test, Julia 1.12: 4/4 passed in 10.8 seconds.
• Focused test, Julia 1.10 LTS: 4/4 passed in 1m53.7s.
• Cold/contention-focused test: 4/4 passed in 5m05.6s, including 293 seconds of child precompilation.
• GROUP=InterfaceI julia +1.12 --project=lib/ModelingToolkitBase -e 'using Pkg; Pkg.test(; coverage = false)': 1498 passed, 5 pre-existing broken, 1503 total, in 38m16.8s.
• Whole-repository Runic 1.5.1 check passed.

[ChrisRackauckas-Claude]

Stacked-PR plan/status:

1. Merge prerequisite Fix SampleTime clock inference formatting #4844 (9d7b22a35bc107a880baa10b54b90cb41f561b3e).
2. Review this PR's added commit 10949fa3552a3cd13aeb7f303d879b12f1a497ee, which only changes the full_equations subprocess test.
3. Keep the setup/readiness wait separate from the 180-second execution timeout and retain subprocess isolation.
4. Monitor CI; the focused Julia 1.12/LTS tests, full ModelingToolkitBase InterfaceI group, and whole-repository Runic check have passed locally.

GitHub does not allow the fork-only #4844 head branch to be selected as this upstream PR's base, so the PR targets master and explicitly depends on #4844.

[ChrisRackauckas-Claude]

CI triage update (exact upstream-master comparison):

The required lib/ModelingToolkitBase [InterfaceI] jobs pass on Julia lts/current/pre. The remaining code-related failures are already present on exact clean upstream master commit 529c98930ee14191a0e90a32e37698ff6dc40309:

• Catalyst Modeling: clean-master job https://github.com/SciML/ModelingToolkit.jl/actions/runs/30363096231/job/90287406626 has the same seven unexpected passes from existing @test_broken assertions.
• ModelingToolkitStandardLibrary Core: clean-master job https://github.com/SciML/ModelingToolkit.jl/actions/runs/30363096231/job/90287406602 has the same test_analysis_points.jl:271 missing operating-point value error.
• StructuralIdentifiability Core: clean-master job https://github.com/SciML/ModelingToolkit.jl/actions/runs/30363096231/job/90287406580 has the same downstream generators ambiguity.
• ModelingToolkitBase QA: clean-master job https://github.com/SciML/ModelingToolkit.jl/actions/runs/30363097045/job/90287762143 has the same Aqua ambiguity/piracy failures.

The PR MTKBase InterfaceII current/pre jobs are infrastructure failures: their check annotations say the self-hosted runner lost communication, and the test steps never completed. Job-specific rerun attempts were rejected because both available GitHub credentials lack Actions write permission. These issues are being kept out of this focused timeout PR.

[ChrisRackauckas-Claude]

Exact clean-master local confirmation for the MTSL downstream failure:

• MTK/MTKBase: 529c98930ee14191a0e90a32e37698ff6dc40309 (detached clean upstream master)
• ModelingToolkitStandardLibrary: c2ee11b86e873820521199f878e9d809f07e164a
• Julia: 1.12.6
• command: official GROUP=Core through Pkg.test(coverage = true)
• result: exit 1, Core/test_analysis_points.jl had 37 passes / 1 error
• error: test/test_analysis_points.jl:271 does not provide the loop-opening parameter feedback₊input2₊u(t) in the operating point

This exactly matches the upstream-master CI failure, so it is independent of this PR. The first local invocation spent most of its one-hour guard precompiling and timed out in sources.jl; I reran with the same resolved manifest and warmed isolated depot, and the official Core group then reached and reproduced the assertion above.

[ChrisRackauckas-Claude]

Exact clean-master local confirmation for the Catalyst downstream failure:

• MTK/MTKBase: 529c98930ee14191a0e90a32e37698ff6dc40309 (detached clean upstream master)
• Catalyst: 53e6482dba3d884e4c40b7120d7d2356ec24ab9f
• Julia: 1.12.6
• command: official GROUP=Modeling through Pkg.test(coverage = true)
• result: exit 1; Conservation Laws summary was 242 passes / 7 errors / 17 broken
• all seven errors were unexpected passes at conservation_laws.jl lines 150, 152, 179, 183, 287, 290, and 325

This exactly matches the upstream-master CI failure, so it is independent of this PR. As with the MTSL audit, a first invocation exhausted its one-hour guard after first-time precompilation; the second invocation used the same resolved manifest and warmed isolated depot and reached the target failure through the official Modeling group.