[ChrisRackauckas-Claude]

🚧 UNREVIEWED — Ignore this PR until reviewed by @ChrisRackauckas.
🤖 This PR was prepared by an AI agent; Chris has not reviewed it.

Local test setup can rewrite tracked source paths or omit solver extensions loaded by earlier official testsets. Add16 lines to AGENTS.md covering portable QA source paths and preservation of the official runner's module-loading context in reduced tests. This keeps setup failures separate from numerical regression evidence. Runtime, tests, dependencies and QA configuration are unchanged.

Verification

The existing official QA runner on the earlier guidance revision, with the identical runtime/test/environment tree, reproduced the path rewrite, and the setup-only change was restored after all test processes ended. Final tracked diff contains AGENTS.md alone; spelling and whitespace checks exit 0. Julia source formatting and a docs build do not apply to this Markdown guidance change.

Julia 1.13.1, four Julia/precompile threads, one BLAS/OMP thread; isolated local validation environment:

JULIA_REVISE_POLL=1 GROUP=QA julia +1.13 --startup-file=no --project="$VALIDATION_ENV" -e 'using Pkg; Pkg.test("ModelingToolkitBase")'

JET Tests  | 54  54  50.3s
Aqua Tests | 21  21  5m47.2s
Testing ModelingToolkitBase tests passed
exit 0

Two earlier runs with default watcher transport passed JET54/54 and failed Aqua20pass/1error at FolderMonitor(start): EMFILE. An independent untouched-base and standalone syscall investigation isolated shared inotify-instance accounting: Julia had 18 file descriptors, no inotify instance and a 1M descriptor limit; ordinary file opens still succeeded. Revise's documented polling transport keeps file monitoring active and runs the same assertions without inotify. The passing run changes only that job-scoped transport setting; no test, tolerance, exception policy or host setting was changed. It does not claim to repair the host's inotify capacity.

No new behavior or public API is introduced. No local Everything, docs, GPU or downstream run is claimed.

CI comparison

Exact reviewed guidance head49aef0358dd742896a25caecf13e25ca3a186caf completed with71 successful checks, ten failures, zero pending and four skipped. Modern ci-compare records all ten failures as pre-existing against default207f39e2d3. Five InterfaceI variants share the causal equation-direction assertion at components.jl:446; three Optimization variants share unsupported literal_pow; one indexing job has the lowered-state BoundsError; Sensitivity Core8 shares the default time-gradient AD failure at mtk.jl:216. No runtime/test/environment file changes.

The current-head CI docs build passed; this is CI evidence, not a local docs run. Current CI, matching default CI, current Sensitivity Core8, matching default Core8.

Independent review

A new fresh Codex CLI0.163.0/gpt-6-astra reviewer inspected the complete16-line diff before author evidence and rated it low risk, high confidence, no blocking findings. It independently verified that only AGENTS.md differs from clean207, all runtime/test/metadata bytes match, Code Generation precedes Discrete System in both runners, and Code Generation imports the public OrdinaryDiffEq solver package. It also confirmed the root macro loads ModelingToolkit while the sublibrary macro differs. Its own spelling and whitespace checks pass. It read the completed54+21 QA logs above; it did not rerun Julia for the new prose. No permission or API change is introduced.

Current reviewer: local session01a15ec6-8592-73e3-a785-76bb35f0e0d8 (no session URL). The earlier eight-line QA-path guidance had a separate fresh Astra review, local session01a15da9-56d1-7771-869f-cde9b62f0166 (no session URL), followed by completed official QA and metadata restoration.

Links

• Tested base: 207f39e
• Separate causal assertion correction: Test causal connector direction against input/output metadata #5298
• Existing Optimization report/fix: OptimizationMOI Core rejects literal_pow expressions with SymbolicUtils 4.52.0 Optimization.jl#1435 ; Handle integer-power codegen in OptimizationMOI expressions Optimization.jl#1436
• Existing lowered-state indexing report: solve(::SteadyStateProblem) returns the lowered problem's solution: sol.u reordered/empty, sol.prob.f.sys time-independent SteadyStateDiffEq.jl#171
• Sensitivity Core8 matching failure: https://github.com/SciML/ModelingToolkit.jl/actions/runs/39585288783/job/119172560011
• Sensitivity Core8 default baseline: https://github.com/SciML/ModelingToolkit.jl/actions/runs/39509026516/job/118925097105
• CI docs pass: https://github.com/SciML/ModelingToolkit.jl/actions/runs/39585288785/job/119172559422
• Revise polling documentation: https://timholy.github.io/Revise.jl/stable/config/
• Independent review evidence: Document local QA setup and reduced-test loading context #5300 (comment)
• Exact reviewed guidance commit: 49aef03

Risk assessment

• Risk: low
• Blast radius: repository contributor guidance only; no runtime, tests, dependency constraints or public API change.
• Evidence: repeated observed setup-only path rewrites, official QA54+21pass with documented polling transport, clean final source/test/environment bytes, spelling/whitespace exit 0.
• Independent review: Codex CLI 0.163.0 / gpt-6-astra rated low risk/high confidence, no findings; required QA condition subsequently met (review evidence).
• Merge: qualified parent handoff under the inherited-red policy: fresh low-risk review and all ten current failures default-matching. Draft remains subject to @ChrisRackauckas review; root does not merge.

🤖 Posted by an AI agent — harness: Codex CLI 0.163.0 · model: gpt-6.1-sol.
Conversation: local session 01a15ca3-5696-7ea3-bbac-ba551b6975af (no session URL).