Propagate release downstream resolver failures

[ChrisRackauckas-Claude]

Summary

• remove the ResolverError catch from the release downstream job
• let dependency-resolution and Catalyst test failures propagate to GitHub Actions
• retain the existing local SciMLBase development and Catalyst test sequence

This is stacked on #1445; the extra concurrency-context commit will disappear from this diff when that PR merges.

Ignore this PR until reviewed by @ChrisRackauckas.

Local verification

• /home/crackauc/.juliaup/bin/julialauncher +1.12 --startup-file=no -m Runic --check .
• reproduced the workflow's downstream environment locally with the branch version of SciMLBase and ran Pkg.test("Catalyst"; coverage=true)
• the command ran Catalyst's real test suite and exited 1 at the already-investigated UndefKeywordError: keyword argument parent_sys not assigned failure (20 passed, 1 errored in Symbolic Stoichiometry), rather than converting it to success

The Catalyst failure is pre-existing on the unmodified base and is tracked separately; this PR intentionally exposes it.

[ChrisRackauckas-Claude]

The newly visible LinearSolve.jl/All failure is another self-hosted runner loss, not a resolver or test assertion. Exact comparison: clean default-branch SHA db9ea39e job 87286414203 failed on self-hosted-4vcpu-8gb, while PR #1445 job 87305228663 used the same SciMLBase/LinearSolve source state on GitHub-hosted infrastructure and ended Testing LinearSolve tests passed. Five other recent 8 GB self-hosted LinearSolve jobs have the same runner-lost annotation, while three recent GitHub-hosted jobs passed. The workflow requests ubuntu-latest; routing is external to this PR. Evidence is being added to the existing infrastructure tracker SciML/.github#116.

[ChrisRackauckas-Claude]

Two additional failures in the current run have the same self-hosted-runner disconnect signature as the already-classified LinearSolve job:

• Downstream (Julia 1), runner suffix 66r4k
• Downstream (Julia pre), runner suffix phpnz

In both jobs the main downstream test step was still recorded as in_progress when the runner stopped communicating; neither job reported a Julia assertion, exception, or package-test failure. I am tracking this shared runner-loss pattern in SciML/.github#116 rather than treating these as downstream-matrix regressions.

[ChrisRackauckas-Claude]

One more job has now ended with the same infrastructure signature: DownstreamAD (Julia 1) lost self-hosted runner czdb2 while test step 12 was still in_progress. It also emitted no Julia test failure. This brings the current run's unambiguous self-hosted disconnects to LinearSolve plus Downstream, Downstream pre, and DownstreamAD.

[ChrisRackauckas-Claude]

Local clean-base verification is now terminal: at clean SciMLBase commit db9ea39e, Julia +1, GROUP=Downstream, JULIA_NUM_THREADS=11, and a clean depot, the full Downstream group exited 0. The final blocks included Initialization 80/80, Table Traits 10/10, SplitODE cache 1/1, and Scalar RODE errors 17/17, followed by Testing SciMLBase tests passed.

That direct run, together with the CI step metadata showing the test step left in_progress, confirms the failed Julia 1/pre Downstream cells are runner losses rather than source failures.

[ChrisRackauckas-Claude]

The matching LinearSolve downstream control is also terminal: a clean current SciMLBase/LinearSolve reproduction using four parallel Julia test tasks completed with exit 0 and Testing LinearSolve tests passed. The CI failure therefore remains classified as runner loss. The cross-repository job evidence, local controls, and hosted/self-hosted routing split are consolidated in SciML/.github#116.

[ChrisRackauckas-Claude]

A late-finishing fourth same-workflow disconnect is now classified: SymbolicIndexingInterface (Julia 1) lost self-hosted runner xq7hr while main test step 12 remained in_progress. It emitted no Julia failure. I have forwarded it for inclusion in the infrastructure rollup.

[ChrisRackauckas-Claude]

Terminal CI classification for this workflow-audit PR: 55 checks succeeded and 8 failed.

Five failures are runner disconnects, not Julia test failures:

• LinearSolve
• Downstream / Julia 1
• Downstream / prerelease
• DownstreamAD / Julia 1
• SymbolicIndexingInterface / Julia 1

In each of those logs, the test step was still marked in progress when the hosted runner stopped responding.

The other three checks reached and ran substantive downstream tests:

• Catalyst's internal and reusable-workflow jobs both report the same Symbolic Stoichiometry error: UndefKeywordError: keyword argument parent_sys not assigned at test/integrator_interface.jl:235.
• NonlinearSolve's Core group reports one failing generalized-Rosenbrock Broyden case (norm(u - u_true) = 4.4 > 0.001; summary: 94 passed, 1 failed, 20 broken).

Those are genuine downstream test failures, so they are deliberately not being hidden or converted to warnings. The workflow change is doing its intended job: each remaining non-infrastructure failure actually entered a package test group.

Local controls also completed successfully on clean current bases:

• SciMLBase GROUP=Downstream with Julia release: package test exited 0 (including Init 80/80, Table 10/10, Split 1/1, and RODE 17/17).
• LinearSolve GROUP=Core with four parallel test tasks: package test exited 0.

The hosted-runner disconnect pattern is tracked in SciML/.github#116.

[ChrisRackauckas-Claude]

Follow-up on the three substantive downstream failures:

• Catalyst's duplicated parent_sys failure is addressed by focused draft Test symbolic stoichiometry through public API Catalyst.jl#1511. It replaces the test's private ModelingToolkitBase AffectSystem inspection with a public end-to-end JumpProblem solve; the changed symbolic-stoichiometry file passed locally 30/30.
• There is also an open NonlinearSolve draft, Update bad Broyden robustness expectations NonlinearSolve.jl#1069, for the Broyden case. That draft adds the failure to the suite's broken_tests list. I am not treating that as an acceptable resolution for this audit: the repository instructions explicitly prohibit marking a newly failing case broken to make the suite pass. A separate underlying-cause reproduction/bisect is queued instead.

This does not change the classification above: all three #1446 failures ran substantive tests rather than silently no-oping.