Re-enable downgrade CI; raise compat floors (LinearSolve 5.4 et al.)

[ChrisRackauckas-Claude]

Summary

Re-enables the downgrade CI workflow (disabled in #162 after the failure reported in #115) and raises six minimum compat bounds to the smallest versions that are both bug-free and mutually resolvable. The key bump is LinearSolve = "5.1" → "5.4": LinearSolve <5.4 calls issquare in LinearSolveForwardDiffExt without importing it (fixed by SciML/LinearSolve.jl#1125), and a runtime-only resolve at the old floor still selects LinearSolve 5.1.0 today, so the broken bound is still reachable. The other five bumps are documentation-only corrections of floors that were already unreachable — at the declared bounds those versions cannot co-resolve with this package's own SciMLBase = "3.55" requirement.

Root cause of the September failure

Historical CI log (run 34304353015, Julia lts, julia-downgrade-compat@30b83cf):

Core/autodiff.jl: Error During Test
  LoadError: UndefVarError: `issquare` not defined
  [1] linearsolve_forwarddiff_solve!(...) @ LinearSolveForwardDiffExt
      ~/.julia/packages/LinearSolve/.../ext/LinearSolveForwardDiffExt.jl:75

That run resolved LinearSolve 5.2.0 + SciMLBase 3.48.0. issquare lives in SciMLOperators and is re-exported by SciMLBase only from 3.50.0 onward; LinearSolve's ForwardDiff extension used it bare until the explicit using SciMLOperators: issquare was added in LinearSolve 5.4.0 (5.2.0 and 5.3.0 both still lack the import — verified against the depot sources). The declared 5.1 floor is genuinely reachable: a scratch env with only the runtime deps at the old bounds resolved LinearSolve 5.1.0.

Verification

All steps on Julia +lts (1.10.12), matching the workflow's julia-version: lts, using the current julia-actions/julia-downgrade-compat@v2.7.0 downgrade.jl (commit 30b83cf94e89744fcbd7ea5496be4967d6fbe540) with mode deps and the effective skip list = Julia stdlibs (no caller skip), exactly as SciML/.github downgrade.yml@v1 passes. Each environment used its own depot under scratch.

Resolution at the new floors (merged main+test project, as v2.7 builds):

⌃ [2569d6c7] ConcreteStructs v0.2.3
⌃ [2b5f629d] DiffEqBase v7.18.1
⌃ [459566f4] DiffEqCallbacks v4.16.0
⌃ [f6369f11] ForwardDiff v1.3.3
⌃ [7ed4a6bd] LinearSolve v5.4.0
⌃ [0bca4576] SciMLBase v3.55.0
⌃ [0c5d862f] SymbolicIndexingInterface v0.3.47

Build: 140 dependencies successfully precompiled in 407 seconds. 114 already precompiled.

Test (CI-faithful, GROUP="" → SciMLTesting default "All" = Core files):

$ GROUP="" JULIA_DEPOT_PATH=<scratch depot>:~/.julia julia +lts --project=. \
    -e 'using Pkg; Pkg.test(allow_reresolve=false, coverage=false)'
...
Test Summary:    | Time
Core/autodiff.jl | None  48.4s
Test Summary: | Pass  Total   Time
Core/core.jl  | 2903   2903  42.5s
Test Summary:                | Pass  Total  Time
Core/developer_interfaces.jl |    5      5  0.3s
Test Summary:               | Pass  Total  Time
Core/precompile_workload.jl |    8      8  3.2s
Test Summary: | Pass  Total     Time
Core/scc.jl   |  411    411  2m07.8s
Test Summary: | Pass  Total   Time
Core/sicnm.jl |   33     33  50.4s
     Testing SteadyStateDiffEq tests passed

The same CI-faithful run at the baseline (pre-bump) floors also passed, because today's registry already forces LinearSolve ≥5.4 through test-only deps (NonlinearSolveFirstOrder/NonlinearSolveQuasiNewton → NonlinearSolveBase ≥1.15). That is exactly why the floor documentation matters: the resolved graph hides that the declared bounds were wrong.

Lint checks on the diff: typos .github/workflows/Downgrade.yml Project.toml → clean; Runic.main(["--check", "src", "test"]) → exit 0 (no .jl files changed); yaml.safe_load on the workflow → OK. The workflow diff is a pure rename of the file disabled in #162.

What I did NOT verify

• The workflow has not actually run on GitHub Actions yet — it was just renamed; this PR's own checks will be the first real run.
• The QA group (GROUP=QA, Aqua/ExplicitImports/JET) was not run at the downgraded floor; the CI-faithful GROUP="" run covers the Core files only, matching the workflow's invocation.
• A live UndefVarError reproduction at LinearSolve <5.4 is not possible with this package's current SciMLBase = "3.55" bound (3.55 re-exports issquare); the mechanism is established from the September log (SciMLBase 3.48.0), the missing import in LinearSolve 5.2.0/5.3.0 source, and the added import in 5.4.0.

What a reviewer should push back on

• Whether the five non-LinearSolve bumps are wanted at all: they are "documentation-only" in the sense that no consistent resolution could ever select below them given SciMLBase = "3.55". They match the floor correction produced by merged_resolve.jl and the precedent set in Re-enable downgrade CI: raise compat lower bounds and use allow_reresolve OrdinaryDiffEq.jl#3701 (bump every unreachable floor to the resolved minimum). Each is the smallest version satisfying the graph: e.g. ForwardDiff 1.0.0 is excluded by PreallocationTools required via SciMLBase ≥3.55; DiffEqBase <7.15 caps SciMLBase below the declared floor.
• LinearSolve = "5.4" rather than a tighter bound: 5.4.0 is the first release with the issquare import; nothing else forces higher.
• The bumps encode today's registry view; a different resolution epoch could derive different floors, but all bounds are lower bounds only.

Closes #115

Please ignore this draft until reviewed by @ChrisRackauckas.

Risk assessment

• Risk: low
• Blast radius: [compat] lower bounds only — no code changes; worst case a user pinned below a new floor cannot upgrade to the next release of this package (that combination was already unresolvable or buggy). Re-enabled workflow adds one CI job.
• Evidence: historical CI failure (run 34304353015, UndefVarError: issquare), source-level fix boundary confirmed in LinearSolve 5.2.0/5.3.0 vs 5.4.0, full Core suite green locally at the new floor with allow_reresolve=false on Julia 1.10.12.
• Independent review: Codex, model gpt-6-astra (a different harness from the Devin author) rated it VERDICT: MERGE, risk low, high confidence, with no blocking findings. It reran the floor resolve and an analytic ForwardDiff probe at the new floors. Review round 1 ran 2026-09-26 against head as of 2026-09-24T20:37Z; no commits since.
• Merge: needs human review — compat policy (which floors to document) is a maintainer judgment call, and the independent rating is pending.

---

🤖 Generated with Devin CLI 3000.11.1 (model: swe-2-max). Session: local transcript at /home/crackauc/sandbox/agent-jobs/SteadyStateDiffEq.jl/jobs/115-devin/log.txt on amdci2.julia.csail.mit.edu

Risk assessment

• Risk: low
• Blast radius: Dependency resolution for users pinned below six raised compat floors; one restored CI workflow. No source/API changes, new dependency-internal access, or weakened tests. Compat tightening excludes some previously permitted environments but introduces no demonstrated runtime/SemVer-breaking behavior.
• Evidence: Head 33d2ac3 has 13 successful checks, including test / Downgrade Tests, Core (lts, 1, x86), and QA. Corresponding Core/QA checks pass on default branch d81bcf1; no test regressions found. Core pre skips execution on both. Dependabot Auto-merge reports startup_failure: pre-existing infrastructure failure supported by earlier identical failures using the same reusable-workflow SHA; direct default-branch comparison is unavailable because this workflow only runs on PRs. Its logs are unavailable. No review comments exist. The description omits this startup failure and is stale about Downgrade CI not having run; its missing failing-before reproduction is explicitly disclosed.
• Independent review: Codex CLI 0.156.1 (Mac) / gpt-6-astra rated it low, high confidence: the complete diff only raises compat floors and restores an unchanged workflow, with executed test checks passing and no newly introduced test suppression or dependency-internal usage.
• Merge: auto-merge candidate

---

🤖 Risk assessment posted by an AI agent (fleet master) — harness: Codex CLI 0.156.1 (Mac) / gpt-6-astra; dispatched by Claude Code head, model claude-opus-5-5[1m]
Conversation: local Claude Code session 3cd6500a-1f81-46b5-ac0b-c466e15b6a53 on Chris's Mac (session ID, no URL)