Fix Downgrade CI: raise compat lower bounds to a resolvable, testable minimum set

[ChrisRackauckas-Claude]

This PR should be ignored until reviewed by @ChrisRackauckas.

Problem

The Downgrade workflow fails on main for all three groups (Core, QA, Applications). It was added on 2026-06-02 by the centralized-workflows migration and has never had a green run: the first completed run on main (7425dea, 2026-06-02T16:58Z) failed, as has every run since (8bc824b 2026-06-07, 476fdbf 2026-06-09). The boundary is therefore the introduction of the workflow itself, not registry drift or a repo commit: the minimum versions allowed by [compat] were never mutually resolvable with the test set, and the workflow is the first thing that ever checked them.

First failing Core job on main (2026-06-02):

ERROR: LoadError: Unsatisfiable requirements detected for package CUDA [052768ef]:
 ├─restricted to versions 4-5 by project, leaving only versions: 4.0.0-5.11.2
 └─restricted by compatibility requirements with GPUArrays [0c68f7d7] to versions:
   [3.3.5-3.13.1, 6.0.0-6.1.0] or uninstalled — no versions left
   └─GPUArrays restricted to versions 8.4.2 by an explicit requirement

Root cause

The downgrade workflow resolves the package manifest to the minimum versions allowed by [compat] (Resolver.jl --min=@deps) and then runs Pkg.test(; allow_reresolve=false), so the test sandbox cannot move those pins. The old lower bounds (SciMLBase = "1", NonlinearSolve = "1", LinearSolve = "2", Zygote = "0.6", ...) fail in several independent ways:

1. The source code requires the SciMLOperators ≥ 1 interface. All FunctionOperator callbacks in the package use the 5-argument matvec(w, v, u, p, t) convention introduced in SciMLOperators 1.x; SciMLOperators 0.3/0.4 call the in-place op as op(v, u, p, t). I verified this empirically: at intermediate minimums (SciMLBase 2.54, LinearSolve 2.36.1 → SciMLOperators 0.3.12) the Core CPU suite throws 8 MethodErrors in every FunctionOperator-based solve (IPM and adjoint paths). Reaching SciMLOperators 1.x requires SciMLBase ≥ 2.100 and LinearSolve ≥ 3.13.
2. The minimal manifest pins force CUDA 6 in the test sandbox. SparseMatrixColorings resolves to 0.4.27 (pulled in via NonlinearSolve), which requires CUDA 6; CUDA 6 in turn needs NNlib ≥ 0.9.34 and DiffEqBase ≥ 6.215, and DiffEqBase ≥ 6.215 needs SciMLBase ≥ 2.143. test/Project.toml gets CUDA = "4, 5, 6".
3. Old minimums break test-time precompilation. E.g. SciMLSensitivity 7.75 references SciMLBase.LinearAliasSpecifier while its compat still allows SciMLBase 2.54 (upstream under-constraint) — precompile failure at minimum versions; SciMLSensitivity ≥ 7.100 (needed to work with FastBroadcast 1.3) requires SciMLBase ≥ 2.147.
4. The test environment needs OrdinaryDiffEq ≥ 6.106 (forced by the DataStructures 0.19.5 / PreallocationTools 1.2 pins), which sets floors for LinearSolve (3.61, via OrdinaryDiffEqDifferentiation ≥ 2.5 with FastBroadcast 1.3), NonlinearSolve (4.12), SimpleNonlinearSolve (2.7), Polyester (0.7.18), FillArrays (1.13), PrecompileTools (1.2.1 — note 1.3.x is Julia 1.12+ only), and TruncatedStacktraces (1.4).
5. NNlib ≥ 0.9.32 requires ForwardDiff 1, and SciMLBase ≥ 2.102 requires Zygote ≥ 0.7 (≥ 0.7.5 once ForwardDiff 1 is in play), so the 0.10.x / 0.6.x lower bounds were unreachable.
6. ParametricMCPs is unresolvable in this stack. Every resolvable release caps ForwardDiff at 0.10 (directly or via SymbolicTracingUtils' SciMLOperators 0.3 cap). This PR cherry-picks the commit from Breaking dependency updates: CUDA 6, DiffEqBase 7, OrdinaryDiffEq 7, Optimization 5 (supersedes #59) #63 that removes the ParametricMCPs comparison benchmark from the test suite (910692a).

Changes

Project.toml [compat] lower-bound raises:

package	old	new
CommonSolve	0.2	0.2.6
FillArrays	1	1.13
ForwardDiff	0.10, 1	1
GPUArraysCore	0.1, 0.2	0.2
LinearSolve	2, 3	3.61
NNlib	0.8, 0.9	0.9.34
NonlinearSolve	1, 2, 3, 4	4.12
PATHSolver	1	1.4
Polyester	0.7	0.7.18
PrecompileTools	1	1.2.1
SciMLBase	1, 2, 3.1	2.147, 3.1
SimpleNonlinearSolve	0.1, 1, 2	2.7
TruncatedStacktraces	1	1.4
Zygote	0.6, 0.7	0.7.5

test/Project.toml: CUDA = "4, 5" → "4, 5, 6", and the ParametricMCPs test dependency removal cherry-picked from #63.

These bounds only drop version combinations that could never have passed the test suite (either unresolvable together with the test dependencies, or runtime-broken against the SciMLOperators 0.x interface), so no working configuration is lost.

Verification (local, Julia 1.10.11)

• Reproduced the CI failure on unmodified main with the workflow's exact procedure (Resolver.jl --min=@deps with the workflow's skip list, then Pkg.test(; allow_reresolve=false)): same minimal pins as CI and the same unsatisfiable-requirements error.
• With this PR, the downgrade resolution succeeds and pins (selection): SciMLBase 2.147.0, NonlinearSolve 4.12.0, SimpleNonlinearSolve 2.9.0, LinearSolve 3.61.0, SciMLOperators 1.22.0, Zygote 0.7.5, ForwardDiff 1.0.1, NNlib 0.9.34, FillArrays 1.13.0, Polyester 0.7.18, PATHSolver 1.4.0, CommonSolve 0.2.6, TruncatedStacktraces 1.4.0, PrecompileTools 1.2.1, GPUArraysCore 0.2.0.
• Pkg.test(; allow_reresolve=false) against that pinned manifest, run to completion for all three Downgrade CI groups (BACKEND_GROUP=CPU):
  • Core: 212 pass, 0 fail, 0 error, 6 broken — passed
  • QA: 5 pass, 1 broken — passed
  • Applications: 41 pass, 1 broken — passed

Relationship to #63: that PR updates the upper bounds (CUDA 6, DiffEqBase 7, OrdinaryDiffEq 7, Optimization 5) for the latest stack; this PR fixes the lower bounds for Downgrade CI on main as it currently stands. They share only the ParametricMCPs-removal commit, so whichever lands first, the other rebases cleanly.

🤖 Generated with Claude Code