Require all components to stall before resetting Broyden

[ChrisRackauckas-Claude]

Please ignore this PR until it has been reviewed by @ChrisRackauckas.

What changed

NoChangeInStateReset now resets Broyden only when every component of the state step or residual change has stalled. The previous any reduction treated one unchanged component as evidence that the entire iteration had stalled, which could consume every reset and fail an otherwise convergent solve.

The regression test uses only public APIs. The existing Helical problem contract is also tightened by removing problem 5 from the alg-3 broken list, and NonlinearSolveQuasiNewton receives a patch version bump to 1.15.2.

A behavior-neutral blank-line removal in NonlinearSolveBase.jl is included deliberately so that this PR triggers the repository's complete source-dependent CI matrix.

Failing before / passing after

The same new test was run against an unfixed clean worktree and this branch:

julia +1.12 --project=lib/NonlinearSolveQuasiNewton -e 'using Test; include("lib/NonlinearSolveQuasiNewton/test/core_tests__item13.jl")'

Unfixed:

Test Failed at core_tests__item13.jl:9
Expression: SciMLBase.successful_retcode(sol)
retcode = ConvergenceFailure
nresets = 1
maximum residual = 0.040265031294151266

Fixed:

Test Summary:                                      | Pass  Total
No-change reset requires every component to stall |    2      2
retcode = Success
nresets = 0
maximum residual = 5.329070518200751e-15

The existing Helical problem 5 with Broyden alg 3 discriminates the same failure mechanism:

unfixed: ConvergenceFailure, 301 evaluations, 100 resets, residual 3.9563
fixed:   Success,             18 evaluations,   1 reset,  residual 2.2373e-16

Local verification

GROUP=Core julia +1.12 --project=. -e 'using Pkg; Pkg.test()'

23 Test Problems: Broyden | 95 pass, 20 existing broken, 115 total
Testing NonlinearSolve tests passed

From lib/NonlinearSolveQuasiNewton:

GROUP=Core julia +1.12 --project=. -e 'using Pkg; Pkg.test()'
GROUP=QA julia +1.12 --project=. -e 'using Pkg; Pkg.test()'

Broyden | 594/594
Klement | 216/216
LimitedMemoryBroyden | 99/99
No-change reset requires every component to stall | 2/2
Testing NonlinearSolveQuasiNewton tests passed
QA | 20/20

From the repository root:

GROUP=QA julia +1.12 --project=. -e 'using Pkg; Pkg.test()'
julia +1.12 --project=@runic -e 'using Runic; exit(Runic.main(["--check", "."]))'
typos <changed files>
git diff --check

QA | 28/28
Runic: exit 0
typos: exit 0
git diff --check: exit 0

Not verified locally

The original Wood failure occurs on macOS after LinearSolve 5.6.0 → 5.7.0. On Linux, Wood succeeds both before and after this patch with zero resets, so local testing does not prove that this fix repairs that architecture-specific trajectory. The deliberate NonlinearSolveBase source touch is present so macOS and the complete downstream matrix can decide that directly.

No GPU tests were run locally. This PR does not change public API or documentation.

Links

• Tracking issue: True-Jacobian Broyden Wood test regressed on macOS with LinearSolve 5.7 #1158
• Original failing macOS CI: https://github.com/SciML/NonlinearSolve.jl/actions/runs/32118981711
• LinearSolve trajectory boundary: SciML/LinearSolve.jl@6bbb233

[ChrisRackauckas-Claude]

The source-triggered macOS Core lanes have now provided the missing platform result: this reset-condition fix does not fix the separate Wood regression. It fails identically on Julia 1, LTS, and prerelease.

23 Test Problems: Broyden | 94 pass | 1 fail | 20 broken
4: Wood function | alg #2: failed

That confirms the PR's local caveat: the one-component false-reset bug is real and fixes Helical, but Wood has an independent LinearSolve-selection regression. I am keeping that correction in a separate focused branch/PR.

CI jobs:

• https://github.com/SciML/NonlinearSolve.jl/actions/runs/32368371779/job/96432457092
• https://github.com/SciML/NonlinearSolve.jl/actions/runs/32368371779/job/96432457316
• https://github.com/SciML/NonlinearSolve.jl/actions/runs/32368371779/job/96432457356

[ChrisRackauckas-Claude]

The source-triggered macOS Core matrix confirms that this reset-condition fix is separate from the Wood regression: all three macOS versions still fail Wood alg #2 with the exact clean-master residual. The inverse-Jacobian initialization fix is isolated in draft #1185 so the two behavior changes remain independently reviewable and bisectable.

[ChrisRackauckas-Claude]

The full source-triggered IntegrationTest matrix has completed. Five downstream groups pass; the three failures are already independently bisected and mapped:

• OrdinaryDiffEq Regression_I (Hairer4/Hairer42): Floor the Newton direction's linear tolerance at sqrt(eps) OrdinaryDiffEq.jl#4301
• OrdinaryDiffEq InterfaceI (exact tstops landing): Correct derivative discontinuity tstop expectation OrdinaryDiffEq.jl#4261
• BoundaryValueDiffEq BigFloat: the exact promote_rule(BigFloat, SparseConnectivityTracer.Dual{...}) ambiguity fixed by Avoid ambiguous reverse promotion in DiffCache PreallocationTools.jl#201

Run: https://github.com/SciML/NonlinearSolve.jl/actions/runs/32368371688. No failure in this integration workflow is caused by the reset-condition diff.

[ChrisRackauckas-Claude]

Follow-up correction: maintainer review rejected changing the inverse-Jacobian solver policy for the separate Wood failure. The broader #1185 is closed; the focused machine-dependent test exclusion is now #1186.

[ChrisRackauckas-Claude]

Rebased onto current master after #1186 merged.

The exact rebased tree was revalidated locally:

GROUP=Core (root)
23 Test Problems: Broyden | 94 pass, 21 broken, 115 total
Testing NonlinearSolve tests passed

GROUP=Core (lib/NonlinearSolveQuasiNewton)
Broyden | 594/594
Klement | 216/216
LimitedMemoryBroyden | 99/99
No-change reset requires every component to stall | 2/2
Testing NonlinearSolveQuasiNewton tests passed

Runic, typos, and git diff --check also pass on the rebased diff. The upstream Wood skip is preserved as skip_tests[alg_ops[2]] = [4, 8]; this PR still tightens the independent Helical problem 5 regression for algorithm 3.

[ChrisRackauckas-Claude]

CI follow-up after rebasing onto current master and adding the deliberate NonlinearSolveBase.jl source touch:

• All six root Core lanes now pass, including both macOS lanes.
• All three NonlinearSolveQuasiNewton Core lanes and its QA lane pass.
• The remaining SimpleNonlinearSolve allocation failure is the already-fixed SciMLBase despecialization regression from Preserve inference for ordinary NonlinearFunction calls SciMLBase.jl#1526; that PR is merged, but this run still resolved a release without the fix.
• The three Integration failures are independent clean-master downstream regressions with focused fixes already available:
  • BoundaryValueDiffEq / BigFloat DiffCache ambiguity: Avoid ambiguous reverse promotion in DiffCache PreallocationTools.jl#201
  • OrdinaryDiffEq InterfaceI derivative-discontinuity tstop expectation: Correct derivative discontinuity tstop expectation OrdinaryDiffEq.jl#4261
  • OrdinaryDiffEq Regression_I Hairer Newton tolerance: Floor the Newton direction's linear tolerance at sqrt(eps) OrdinaryDiffEq.jl#4301
• Chris’s requested handling of the machine-dependent Wood case was merged separately in Skip machine-dependent Wood test for true-Jacobian Broyden #1186.

At this snapshot, only FirstOrder downgrade/Core and the CUDA lanes remain pending; all other NonlinearSolve-owned checks have completed successfully.