DiffEqBase-v7.21.2

DiffEqBase DiffEqBase-v7.21.2

Diff since DiffEqBase-v7.21.1

Merged pull requests:

• Reuse the finite-difference JVP base evaluation across a linear solve (#4287) (@ChrisRackauckas-Claude)
• Add adaptive step size control to SDC (#4346) (@singhharsh1708)
• Fix SBDF order ramp, history-shift timing, and dt-change reset (#4366) (@niy-ati)
• Despecialize callbacks across solver specialization levels (#4376) (@ChrisRackauckas-Claude)
• Do not seed a retry from the previous-step interpolant (#4473) (@singhharsh1708)
• NonlinearSolveAlg: stale-Jacobian retry + rejected-step W reassembly (carries #4282 to current master) (#4505) (@ChrisRackauckas-Claude)
• Allow JET 0.12 in every QA test environment (#4513) (@ChrisRackauckas-Claude)
• Resolve OrdinaryDiffEqDefault against the local sublibraries (#4514) (@ChrisRackauckas-Claude)
• OrdinaryDiffEqDefault: require OrdinaryDiffEqCore 4.17.1 (fixes registry precompile failure) (#4520) (@ChrisRackauckas-Claude)
• Fix baseline spelling and whitespace checks in tests (#4524) (@ChrisRackauckas-Claude)
• Require OrdinaryDiffEqCore 4.17.1 where its new internals are used (#4525) (@ChrisRackauckas)
• Carry the QNDF2 difference array across steps (#4526) (@ChrisRackauckas-Claude)
• Say why the linear path ignores new_W_dt_cutoff (#4527) (@ChrisRackauckas-Claude)
• NonlinearSolveAlg: reuse a split W (LHLFactorization) as the operator (#4528) (@ChrisRackauckas-Claude)
• Fix QNDF/QBDF accept-path step and order selection for backward solves (#4529) (@ChrisRackauckas-Claude)
• Apply the dtmin floor to backward-in-time solves in fix_dt_at_bounds! (#4530) (@ChrisRackauckas-Claude)
• Select extrapolation order on |dt| in the Hairer-Wanner controller (#4531) (@ChrisRackauckas-Claude)
• Invalidate the cached factorization in the in-place Deuflhard order extension (#4532) (@ChrisRackauckas-Claude)
• Do not re-factorize a reused W in the in-place RadauIIA3 step (#4533) (@ChrisRackauckas-Claude)
• Pass Θ = 0, not the absolute time, to the AN5 startup interpolant (#4534) (@ChrisRackauckas-Claude)
• Make the stabilized-RK step cap and eigenvalue tolerance direction-agnostic (#4536) (@ChrisRackauckas-Claude)
• Accept a callable stage predictor on the nlsolve algorithms (#4539) (@ChrisRackauckas-Claude)
• NordsieckBDF: rebase dense output when a callback truncates the step (#4540) (@ChrisRackauckas-Claude)
• Fix Runic formatting and typos failures on master (#4541) (@ChrisRackauckas-Claude)
• Stop tying the state Jacobian's finite-difference direction to tdir (#4543) (@ChrisRackauckas-Claude)
• Release OrdinaryDiffEqExponentialRK 2.4.0 (#4545) (@ChrisRackauckas)
• Release OrdinaryDiffEqFIRK 2.8.7 (#4546) (@ChrisRackauckas)
• Extend OrdinaryDiffEqCore.dtnew_modification instead of shadowing it (#4547) (@ChrisRackauckas-Claude)
• Release OrdinaryDiffEqNonlinearSolve 2.9.7 (#4548) (@ChrisRackauckas)
• Pick the DAE initialization dt by magnitude, not signed min (#4549) (@ChrisRackauckas-Claude)
• Apply stage limiters in every StabilizedRK method (#4550) (@ChrisRackauckas-Claude)
• Run the QA test group last in every sublibrary that ran it first (#4551) (@ChrisRackauckas-Claude)
• Release OrdinaryDiffEqNonlinearSolve 2.9.8 (#4552) (@ChrisRackauckas)
• Declare OrdinaryDiffEqCore.dtnew_modification public (#4553) (@ChrisRackauckas-Claude)
• Raise OrdinaryDiffEqNonlinearSolve floor to 2.9.8 for stage_predictor (#4554) (@ChrisRackauckas-Claude)
• Fix the sub-cache used by in-place interpolation for alg_choice 2 and 4 (#4556) (@utkuyilmaz1903)
• Release OrdinaryDiffEqCore 4.17.3 (#4557) (@ChrisRackauckas)
• Release OrdinaryDiffEqBDF 2.4.10 (#4558) (@ChrisRackauckas)
• only allow diagonal M for ESDIRK (#4560) (@oscardssmith)
• Release OrdinaryDiffEqCore 4.17.4 (#4562) (@ChrisRackauckas)
• Release OrdinaryDiffEqDifferentiation 3.12.1 (#4563) (@ChrisRackauckas)
• Release OrdinaryDiffEqSDIRK 2.9.6 (#4564) (@ChrisRackauckas)
• Release OrdinaryDiffEqDifferentiation 3.12.2 (#4565) (@ChrisRackauckas)
• Release OrdinaryDiffEqRosenbrock 2.7.4 (#4566) (@ChrisRackauckas)
• Tolerate roundoff-scale dt differences in SBDF dt-change reset (#4574) (@ChrisRackauckas-Claude)
• Fix NordsieckBDF trial-step interpolation for delay equations (#4575) (@ChrisRackauckas-Claude)
• Release OrdinaryDiffEqBDF v2.4.11 (#4576) (@ChrisRackauckas)
• Release OrdinaryDiffEqLinear v2.3.1 (#4577) (@ChrisRackauckas)
• Release OrdinaryDiffEqStabilizedRK v2.7.1 (#4578) (@ChrisRackauckas)
• Require the FiniteDiff release that scales the JVP step (#4583) (@singhharsh1708)
• Release DiffEqBase v7.21.2 (#4584) (@ChrisRackauckas)
• Release DelayDiffEq v6.4.1 (#4585) (@ChrisRackauckas)

Closed issues:

• NonlinearSolveAlg: extend W-matrix reuse to nlstep_data (MTK) cases (#3696)
• SDIRK smooth_est: W⁻¹-filtered error estimate underestimates VDP local error by ~3 orders; silently disabled under NonlinearSolveAlg (isnewton gate) (#3821)
• Tridiagonal jac_prototype crashes in build_J_W (fill! on structured matrix) — regression from v6.87 (#3937)
• AutoFiniteDiff JVP recomputes f(x) on every Krylov matvec; nf accounting double-counts it (#4162)
• log_numerical_instability scalar-indexes CUDA state in WeakAdaptiveGPU (#4217)
• SBDF solvers do not achieve expected convergence order (#4329)
• QNDF2 and QBDF2 reject many more steps than expected (#4332)
• Predictor.MaxOrder drives dt below eps on several SDIRK methods (#4472)
• OrdinaryDiffEqBDF QA (JET) fails on master: FBDF time_filter calls undefined functions in OOP path (#4518)
• OrdinaryDiffEqDefault 2.6.1 under-bounds OrdinaryDiffEqCore (uses _is_identity_massmatrix from Core 4.17.1, allows 4.11) (#4522)
• OrdinaryDiffEqStabilizedRK shadows dtnew_modification instead of extending it, so the stability-limit dt cap never runs (#4535)
• solve(prob) is not time-reversible: finite-difference Jacobians take their perturbation direction from diffdir (#4537)
• DAE initialization dt picks the larger magnitude for backward solves; mass-matrix FBDF loses time-reversal symmetry (#4538)
• Pkg.test("OrdinaryDiffEqDifferentiation") fails because JET runs before the functional tests in the same process (#4542)
• In-place interpolation uses the wrong sub-cache for alg_choice 2 and 4 with the default algorithm (#4555)
• OrdinaryDiffEqBDF.jl v2.4.10 changed behavior significantly (#4573)