ripopt v0.8.2

[0.8.2] - 2026-05-22

Patch release: the pure-Rust feral sparse linear-solver backend is
bumped from 0.4 to its published 0.5.0 crates.io release. No ripopt
source code changed — this is a hardening-and-polishing release for the
linear-solver layer. The full benchmark suite was re-run to confirm the
bump is regression-free: ripopt's strict-Optimal count is unchanged at
568/745 (CUTEst 551/727), and no problem crossed the Optimal boundary
in either direction.

Workspace version bumps: ripopt 0.8.1 → 0.8.2, ripopt-py 0.8.1 → 0.8.2, pyomo-ripopt 0.8.1 → 0.8.2. rmumps held at 0.1.2 (no rmumps
changes) and Ripopt.jl held at 0.8.0 (no Julia-affecting C-ABI
changes).

Changed

• feral 0.4 → 0.5.0 (crates.io). The local [patch.crates-io]
  override that had tracked the pre-release feral checkout is removed;
  ripopt now builds against the published 0.5.0 crate.

Fixed

• MC64 scaling-cache correctness, via feral 0.5.0 (feral #47, #49).
  feral's B2 value-bounded scaling cache no longer populates on
  non-MC64 scalings (InfNorm, user-supplied External), and
  pick_scaling_strategy is now value-aware so explicit stored-zero
  KKT (2,2)-block entries no longer mis-route a constraint column to
  MC64. The defect was latent — a stale InfNorm vector could be
  replayed as an injected External scaling on a drifted iterate,
  benign on every measured ripopt KKT but unsound. The fix shifts
  ripopt's numerical path on ~12 CUTEst problems; the net effect on the
  strict-Optimal count is zero, with FEEDLOC notably improving
  (2193 → 70 iterations).

Performance

• feral Schur trailing-update kernel widened (feral #44) to a quad
  NEON-tile inner loop. The gain is concentrated on large dense Schur
  updates; ripopt's CUTEst corpus is capped at n ≤ 100, too small for
  the widened tile to engage, so suite-level timing is unchanged within
  measurement noise.

Notes

• This release closes the v0.8 linear-solver track: the feral backend
  is considered solid. The remaining performance and robustness gaps
  versus Ipopt 3.14 — concentrated at large scale and on the
  RestorationFailed tail — are IPM-algorithm gaps, not linear-solver
  gaps.