Skip to content

v0.2.0 — first tagged release

Choose a tag to compare

View all tags
@jkitchin jkitchin released this 25 May 22:08
· 771 commits to main since this release
v0.2.0
c01b5cb

[0.2.0] — 2026-05-25

First tagged release. The 0.1.0 work-in-progress version was never
tagged; everything below summarizes the state of main as of this
release.

Solver core

  • Full Ipopt-parity C ABI: CreateIpoptProblem, IpoptSolve,
    AddIpoptStrOption / AddIpoptNumOption / AddIpoptIntOption,
    OpenIpoptOutputFile, SetIpoptProblemScaling,
    SetIntermediateCallback, GetIpoptCurrentIterate,
    GetIpoptCurrentViolations, plus a new IpoptSolver session
    handle (IpoptSolverSolve, IpoptSolverResolve,
    IpoptSolverKktSolve, IpoptSolverParametricStep).
  • Restoration phase wired through IpoptSolve with the soft
    restoration line search; nested IPM honors the parent's
    print_iter_output gate.
  • Rapid infeasibility detection in the main loop; convergence
    statuses certified against upstream Ipopt.
  • Option-parity (tier-A waves 1-4): convergence options
    (tol, acceptable_tol, etc.), mu/watchdog/output toggles,
    iteration-output flags, warm-start machinery,
    fixed_variable_treatment, nlp_*_bound_inf,
    barrier_tol_factor, sigma_min / sigma_max for the adaptive
    quality-function oracle.
  • Sensitivity (sIPOPT): Phase D landed — convenience API,
    eigendecomposition, fixed-variable lifting, boundcheck. New
    Solver session API on top: value-typed Factorization handle
    in pounce-linsol enables factor-once / solve-many; Solver
    exposes kkt_solve, parametric_step, and
    compute_reduced_hessian without callback shapes.
  • Presolve crate (pounce-presolve) as an opt-in TNLP wrapper.

Backends and bindings

  • Python (pounce-solver): PyO3 bindings with a cyipopt-style
    Problem class and a scipy-style minimize() facade. The wheel
    bundles the pounce CLI executable.
  • Python session API (pounce.Solver): pyclass that wraps the
    Rust Solver, enabling warm-start sequences (MPC / parametric /
    B&B) and many-RHS sensitivity workflows without the
    callback-based shape.
  • pyomo-pounce (pyomo-pounce): Pyomo SolverFactory plugin
    that drives the pounce CLI on the user's PATH.
  • GAMS link: native solver link (libGamsPounce) for GAMS;
    Jacobian eval skips dense memsets and pure-linear rows.
  • CLI: bundled pounce binary writes AMPL .sol solution
    output; new --about prints version / build / features / paths;
    --dump writes per-iteration KKT artefacts; the sIPOPT
    sensitivity step is folded in.

Linear-solver layer

  • Public Factorization in pounce-linsol: factor once,
    back-solve many RHS, refactor with new values reusing the
    symbolic factor / AMD ordering.
  • MA57 backend (pounce-hsl) honors the linear_solver
    option default ("ma57").
  • Feral backend: cascade-break and FMA default off (opt-in via
    env); near-singular factorizations are flagged via an absolute
    pivot floor; explicit-zero stripping before KKT factor; skips
    refactor on same-matrix back-solve.

Numerical robustness

  • TNLP eval_* user-callback failures surface as NaN instead of
    panicking.
  • Round-off-tolerant Compare_le in the Armijo line-search test.
  • Unconstrained problems routed through the IPM (no degenerate
    paths).
  • push_x_into_interior uses dim() (not values().len()),
    fixing a subtle off-by-one on partially-filled vectors.
  • OrigIpoptNlp::eval_h always uses the h_entry_in_full
    mapping; closes the panic when an entire Hessian row sits on a
    fixed variable.

Benchmarks

  • Composite report (make benchmark →
    benchmarks/BENCHMARK_REPORT.md) covering 9 suites: CUTEst (727
    curated; 1542 full sweep), Mittelmann LP/QP, water-network
    design, gas-network, electrolyte, grid, CHO, large-scale, and
    the GAMS link.
  • Incremental per-suite targets: make benchmark-<suite>
    skips when results.json is fresh; make benchmark-<suite>-rerun
    forces a rebuild.
  • MA57 baseline integrated into the composite report.

Studio & tooling

  • studio/mcp MCP server (pounce-studio-mcp) with
    analyze, run, explain, citations tools and an embedded
    glossary; backed by pounce-studio-core via PyO3.
  • Linear-solver post-mortem aggregated end-to-end and
    surfaced through the studio.

Infrastructure

  • CI workflow with format / clippy / build / test, plus
    wheel-smoke for pounce-solver and pyomo-pounce.
  • mdbook documentation built and deployed to GitHub Pages via the
    new docs.yml workflow.
  • Zenodo metadata (.zenodo.json) and CITATION.cff for
    archival on every GitHub Release.
Assets 2
Loading