discopt v0.3.0

Highlights

discopt 0.3.0 is a feature-heavy release that broadens the solver surface and adds two entirely new modeling capabilities on top of the v0.2.5 MINLP core:

• discopt.mo -- multi-objective optimization via scalarization (weighted sum, AUGMECON2 ε-constraint, weighted Tchebycheff, NBI, NNC) with a ParetoFront container and hypervolume / IGD / spread / ε-indicator quality metrics. Every scalarized subproblem inherits the full MINLP / differentiable / global-solver stack.
• discopt.doe -- model-based design of experiments. Identifiability, estimability, and profile-likelihood diagnostics; five model-discrimination criteria (HR, BF, JR, MI, DT) with sequential-design and Akaike-weight model selection; batch / parallel experimental design.
• AMP -- Adaptive Multivariate Partitioning global MINLP solver with the soundness guarantee LB_k <= global_opt <= UB_k at every iteration.
• SUSPECT-style convexity detector with sound certificates feeding the convex NLP fast path, tighter αBB underestimators, and MO scalarization fast paths.
• Claude Code skills + CLI installer: 20 expert agents, plus discopt install-skills to drop them into ~/.claude/skills/.
• ripopt 0.6.1 → 0.7.0 with a number of correctness fixes (HiGHS LP/QP false-optimality on wide bounds, single-solve starting points, RLT cuts on bilinear terms with no auxiliary index, residuals using all array observations).

The companion manuscript at manuscript/discopt.org is updated with sections covering the convexity certifier, AMP, the DOE diagnostics, and multi-objective optimization, exported to manuscript/discopt.pdf (36 pages).

This release skips the never-tagged 0.2.6 and folds its draft entries into 0.3.0. The full changelog is below.

---

[0.3.0] - 2026-04-24

This release skips the never-tagged 0.2.6 and folds its draft entries into 0.3.0 along with the post-0.2.6 feature and infrastructure work.

Added

• discopt.mo -- multi-objective optimization (feat(mo): multi-objective optimization via scalarization). Weighted-sum, AUGMECON2 ε-constraint, weighted-Tchebycheff, NBI, and NNC scalarizations; ideal/nadir payoff-table utilities; ParetoFront container; hypervolume / IGD / spread / ε-indicator quality metrics under discopt.mo.indicators.
• discopt.doe -- model-based design of experiments. Identifiability + estimability + profile-likelihood analysis (feat(doe): identifiability + estimability + profile likelihood, #48); model discrimination criteria + selection + sequential-design loop (feat(doe): model discrimination, #49, #50); batch / parallel experimental design (feat(doe): batch / parallel experimental design).
• AMP -- Adaptive Multivariate Partitioning global MINLP solver (feat(amp), #44). Iterates MILP relaxation -> NLP subproblem -> partition refinement with the soundness guarantee LB_k <= global_opt <= UB_k at every iteration.
• SUSPECT-style convexity detector with sound certificates (#46). Structural convexity / concavity / monotonicity proofs for use by the convex NLP fast path and discopt.mo reformulations.
• Claude Code skills + CLI installer (feat(cli): ship Claude Code skills in package + discopt install-skills, feat(skills): 20 discopt feature / algorithm expert agents). 20 expert agents shipped in the package and installable into a user's ~/.claude/skills/ via discopt install-skills.
• Crucible knowledge base tracked in git (feat(crucible): track wiki, bib, and 3 new articles in git).
• Zenodo metadata and refined manuscript sections (#47).
• RELEASE.md -- authoritative release checklist documenting the procedure for cutting a discopt release.
• CHANGELOG.md in Keep a Changelog format.
• Local cargo-fmt pre-commit hook so Rust formatting is enforced alongside ruff and mypy.

Changed

• ripopt workspace dependency 0.6.1 -> 0.7.0 (via 0.6.2; Cargo.toml, Cargo.lock). The 0.6.2 step transitively updated rmumps 0.1.0 -> 0.1.1; the 0.7.0 step adapted crates/discopt-python/src/ripopt_bindings.rs to the new NlpProblem trait signatures: evaluation methods (objective, gradient, constraints, jacobian_values, hessian_values) now take an explicit new_x: bool flag and return bool (success / evaluation-failure), matching Ipopt's TNLP contract. Added match arms for the new SolveStatus::Acceptable, SolveStatus::EvaluationError, and SolveStatus::UserRequestedStop variants, surfaced as "acceptable" / "evaluation_error" / "user_requested_stop" on the Python side; acceptable maps to SolveStatus.OPTIMAL (KKT residuals within Ipopt's relaxed-acceptable-level tolerances).
• _solve_continuous (pure-continuous NLP fast path) now promotes the default nlp_solver="ipm" to "ipopt" for single-problem solves. The pure-JAX IPM's acceptable-tolerance check only covers variable-bound complementarity, so on problems with unbounded variables plus inequality constraints it could terminate at a non-KKT point and report OPTIMAL. Ipopt is more reliable for single solves; the JAX IPM remains the default for B&B subproblems.
• differentiable_solve and differentiable_solve_l3 default backend changed from "ipm" to "ipopt" for the same reason.
• solver now routes pure-MILP problems through HiGHS MIP with a B&B fallback (fix(solver): route MILP through HiGHS MIP with B&B fallback).
• DAE collocation perf (perf(dae): vectorize collocation and fix sparse Jacobian for NMPC warm solves). Vectorized collocation residuals; sparse Jacobian assembly fixed so NMPC warm-start solves don't densify.
• manuscript/discopt.tex is no longer tracked -- it is generated from manuscript/discopt.org.

Fixed

• Jupyter Book docs build with zero warnings. Cleaned up RST-formatting issues in module docstrings across benchmarks/, modeling/, ro/, solvers/, doe/, and mo/; suppressed autoapi import-resolution warnings for the compiled discopt._rust extension; escaped **kwargs parameter entries to keep Sphinx from parsing the leading ** as inline strong.
• HiGHS LP/QP false optimality on wide bounds: solvers/qp_highs.py and solvers/lp_highs.py now clip any bound with magnitude >= 1e15 to highspy.kHighsInf before passing to HiGHS. Bounds like discopt's default +/-9.999e19 fall just below HiGHS's internal infinity threshold (1e20) and caused HiGHS to return false-optimal solutions on convex QPs with unbounded variables.
• Single-solve starting point: _solve_continuous now clips the default starting point to +/-10 (respecting actual bounds) instead of the previous +/-100, preventing ipopt from exploding on exp/log NLPs with one-sided large bounds.
• Stationary-point starting point: Fully unbounded variables (|lb| > 1e15 and |ub| > 1e15) now start at 0.5 instead of the midpoint of 0. Zero is a stationary point of periodic functions (sin, cos) and even functions generally; starting at 0.5 lets first-order NLP methods pick a descent direction and escape local maxima of the objective. Same fix applied in _jax/differentiable.py::_safe_x0.
• _solve_qp_highs and _solve_qp_jax now set SolveResult.convex_fast_path = True when solving a detected convex QP directly, matching the semantics of the convex NLP fast path.
• Cutting planes with bilinear terms (#35): _jax/cutting_planes.py::generate_rlt_cuts no longer emits unsound inequalities when a bilinear term has no auxiliary w_index.
• fix(estimate): discopt.estimate now uses all array observations in residuals and Fisher information, instead of dropping all but the first row.
• fix(ci): cleared a clippy collapsible_match and repaired the T24 vmap path after the MILP rerouting change.

---

Full changelog: v0.2.5...v0.3.0