v1.5.0

Added

• New documentation site at libqif.chatzi.org, with one page per module.
• qif.lp: HiGHS solver.
• method argument in channel.factorize, channel.left_factorize ("lp", "subgrad") and metric.kantorovich ("lp", "fastemd").
• bayes_vuln.mult_capacity_bound_cap in python.
• metric.manhattan in python, rat support in metric.l1, total_variation, convex_separation(_quasi), kantorovich, from_distance_matrix and to_distance_matrix.

Changed

• metric.euclidean(uint) and metric.discrete(uint) return distances of the default type (rat after set_default_type(rat), previously always double).
• qif.lp: the default solver is CLP (instead of GLOP), and HiGHS for interior point.
• qif.metric: the combinators work for metrics on any domain.
• channel.hyper removes zero-probability columns, bayes_risk.mult_leakage returns inf for zero posterior risk.
• Python: more consistent argument names.
• Upgraded Armadillo to 15.6.2 and OSQP to 1.0.0. CMake 4 works without extra flags.
• Tests are in python (tests/, pytest), tests_cpp only covers C++ internals (LP/QP solvers, equal).

Removed

• C++: refinement::refined_by(A, B, G), use refined_by(A, B, G, R) instead.
• C++: metric::threshold_bin.

Fixed

• mechanism.l_risk ignored n_guesses.
• Windows: functions using linear system solving (channel.factorize with the subgradient method, left_factorize for large channels, mechanism.d_privacy.tight_constraints) failed, the wheels were built without LAPACK.
• rat infinity (e.g. bayes_risk.mult_capacity of channels with zero risk) was 2^31-1 on Windows, 2^63-1 elsewhere.
• channel.left_factorize failed for matrices with 1000 or more elements (and channel.factorize with col_stoch using the subgradient method).
• channel.iterative_bayesian_update and channel.posteriors silently computed in double for rat arguments, unless all optional arguments were given.