Relation to pathwise

pathwise is a separate crate with generic search and optimization building blocks: a Problem trait (state, moves, cost or score) and strategies that run against it interchangeably, such as A*, branch and bound, local search and simulated annealing. unifier depends on it, but uses much less of it than the name “built on pathwise” might suggest.

What unifier uses

Two CSP-independent coordination primitives:

In pathwise	In unifier	Purpose
pathwise::core::cancellation::CancellationToken	re-exported as unifier::CancellationToken	A cloneable flag to stop a running search from another thread.
pathwise::core::incumbent::SharedIncumbent	wrapped by unifier::solver::SharedIncumbent	The best solution found so far, shared between portfolio workers.

SharedIncumbent in unifier is a thin wrapper that specializes the generic type to unifier’s assignment (HashMap<VariableId, i64>) and score (HardSoftScore). Using both from pathwise avoids keeping a second copy of the same logic.

What unifier does itself

unifier does not implement pathwise’s Problem/OptimizationProblem traits and does not route its search through pathwise’s generic algorithms. Its solvers are specialized for constraint models:

• constraint propagation, with generalized arc consistency for AllDifferent (Régin’s matching and SCC algorithm), overload detection for Cumulative and NoOverlap, edge-finding for NoOverlap, and AC-3 otherwise;
• dom/wdeg and MRV variable ordering;
• reversible domains (checkpoint and undo) instead of cloning the domains at every search node;
• Branch & Bound with its own optimistic-bound pruning over hard/soft scores.

A generic state-space search cannot see domains or constraints, so it cannot prune with propagation. That is why this layer lives in unifier rather than in pathwise.

In practice

You don’t need to use pathwise directly. unifier::CancellationToken is the type you pass in SolverOptions, and SharedIncumbent is set up for you by ParallelSolver.