solverforge 0.26.x: Replica Directors Without Descriptor Types, Proportional Debug Info, and Late Acceptance That Keeps Its History

solverforge 0.26.x follows solverforge 0.25.1; the line was published on crates.io from 2026-11-10 through 2026-11-11 and closes with 0.26.1, with a Rust 1.95 floor. At publication cargo add solverforge installed 0.26.1, and docs.rs served the 0.26.1 API. The public modeling, stream, configuration, and solving APIs are unchanged across the line; the one breaking change, in 0.26.0, is in the lower-level solverforge-scoring score-director constructors. 0.26.1 changes no public API: late acceptance keeps its score history unless a hard or medium level of the best score changes, and scalar ruin-and-recreate draws its batches without an entity-sized permutation. solverforge-cli 4.0.4 scaffolds 0.26.1, and all eight worked use cases pin it.

What Changed

Late acceptance keeps its score history (0.26.1)

late_acceptance keeps its score history across a new best score unless a hard or medium level of the best score changes: the hard level of a hard/soft score, the hard and medium levels of a hard/medium/soft score, or a hard level of a bendable score. Only then is the history refilled with the new best, so entries from before a feasibility change cannot admit moves that keep the new hard level but give up the soft levels. An improvement in soft levels alone leaves the history as it is, and a score with only soft levels never refills it.

Through 0.26.0 the acceptor refilled the whole history on every new best score. After the first improvement every entry equalled the best, so only moves at least as good as the best were accepted: from then on the phase ran as hill climbing with sideways moves, and searches that rely on accepting worse moves to leave a local optimum settled early. The configuration is unchanged: late_acceptance_size still sets the history length, and diversified_late_acceptance is unaffected.

The fix was contributed by slapers in #242, together with a follow-up that refills the history of a bendable score only on its hard levels.

Scalar ruin batches without an entity-sized permutation (0.26.1)

ruin_recreate_move_selector draws each batch of entities to clear from a partial shuffle that records only the positions it displaces, so a batch costs time and memory proportional to the ruin count instead of the entity count, and opening the move cursor no longer resets an entity-sized permutation for every batch. The batches and the random draws they consume are unchanged, so a seed selects the same entities as before. In the zero-regression gate’s ruin case, a 15-trial run against 0.26.0 measured 41.8% less wall time and 55.5% fewer instructions (#249).

Replica-forking directors — breaking change (0.26.0)

solverforge-scoring removes ReplicaDescriptor. A score director that forks evaluator replicas is built with ScoreDirector::with_replicas, with_replicas_weighted, or with_replicas_in_mode, which take the constraint-set constructor beside a borrowed descriptor:

ScoreDirector::with_replicas(
    solution,
    constraints,            // normally constraint_builder(&solution)
    constraint_builder,     // fn(&S) -> C
    &descriptor,            // &SolutionDescriptor
    entity_counter,         // fn(&S, usize) -> usize
)

The constructor is a field of the director instead of part of its descriptor type, so a replica-capable director is ScoreDirector<S, C, &SolutionDescriptor> and names its constraint set once. DirectorDescriptor has no type parameters, and DirectorDescriptor::replica() is replaced by share(). Forking with Director::fork_replica() and Director::fork_for_solution(), and rebuilding the constraint set on reset(), behave as before; a director not built by with_replicas* returns None from both forks.

Macro-generated solves and the configured entry points build their directors themselves, so application code that does not construct a ScoreDirector directly needs no change. See Extend the Solver.

Debug info stays proportional to the constraint set

The stream combinators build the closures they store (filters, weights, extractors, joined-row adapters) in functions generic only over what each closure captures. A closure built inside a generic method used to name every upstream stream type, so each chained filter, penalize, or named repeated the whole stream type, and debug-info type names grew with the length of every constraint chain. The public stream API is unchanged.

On a 30-constraint application test build, the line-tables debug strings drop from 2.82 GB to 0.37 GB and the longest type name from 8.0 MB to 1.0 MB, and the test binary links and passes with full debug info.

Generated items are documented or hidden

Every public item the planning macros generate carries documentation or is #[doc(hidden)]: descriptor and entity-count plumbing is hidden, and the collection source methods and the generated <Solution>ConstraintStreams trait and its methods are documented. A crate that enables missing_docs gets no warnings from the macros (#235).

Module roots hold declarations only

Every mod.rs and lib.rs in the workspace holds only module declarations and re-exports, and each crate’s lint levels live in its Cargo [lints] table. Public paths and behaviour are unchanged.

Published Version Boundaries

All nine workspace crates (solverforge, solverforge-core, solverforge-config, solverforge-scoring, solverforge-solver, solverforge-macros, solverforge-bridge, solverforge-console, and solverforge-cvrp) publish 0.26.0 and 0.26.1.

The registry records the facade crate’s publication at 2026-11-10T10:33:40Z. Its artifact SHA-256 is 62181b0bb54c94833633b2b69cb0592d98d4532e0ba204f7cce828966cc0f773; .cargo_vcs_info.json binds it to 720fdaded42f16a131cdb8820b624b407b275d5f, the v0.26.0 tag commit. The GitHub Release was published at 2026-11-10T10:33:56Z.

On that commit the CI run passed. The Release workflow re-ran the CI matrix on Linux, macOS, and Windows with the Miri, MSRV, docs, and package jobs, verified the release, published to crates.io, and created the GitHub Release; every job succeeded.

The registry records the 0.26.1 facade publication at 2026-11-11T02:49:52Z. Its artifact SHA-256 is a00f06d8db244d619f556a25165115015bda50f01c811db0d515c9e6e3c28458; .cargo_vcs_info.json binds it to f3b41788e7519f3b707f8a9aa450fc545e0d5c08, the v0.26.1 tag commit. The GitHub Release was published at 2026-11-11T02:50:11Z. On that commit the CI run passed, and every job of the Release workflow succeeded: the Linux, macOS, Windows, Miri, MSRV, docs, and package jobs, release verification, crates.io publication, and the GitHub Release. The 0.26.1 range also changes the zero-regression gate’s own tooling (timed trials rotate through the code layouts, and hardware branch misses are reported rather than graded); no published crate behaviour depends on it.

Patch History

Upgrade Checklist

• From 0.26.0, move the solverforge pin to 0.26.1; no code changes. A late_acceptance phase now keeps accepting worse moves after its first improvement, so re-check solve quality and any tuned late_acceptance_size or termination budget on models that use it.
• Replace ScoreDirector::with_descriptor(solution, constraints, ReplicaDescriptor::new(&descriptor, constraint_builder), entity_counter) with ScoreDirector::with_replicas(solution, constraints, constraint_builder, &descriptor, entity_counter), and the weighted or mode-generic forms with with_replicas_weighted or with_replicas_in_mode.
• Spell a replica-capable director’s type as ScoreDirector<S, C, &SolutionDescriptor> (plus the weight mode, if not the default) instead of naming ReplicaDescriptor.
• Drop the type parameters from DirectorDescriptor bounds, and call share() where a custom descriptor implementation or caller used replica().
• Application code that only uses the facade, the planning macros, and the configured solve entry points upgrades by moving the solverforge pin.

Where to read next

Read the core manual, Local Search, and Extend the Solver. The 0.26.0 compare covers its 8 commits and the 0.26.1 compare its 8 commits; the CHANGELOG lists the breaking change and both releases.