solverforge 0.25.x: Literal Configuration, Framework Partitioned Search, and Automatic Parallel Evaluation

solverforge 0.25.x follows solverforge 0.24.0; the line was published on crates.io on 2026-11-09. The first release is 0.25.0, with a Rust 1.95 floor, and the line closes with 0.25.1. At publication cargo add solverforge installed 0.25.0, and docs.rs served the 0.25.0 API. solverforge-cli 3.9.0 scaffolds target 0.25.0 and solverforge-cli 4.0.0 and 4.0.1 scaffold target 0.25.1; existing applications keep their pins until they upgrade.

What Changed

Termination means what the configuration says — breaking changes

• Every configured solver-level criterion applies, combined with OR: the first one reached stops the solve. A configuration with several criteria can stop earlier than before.
• Nothing adds a time limit the configuration does not set. A solve with only best_score_limit, step_count_limit, or an unimproved limit is no longer capped at 30 or 60 seconds.
• seconds_spent_limit = 0 and minutes_spent_limit = 0 are a zero time limit, not “no limit”; time limits are literal and saturate instead of overflowing.
• The solverforge.execution_policy candidate-trace header carries one attribute per criterion, and termination_composition names every installed criterion (for example time_or_step_count). The criterion and criterion_target attributes are removed.
• try_run_solver_with_config_and_search, try_run_weighted_solver_with_config_and_search, and try_run_dynamic_solver_with_config_parts lose their default_time_limit_secs parameter.

See Termination.

Parallel evaluation is automatic — breaking change

The default move_thread_count = "auto" evaluates local-search steps on up to four evaluators, choosing per step between sequential evaluation and parallel windows from measured trial and coordination costs. A parallel window forks up to three working-solution replicas. "none" keeps evaluation sequential. { count = N } needs N >= 1: MoveThreadCount::Count and ThreadCount::Specific hold a NonZeroUsize, and { count = 0 } is a configuration error. The thread policy is resolved when the local-search phase solves, on the pool that runs it.

See Local Search.

Environment modes seed and check — breaking changes

• reproducible, fast_assert, and full_assert without random_seed solve with the fixed seed REPRODUCIBLE_RANDOM_SEED; only non_reproducible draws an operating-system seed.
• fast_assert panics on score drift at committed steps and best-solution updates.
• full_assert panics on drift in any score-director calculation, including move evaluations and evaluator replicas. Director gains the provided method set_score_assertion.

Partitioned search runs in the framework — breaking changes

SearchBuilder::partitioned_phase is removed. Register a SolutionPartitioner with SearchBuilder::partitioner; the framework builds and runs the phase from the partitioned_search configuration, and the phase termination bounds every partition. SolutionPartitioner::recommended_partition_count returns Option<NonZeroUsize>, FunctionalPartitioner::with_recommended_count takes a NonZeroUsize, ThreadCount::resolve takes the recommendation as its second argument, and auto threads are bounded by the recommended count. A partitioned_search termination with best_score_limit, unimproved_step_count_limit, or unimproved_seconds_spent_limit is a configuration error.

Configuration that used to be ignored is rejected — breaking changes

• group_candidate_limit without group_name.
• value_candidate_limit, group_candidate_limit, or a non-default construction_obligation on a list construction heuristic.
• simulated_annealing with both calibration and level_temperatures.
• step_counting_hill_climbing.step_count_limit = 0.
• A non-positive or non-finite great_deluge.water_level_increase_ratio.
• A score_director.constraint_provider that names no constraints or constraint_provider path declared on the solution (RuntimeBuildError::UnknownConstraintProvider).

A fixed-weight union needs at least one positive weight; a zero weight disables its child instead of failing the configuration.

Score director and analysis — breaking changes

score_director.constraint_provider selects which declared constraint set a solve uses, and score_director.constraint_match_enabled defaults to true. Setting it to false disables SolverManager::analyze_snapshot for the job, which returns the new SolverManagerError::SnapshotAnalysisDisabled. SolverRuntime::bind_snapshot_analysis takes Option<fn(&S) -> ScoreAnalysis<S::Score>>, and try_run_dynamic_solver_with_config_parts takes &[DynamicConstraintProvider].

Runtime overrides

SolverConfig::with_override applies a SolverConfigOverride to a configuration, and SolverManager can run a retained job under an override. See SolverManager.

Search

• VND phases honour score_tie_break, whose default is random; set score_tie_break = "first" to apply the first of several equally scored best improving moves.
• list_k_opt construction accepts k from 2 to 5, and polishes routes with k-opt for k from 3 to 5.

Scoring — breaking changes

• SharedOperatorTerminal requires its operator to implement GroupedOperator, and changed_key_count sums changed group keys instead of counting row-changing notifications.
• A Projection that emits more rows for one input than its MAX_EMITS panics; raise MAX_EMITS to the real maximum. Row lists are sized by MAX_EMITS.
• ConstraintWeight is documented and exported at the crate roots beside FixedWeight and HardWeight.
• A score director can fork over another solution.
• Scoring production code forbids unsafe.

Performance

The 0.25.1 patch keeps the nearest destinations of the nearby list change and swap cursors in one bounded buffer per cursor, described below. In 0.25.0, construction reuses its candidate and value buffers across placements, a solver runner keeps its declaration only when it runs partitioned search, and the global Rayon pool is no longer initialized while a solver compiles.

0.25.1: bounded nearby list candidates

The nearby list change and swap cursors keep only the max_nearby nearest destinations of a source position in one buffer each cursor owns, ordered by ascending distance with offer order between equal distances. That is the same set and order the earlier sort-and-truncate produced, so candidate streams are unchanged; the buffer is reused across source positions and moves read from it directly, with no per-position allocation (#228). The public API is unchanged.

Published Version Boundaries

Surface	Version	Notes
Core runtime	solverforge 0.25.0	Rust 1.95; literal termination, automatic parallel evaluation, framework partitioned search, stricter configuration
Core patch	solverforge 0.25.1	Bounded nearby list candidate buffer; API unchanged; scaffolded by solverforge-cli 4.0.0 and 4.0.1
Previous core runtime	solverforge 0.24.0	Explicit configuration and constraint weights; see the 0.24.x release note
CLI	solverforge-cli 3.9.0	Fresh scaffolds target core 0.25.0
Previous CLI	solverforge-cli 3.8.3	Scaffolds core 0.24.0; existing projects retain their pins
Python	solverforge-py 0.7.2 / PyPI solverforge 0.7.2	Still compiled onto core 0.23.0; upgrading the Rust core does not retarget these wheels
Worked use cases	November 9 app releases	All eight target core 0.25.0 and UI 0.9.0; see the use-case 3.0 release note
UI	solverforge-ui 0.9.1	Unchanged
Maps	solverforge-maps 2.3.3	Unchanged

All nine workspace crates (solverforge, solverforge-core, solverforge-config, solverforge-scoring, solverforge-solver, solverforge-macros, solverforge-bridge, solverforge-console, and solverforge-cvrp) publish 0.25.0.

The registry records the facade crate’s publication at 2026-11-09T03:06:41Z. Its artifact SHA-256 is 4a3ecc7e869c19a21e06386736824f230ad106941188b861424389556456737c; .cargo_vcs_info.json binds it to 3115b4dd6b40bc72a472b042f5e6b665cb9720c7, the v0.25.0 tag commit. The GitHub Release was published at 2026-11-09T03:06:58Z.

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.

All nine workspace crates publish 0.25.1. The registry records the facade crate’s publication at 2026-11-09T17:38:50Z; its artifact SHA-256 is c899657c17ae3a41d67b583ec400e5d79fe253f0ed62a38a2efcb5d04d1171bb, and .cargo_vcs_info.json binds it to 6069763dffdf57786b703f113e5d290f3695ed02, the v0.25.1 tag commit. On that commit the CI run passed, and 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 at 2026-11-09T17:39:07Z; every job succeeded.

Patch History

Version	Date	Notes
0.25.1	2026-11-09	The nearby list change and swap cursors keep the nearest destinations in one bounded buffer per cursor; candidate streams and API unchanged.
0.25.0	2026-11-09	Every termination criterion applies with no hidden time limit; automatic measured parallel evaluation; fixed seeds and stronger checks in the assert modes; partitioned search run by the framework; ignored configuration rejected; snapshot analysis gated on constraint_match_enabled; configuration overrides.

Upgrade Checklist

• Read every [termination] table: all criteria now apply, a work or score criterion without a time limit can run indefinitely, and a zero time limit stops at once.
• Replace move_thread_count = { count = 0 }, and set move_thread_count = "none" where a run must stay sequential.
• Set random_seed explicitly if a reproducible or assert-mode run relied on a fresh seed each time.
• Move partitioned search to SearchBuilder::partitioner and the partitioned_search configuration; pass NonZeroUsize partition counts.
• Remove the configuration combinations listed above, and give a union with fixed weights at least one positive weight.
• Handle SolverManagerError::SnapshotAnalysisDisabled in exhaustive matches.
• Raise a Projection’s MAX_EMITS to its real maximum, and read changed_key_count as a count of changed group keys.
• Add score_tie_break = "first" to a VND phase that must keep the earlier choice among equal best moves.

Where to read next

Read the core manual and the solver configuration guide. The full compare covers all 54 commits of 0.25.0, the 0.25.1 compare covers the patch, and the CHANGELOG lists every breaking change.