SolverForge
Search docs Ctrl K

releases

solverforge 0.22.x: The Canonical Grouping Kernel

The 0.22.x line routes every grouped stream through one accumulator kernel, retracts group and hash memberships by position, owns each grouping key once, and drops stale key Clone bounds in complement weights.

solverforge 0.22.x follows solverforge 0.21.3; the line was published on crates.io on 2026-10-18. cargo add solverforge installs 0.22.0 and docs.rs serves the 0.22.0 API. The Rust 1.95 floor remains. The 0.21.x line’s joined-tuple grouping, live-solution lending, owned search state, and net-route-delta precedence scoring carry into this line unchanged.

What Changed

One grouping execution path

Every grouped stream — unary groups, cross-join groups, projected groups, and every complemented variant — lowers onto the same GroupNode accumulator in stream::relational::operator. Earlier releases kept per-family grouped engines; they are gone. Each group keeps an explicit contributor list with per-input accumulator tokens, so a full evaluation publishes exactly one replacement row per group and an incremental update touches only the groups its inputs belong to. #[solverforge_constraints] compiles chained grouped terminals against one shared terminal owner and scores, orders, and explains each terminal independently.

The macro’s grouped-sharing recognition for complemented sources no longer requires project(...) or join(...) earlier in the chain: a plain group_by chain qualifies. SharedNodeDiagnostics, SharedNodeId, and SharedNodeOperation remain the shared-node diagnostic vocabulary, with the cross-complemented and projected-complemented variants documented in the facade.

Membership retract by position

Group, hash-bucket, and relational-link membership changes unlink in O(1) by stored position instead of scanning buckets for the unlinked row: removal swaps the contributor’s successor into place and patches its index, hash buckets remove the exact slot, and complement/links restore through retained keys. Grouping keys are owned once per group (no per-contribution clones), and the retained grouping record is slimmer per row. Complement weight closures dropped a stale Clone bound on the key, so key types that are neither Copy nor Clone now work in complemented grouped weights.

Scores, match results, and candidate ordering semantics are unchanged; regression suites pin seeded parity on every reworked family.

Published Version Boundaries

Surface Version Notes
Core runtime solverforge 0.22.0 Rust 1.95; docs.rs serves the 0.22.0 API
Previous core runtime solverforge 0.21.3 Registry and release artifacts remain available
Python solverforge-py 0.6.15 / PyPI solverforge 0.6.15 Runtime 0.22.0, UI 0.9.0
CLI solverforge-cli 3.5.2 Fresh scaffolds target runtime 0.21.3; UI 0.9.0, maps 2.1.4, MCP rmcp 3.5.1
Worked use cases Eight released apps Runtime 0.22.0, UI 0.9.0

The 0.22.0 crate artifact’s SHA-256 is 5c882fc55b59e28632a276df4e8f820ec0a3b52ecaf5d4a7afe69139c954db6b, and .cargo_vcs_info.json binds it to commit df1ce37b4a69a61fe249e71852ea694de94b4493 (the v0.22.0 tag points at the same commit’s workspace snapshot). The publish and release workflows completed on the tag commit. Independent CLI, Python, and app lines move only through their own releases: CLI 3.5.2 still scaffolds core 0.21.3; Python 0.6.15 and the October 18 app train compile against core 0.22.0. Existing applications retain their own dependency pins until deliberately upgraded.

Patch History

Version Date Notes
0.22.0 2026-10-18 One grouping accumulator kernel across all grouped families; membership unlink by exact position; grouping keys owned once; complement weights drop stale key Clone bounds.

Upgrade Checklist

  • Update direct Rust dependencies and confirm Cargo.lock resolves 0.22.0.
  • Grouped complemented weight closures with non-Clone keys: remove the workaround clones.
  • Grouped sharing recognition for a complemented source now accepts plain group_by chains; a #[solverforge_constraints] function that duplicated work to satisfy the old requirement can share again.
  • Scores and decisions are unchanged; re-running your regression suite is sufficient for behavior parity.

Read Node Sharing for the shared terminal contract, Collectors for accumulator shapes, and Joiners for the join surfaces above the shared kernel. The final tag diff records the exact code changes.