Release v2.5.2 — consolidation v1: the brain that learns from its own outcomes

[RodCor]

What ships

Consolidation v1 (the headline)

• kimetsu brain reinforce (new command): offline consolidation pass.
  --staple merges repeatedly co-cited memories into single fact memories (precomputed multi-hop joins, originals kept, provenance auditable). --routes builds a query-routing index from citation history.
• brain cite now takes repeated --memory-id (grouped citations = the co-citation signal) and --query (feeds routing).
• Retrieval-time routing boost: bounded (fixed budget + per-candidate cap + min-support + power-law decay). Schema v9 → v10, automatic.

Learning-loop guardrails (validated on a k=5 iteration harness)

• Usefulness boost gain capped — citation feedback can reorder near-equals but never flood retrieval (fixes a measured collapse mode).
• Citation-history-scaled failure penalty: -1.0 / (1 + priors/3) — flaky runs can't prune proven memories.
• All scoring constants centralized and documented in scoring.rs.

Maintainability

• project.rs (8.3k lines) split into 7 focused modules; CLI main.rs (13.8k) split into 10 command modules. Pure moves, public API unchanged.

Supply chain + UX

• npm publishes now carry Sigstore provenance attestations.
• Schema-mismatch errors name the offending file and the fix.

Benchmarks

• LoCoMo 89.4% on the standard 1,540-question set (published on kimetsu.dev).
• Learning-curve validation: feedback never degrades retrieval; train/holdout split proves generalization.

Gates

• 522 kimetsu-brain + 253 kimetsu-cli tests green (embeddings, single-threaded), clippy -D warnings clean on lean + embeddings, fmt clean.
• Changelog updated (incl. retroactive v2.5.0 entry).