claude-flow

Ruflo v3.10.9 — per-task bandit learning + honest intelligence labeling

Completes the intelligence-audit punch-list (docs/reviews/intelligence-system-audit-2026-05-29.md): fixes the genuinely-fixable items and documents honest ceilings for the rest — without fabricating any signals.

Fixed

• Per-task bandit priors (ADR-142) — model-routing Beta priors are now keyed by complexity bucket (low/med/high) instead of global-per-model. Failures on one task type no longer suppress a model for all task types (the "8 Haiku failures suppress Haiku everywhere" defect). Backward-compatible schema migration (v1 flat → seed all buckets, version:2); proven by a per-bucket isolation test + a migration test.
• EWC++ honesty — clarifies F_i is a heuristic embedding-importance proxy (embedding_i²), not true gradient-curvature Fisher.
• HNSW backend label — ruvector-native vs ruvector-stub-search-disabled; isWasm()===true means the broken stub (search returns []), not acceleration, so a regression is now visible.
• MicroLoRA — fixed the genuine applyUpdates() signature bug (passed a Float32Array where the WASM runtime wants a scalar learning rate).

Honest ceilings (NOT fixable in-repo — documented, not faked)

• WASM-accelerated HNSW: no WASM HNSW build exists in the stack; native NAPI is already fastest (~1.9–6.5× at N=20k); "150×–12,500×" is unreachable here.
• WASM MicroLoRA apply(): empirically still inert after the flush (Δ=0 after 200 adapts). We deliberately do not synthesize a gradient from the scalar quality to fake adaptation — that would be a fabricated signal (the same class as the Flash Math.random metric the audit removed).

All three packages published at 3.10.9 (latest/alpha/v3alpha in lockstep). CI: 29/29 green. This closes the intelligence punch-list — every item is now fixed, shipped, or documented with an honest reason it's deferred.

🤖 Generated with RuFlo