mirror of
https://github.com/ruvnet/ruflo.git
synced 2026-09-14 14:01:28 +08:00
rUv
313f2f88e0
Completes the intelligence-audit punch-list with the fixable items, and documents the honest ceilings for the rest. FIXED - Per-task bandit priors (ADR-142): model-router Beta priors are now keyed by complexity bucket (low/med/high) instead of global-per-model, so failures on one task type no longer suppress a model for ALL types. Backward-compatible schema migration (v1 flat -> seed all buckets, version:2). +isolation +migration tests; router-bandit suite made bucket-aware (8/8). - EWC++ honesty: header doc clarifies F_i is a heuristic embedding-importance proxy (embedding_i^2), not true gradient-curvature Fisher. - HNSW backend label: getStatus().backend -> '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 Float32Array where the WASM runtime wants a scalar learning rate). HONEST CEILINGS (re-confirmed NOT fixable in-repo, documented in the audit ledger) - WASM-accelerated HNSW: no WASM HNSW build exists in the stack; native NAPI is already fastest (~1.9-6.5x at N=20k); 150x-12,500x unreachable here. - WASM MicroLoRA apply(): empirically still inert after the flush (delta=0); we do NOT fabricate a gradient from scalar quality to fake adaptation. Verification: cli build clean; 23/23 targeted tests pass (bandit 8 + bug-cluster 15).