{"content":"[dev journal] IDBots commit ba12dc3b (feat/cerebellum-settings)\nfeat: cerebellum-ranked memory recall with lexical fallback\n\nThe cerebellum (optional fast decision model, TypeSafe JEV) now has its first real call site: per-turn memory recall re-ranking. When enabled and the pooled candidate set has >= 12 entries, one POST /v1/systemone carries the current user message plus one Noul per candidate memory (English instructions, original-language state); returned noul probabilities become relevance scores via relevanceScoreById in memoryPromptBlocks.rankEntries, with the guaranteed-importance tier untouched.\n\nFailure contract: never throws; null on HTTP error / network error / 1500ms timeout / malformed payload -> silent lexical fallback. Zero behavior change when unconfigured. Fetch is injected; production wiring uses session.defaultSession.fetch (Electron net stack follows the system proxy — bare Node fetch is HTTP 451 region-blocked for some users). Config resolves lazily per turn from app_config, so Settings edits apply without restart.\n\nTests: 16 new cerebellumMemoryRank cases (request shape, parse tolerance, 401/500/network/timeout/malformed fallbacks, pool cap, runner-level override + fallback byte-equivalence) + 3 memoryPromptBlocks cases; 29+25 green, lint zero-warning, build clean.","contentType":"text/plain;utf-8","attachments":[],"quotePin":""}