docs(review round 20): namespace-from-real-engine-args wording (fp8 wrong-hit risk called out), CLI-error-aborts behaviour, model-name prefix coherence (mimo-pro/<hash> both paths), KV_PERSIST_BLOCK_SIZE, MAX_NUM_SEQS=2 + KV_CACHE_DTYPE=auto note, How-the-block-is-tested section, test counts synced (143/160/177 + recipes 302/307)
docs(review round 19): Recipes-Integration rewritten (namespace mechanism, 4 things KV_PERSIST=1 does, port-name mapping, two-paths table), Quick-Start launcher wording fixed (not a mirror), test counts corrected
docs(review round 18): scenario A = L1 memory hit is expected (not a failure), tmp_guard liveness works in docker mode via pid-namespace + heartbeat, dashboard honesty note (4/28 panels remapped to 0.5.3 metrics, Coordinator stays empty)
docs(review round 17): bfloat16 CacheDType spelling, LMCACHE_LOG_LEVEL knob, restart-accounting caveat, systemd 0700 enforcement, fingerprint comment fix
docs(review): registry liveness-by-process wording, advanced identity fields in Configuration x3, sync test counts to 157
docs(review): Security flat-layout purge safety, Architecture logical namespace -> model_name mapping, active-write registry replaces heuristic wording, real dedup test references
docs(review): flat cache layout + purge --namespace=NAME, quotas = policy layer not enforcement, LRU = production eviction / retention.py = planner, KV-dtype vs L2-codec split, honest tmp-guard heuristic, sync test counts, fix Home.ru grammar
docs(review): lossless codec per-codec dependencies, tmp-file accounting + active-write safety, monitoring stack (Grafana), FAQ updates
wiki: sync test count to 59, drop 4-GPU host phrasing (universal GPU host wording per repo 3076adc)
wiki: universalize wording -- any model, any GPU host, generic disk names; same as repo commit 12a1955
wiki: review round 7 -- shared-VM-disk wording, cache-miss matrix x3, verify-engine gate in Recipes-Integration, unit-test count 52 sync
wiki: review round-6 -- TTFT-UP->DOWN typo fixed, Retrieved->L2-read evidence wording, Restart-accounting NEEDS TESTING, Recipes-Integration verify-engine gate
wiki: review round 5 -- verify-engine gate in step 8 (x3), eviction math ~0.81 not 0.85, budget 1500 everywhere, KVP_CACHE_DEVICE on step 2, RU language polish
wiki: evidence-criteria wording aligned with actual test-recovery pattern (not just Retrieved)
wiki: review round 3 -- Quick Start step 7 preconditions (recipe clone, API_KEY, MODEL_DIR), no manual digest copy in step 5, Python 3.12 note, terminal exports, detached service note
wiki: remove all chat-app branding; Apache-2.0 + NOTICE attribution; 1.5 TB budget + context-limit clarification
wiki: strict recovery protocol v2 + one-command identity setup + pinned lmcache install - Quick-Start (x3): identity now filled AND verified by scripts/setup-identity.sh (all mandatory fields, no manual paste); lmcache installed from pinned requirements-lmcache.txt - Persistence-and-Recovery (x3): strict acceptance protocol -- log-position markers, store-before-restart, L2-vs-L1 evidence distinction, hard-fail readiness, machine PASS/FAIL verdict
wiki: full documentation in EN/RU/ZH (10 pages x3) Home, Quick-Start, Architecture, Configuration, Persistence-and-Recovery, Storage-and-Eviction, Recipes-Integration, Security, Troubleshooting, FAQ. All commands as-is (no placeholders), all links verified internally. Recipes integration (KV_PERSIST=1) documented in all three languages.
Initial Home page