docs: 8.10.2 consumer release notes — update() write granularity, PathResolver idle-log fix, graph-lsm key recognition

RELEASES.md

@@ -31,6 +31,30 @@ is sometimes cited as a 7.x removal — those methods never existed on 7.x; the
 
 ---
 
+## v8.10.2 — 2026-07-29 (metadata-only updates stop rewriting the vector record)
+
+From a production incident on a large deployment: a read-heavy sweep that bumped
+per-entity stats (metadata-only `update()` calls) saturated the disk — 5.8GB written
+in 40 minutes — because every `update()` unconditionally re-persisted the WHOLE noun
+record, unchanged vector included, fsynced.
+
+- **`update()` write granularity fixed at the core.** A metadata-only update (no new
+  `data`, `vector`, or `type`) now writes the metadata leg and index deltas ONLY —
+  the vector-bearing noun record is never rewritten. Vector-side writes and HNSW
+  reindexing still happen exactly when the vector side actually changed. Regression
+  pins: `tests/integration/update-write-granularity.test.ts`.
+- **Consumer guidance:** per-entity stat touches are now cheap, but batch them anyway
+  (one `transact()` instead of N `update()` calls) — granularity fixes the cost per
+  touch; batching fixes the count.
+- Idle VFS `PathResolver` no longer logs `NaN% hit rate` once a minute (stats log
+  only on new traffic, at debug level).
+- Native graph providers' `graph-lsm-*` storage keys are recognized as system
+  resources — the per-boot `Unknown key format` warning for them is gone.
+
+Pairs with the native accelerator's same-day patch release; adopt as one bump.
+
+---
+
 ## v8.10.1 — 2026-07-24 (the no-hot-retry contract + warm()'s metadata surface under native providers)
 
 From a production incident: a native-provider op ground 38-40s inside a transaction,