Merge pull request #358 from rush-db/changeset-release/main

chore(release): version packages v2.10.2

Author: 1pxone

.changeset/rare-frogs-knock.md

@@ -1,5 +1,20 @@
 # rushdb-core
 
+## 2.10.3
+
+### Patch Changes
+
+- 8e020f2: Dashboard: reworked project Getting Started into a "Connect this project" page with a connection panel (API key, URL, status) and SDK-first setup tabs; neutral, non-shifting feedback for copy/test-connection buttons; unified JSON + language tabs in the Search Query modal.
+
+  Core: fixed flaky post-import side effects — record/property counts, schema recalculation, and relationship suggestions could be skipped until records were touched again. Side effects now run on a guaranteed post-commit hook (no polling), with recount → schema recompute → relationship analysis correctly ordered and isolated so one failure no longer cascades.
+
+- 8e020f2: Make schema recalculation scale with project size and stop AI endpoints from timing out on large datasets
+
+  - **Schema recalculation is now labels-first and incremental in shape**: instead of one monolithic full-graph scan inside a single transaction, the rebuild fans out small per-label statements — property inventory via pure VALUE-edge counts (no value reads), exact streaming min/max for number/datetime properties only, and split relationship topology/property queries — each running as its own short transaction on dedicated read sessions. No single unit of work can approach the server-side transaction time budget regardless of project size; on a 40k-record / 320k-value dataset a full forced rebuild completes in ~0.5s.
+  - **`POST /ai/schema` and `/ai/schema/md` no longer block or 408 on large projects**: a stale cache is served immediately while a single-flight background refresh recomputes it; concurrent requests share one rebuild instead of stacking full-graph scans. `force: true` still waits and now guarantees read-your-writes — a force call issued after a write never returns a schema computed before that write committed.
+  - **Smart Search (`POST /ai/search-query`) no longer fails with “transaction time budget” errors**: the endpoint releases the idle request-scoped transaction before the LLM round-trip (previously the transaction expired while waiting on the LLM and the request failed at commit), and query generation uses the cached schema instead of triggering a synchronous recompute.
+  - Behavior note: `isArray` detection for string/boolean properties is now based on the bounded sample window (first 100 records per property) rather than a full column scan; number/datetime `isArray`, all min/max ranges, record counts, and relationship counts remain exact.
+
 ## 2.10.2
 
 ### Patch Changes

.changeset/thin-foxes-scream.md

@@ -1,6 +1,6 @@
 {
   "name": "rushdb-core",
-  "version": "2.10.2",
+  "version": "2.10.3",
   "description": "RushDB Core",
   "private": true,
   "license": "Elastic License 2.0",

platform/core/CHANGELOG.md

@@ -1,5 +1,13 @@
 # rushdb-dashboard
 
+## 2.10.3
+
+### Patch Changes
+
+- 8e020f2: Dashboard: reworked project Getting Started into a "Connect this project" page with a connection panel (API key, URL, status) and SDK-first setup tabs; neutral, non-shifting feedback for copy/test-connection buttons; unified JSON + language tabs in the Search Query modal.
+
+  Core: fixed flaky post-import side effects — record/property counts, schema recalculation, and relationship suggestions could be skipped until records were touched again. Side effects now run on a guaranteed post-commit hook (no polling), with recount → schema recompute → relationship analysis correctly ordered and isolated so one failure no longer cascades.
+
 ## 2.10.2
 
 ### Patch Changes

platform/core/package.json

@@ -1,6 +1,6 @@
 {
   "name": "rushdb-dashboard",
-  "version": "2.10.2",
+  "version": "2.10.3",
   "description": "RushDB Dashboard",
   "type": "module",
   "private": true,