fix(annotations): repair the main merge and port the sampling tests (#1277)

The last merge of main kept four tests on a DOM harness that main had
removed, so they threw ReferenceError and CI failed. They now run on the
store-backed board. The merge had also dropped main's empty-board publish,
the leaving/byKey clear in destroy, and three unrelated CURRENT-STATE
paragraphs; all are restored. The projection now builds each height-cache
key once per point, and the changelog entry says what changed.

docs/CURRENT-STATE.md

@@ -127,6 +127,24 @@ broader suffix variants. Follow-ups retain boundary provenance, and cancelled
 lookups preserve the previous analyst cohort and boundary source. Navigation
 search keeps a successful geocode, with no search or view-based region recovery.
 
+Manager-owned periodic network updates pause while the document is hidden. Returning to a visible tab runs one overdue update per enabled layer without overlapping an active refresh and restarts that layer's polling schedule from the return, so the next poll follows one interval later whether or not the catch-up update succeeds; explicit refresh requests remain available. Transit also defers camera-driven feed admission while hidden and rechecks coverage on return. Visibility listeners are released with their polling intervals.
+
+Overpass reserves each normalized query before its disk-cache lookup. The
+initial disk lookup, upstream quota admission, fetching and memory publication
+are coalesced under that reservation. A caller that joined a reservation
+rejected by another client's quota retries admission with its own quota.
+Stale fallback runs per response after the reservation is released, and disk
+writes finish asynchronously. Settled and rejected reservations are removed
+only while they still own their map entry.
+
+External ALPR tags, surveillance-tool fields, voice context text and compact
+analyst strings become one line of at most 180 characters (scene labels: 120).
+Empty ALPR values stay null, and the cap never splits a surrogate pair.
+Canonical analyst identifiers such as `id` and `routeId` stay unchanged, because
+voice follow-ups match on them and bikeshare uses the raw station name as its
+ID. The text is still untrusted: normalization does not prevent prompt
+injection.
+
 Screen annotations skip hidden records and, before surface sampling, reject an
 unsettled anchor only when its whole accepted height range (-430 to 9000 m) is
 behind the camera or below the globe horizon; a settled height is tested as the