Summary

Fixes two issues in collection item fetching: searching a collection's records could fail with a "Failed to count collection items" error, and reference-field text could fail to load on the canvas because some referenced collections received zero items.

Changes

• Filter collection item search through an embedded inner join on collection_item_values instead of collecting matching item ids and using .in('id', ...). The values table has no collection_id, so the old approach matched values across every collection and built an id list large enough to overflow the request URL and fail the count query. The embedded join scopes matches to the collection via the parent FK.
• Fix the top-items-per-collection fallback (used when the Postgres RPC is unavailable): a single shared query with a global limit let an early-sorting collection consume the whole row budget and starve later collections to zero rows, breaking reference-field lookups. Each collection's top N is now fetched independently.

Test plan

• Search records in the dynamic page selector with short terms on a large collection — no "Failed to count collection items" error
• Verify search results are scoped to the current collection only
• Open a page binding dynamic text to a reference field — referenced values render on the canvas
• Confirm reference fields across multiple collections all resolve (none starved to zero items)