fix: honour keyset cursor in paginated live queries

[ivasilov]

Makes cursor-based pagination correct under Supabase Realtime by fixing three defects in how the adapter handled core's keyset hints. loadSubsetOptionsToSearch now skips offset whenever a cursor is present, so keyset requests no longer re-skip rows the cursor already moved past. supabaseQueryFn issues a second, unlimited whereCurrent request in parallel for boundary ties, so rows sharing the boundary orderBy value are no longer dropped. subsetOptionsToQueryKey now folds whereFrom/offset into the cache key so distinct windows of one subset no longer collide and delete each other's rows (the key still starts with [tableName] so realtime prefix-matching is unaffected).

Cursors are rendered through the adapter's own toPostgrestParams (strict), which handles the composite or(...) cursors core emits for multi-column orderBy and serialises Date boundaries as ISO 8601; the query-key function never throws (it falls back to a deterministic discriminator on an un-pushable cursor); and limit: 0 is treated as a distinct subset so an empty result cannot take over and delete the unlimited query's rows.

The @tanstack/db dependency is bumped to the 0.9 line, with the range set to >=0.6 <1.0 (0.5.12 was the first version exposing the cursor whereFrom/whereCurrent expressions; 0.6 is the practical floor).

Tests. Adds end-to-end coverage of windowed (infinite) pagination via createLiveQueryWindowController: a single page, window growth by keyset cursor (no offset), and the boundary tie class (whereCurrent). Note that on 0.9 the ordered-source loader issues extra reads around an ordered+limited query — a boundary probe (col=eq.<last>) to settle hasNextPage, or a full-source load when the order isn't keyset-expressible (e.g. a string column under the default locale sort) — so five exact single-URL query-generation tests no longer hold and are marked test.todo pending a decision on how to assert generation under the new loader behaviour.

Note

The unlimited whereCurrent tie request must be routed through PR #17's page-size loop once that PR lands — a tie class larger than the API row cap would otherwise be truncated. That change belongs to #17.

🤖 Generated with Claude Code