[joker-777]

With _unsafeBatchPerCollection, pulled changes are applied in batches of up to 5000 records. This PR commits those batches one after another and sends observers one notification per pull.

Changes

• Batches are now prepared and committed one after another, instead of all at once with Promise.all. A failed batch stops the apply: earlier batches stay committed, and later ones are never prepared, so no record is left with a pending prepared state
• The whole apply runs inside database.experimentalBatchNotifications(), so observers fire once per pull instead of once per table and batch
• experimentalBatchNotifications() now merges everything changed inside it into one notification: one call per experimentalSubscribe subscriber and one collection.changes emission per table. Before, it only deferred the per-batch notifications and replayed them one by one
• Docs (Sync/Frontend.md and the SyncArgs comments) now say that the option shrinks the batch payload sent to the native side, not JS memory, and that it isn't transactional

Tests

• A batched pull of 10001 records (5 batches) sends one notification
• An adapter that fails the second batch shows that batches never overlap, that only the first 5000 records land, and that later records aren't in the cache
• A Database test covers merging notifications
• All three fail on master: 5 notifications, 3 batches called, 3 subscriber calls

Verification

• On this branch: Jest passes on the host, 79 suites (the sqlite-node suite was left out because better-sqlite3 does not build on the host's Node; this change does not touch the adapters). flow check is clean, and ESLint is clean on the changed files
• The four sync PRs from us ([Sync] Apply batched pulls sequentially and notify observers once per pull #1984, [Sync] Validate the whole pulled changeset before applying any of it #1985, [Sync] Track changed columns on locally created records #1986, [Sync] Make sync bookkeeping linear-time #1987) touch applyRemote.js and its tests in separate places. They merge into master together without conflicts, and Jest passes on the result

🤖 Generated with Claude Code

[StasDoskalenko]

Hi @joker-777, thank you for all the sync work, these PRs are really well explained!

I maintain NitromelonDB, a maintained fork of WatermelonDB (New Architecture, Nitro Modules). This one applies there too: _unsafeBatchPerCollection still commits batches with Promise.all, and experimentalBatchNotifications() still replays notifications one by one. It would also fit nicely with our plans for paginated sync (StasDoskalenko#50).

Our fork was rewritten in TypeScript (strict mode) to reduce the chance of bugs like this one slipping through, so I think the community would really benefit if you opened your PRs against it too. The code lives in src/sync/impl/applyRemote.ts there, and the logic should carry over as is.

If you don't have the time or would rather not, no worries at all, I'm happy to port it myself. Just let me know. Thanks a lot!