SQLite persistence: after two quick writes to one row, the collection briefly shows the state before the second write #2201

• I've validated the bug against the latest version of DB packages

Describe the bug

An Electric collection with persistedCollectionOptions inserts a row and updates it right away. Both handlers await awaitTxId. After the update is confirmed, the collection shows the pre-update state for 10–30 ms, then the updated state again. Without persistence, this never happens.

I hit this in a shopping list app. A user adds an item and ticks it off. The tick disappears for a moment, then comes back. An action that runs during that gap reads the old state: "clear bought items" saw the item as not bought and left it on the list. With 10× CPU throttling, the gap reached 263 ms in my app.

To Reproduce

Repro: https://github.com/daniel-fahl/tanstack-db-flicker-repro (StackBlitz). It needs no backend: a fake Electric server runs in the page.

1. Open the repro.
2. Click Add and tick 30 times. Each run inserts a todo and updates done to true without waiting in between.
3. See the result: 30 of 30 runs flicker. Each run logs the states the collection showed, for example missing → unticked → ticked → unticked +24 ms → ticked +41 ms.
4. Switch to Control (?persistence=off) and run again: 0 of 30 flicker.

Expected behavior

When awaitTxId(txid) resolves and the handler returns, the collection never shows a state older than that write.

Screenshots

None. The repro's log shows every state with its timing.

Desktop (please complete the following information):

• OS: macOS; my app's CI also hits it on Ubuntu (GitHub Actions ubuntu-latest)
• Browser: Chromium, via Playwright
• Version: Chromium as bundled with @playwright/test 1.64

Smartphone (please complete the following information):

Not tested on a device yet. 10× CPU throttling stands in for a slow phone.

Additional context

Versions: @tanstack/db 0.15.1, @tanstack/electric-db-collection 0.6.2, @tanstack/browser-db-sqlite-persistence 0.2.36, @tanstack/offline-transactions 1.1.3, @electric-sql/client 1.5.28. The same result on @tanstack/db 0.15.0 with @tanstack/electric-db-collection 0.6.1, and on 0.13.0 with 0.5.9.

Cause, in brief: the update's txid arrives while SQLite still stores the insert's synced transaction. awaitTxId resolves, but the update's synced transaction waits for applyMutex, so it is not in pendingSyncedTransactions yet. When onUpdate returns, recomputeOptimisticState finds no committed pending synced transaction for the key and drops the optimistic row. The repro's README has the full trace with timings from my app.

Possible fix: I am not sure where it belongs. Two ideas:

• awaitTxId resolves only once the synced transaction for that txid has reached the collection, at least as a committed pending transaction.
• recomputeOptimisticState also keeps the optimistic row while a synced transaction for that key still waits on persistence.

Waiting until the rows are fully applied would deadlock: the collection holds back synced rows while a transaction is persisting, and the handler's own transaction is persisting while it awaits the txid.

My workaround: the handler waits 150 ms after awaitTxId resolves. That was the shortest wait that removed every flicker in my app at 10× CPU throttling.

Related: #2120 may be related. It covers changes that progressive mode applies and confirms during the background full sync.