v0.66.0

Contended decisions have their own error

A transaction that finds its decision rows already being evaluated by another now fails with decision_contended instead of stale_context. Nothing has changed when this happens, so the error reports retryable: true and the unchanged request can be sent again shortly. Functional updates retry it automatically. Code that caught AbloStaleContextError to handle contention should also check for the new code, which arrives as a plain AbloError. stale_context now means only that a row changed after it was read. The error contract version is 2026-09-20.

Reconnect catch-up arrives in bounded chunks

A client returning from a disconnect now receives the changes it missed as a sequence of bounded chunks rather than one message covering the whole gap. Each chunk is stored locally before the client acknowledges its position, so an interrupted catch-up resumes from the last stored chunk and a replayed chunk does no harm. A chunk that cannot be stored stops the catch-up with a sync error instead of being skipped. The client reports catch-up complete only once every chunk is stored. The exchange is version 2 of the sync protocol; servers continue to serve version 1 clients, so applications can upgrade on their own schedule.

Commit latency covers batched model writes

onCommitLatency now reports model writes that share a batch, with one sample per local transaction. A retried seal keeps its first start time, so the local measurement includes the wait.

Large bootstrap scopes use request bodies

Browser clients with many subscriptions now send bootstrap scopes in a request body, avoiding proxy URL limits. The matching sync-server change must be live before these clients upgrade.