[kyscott18]

Summary

Accepted mutations committed to Postgres while holding the speculative-state lock, so submit, watch, and the next batch waited on the commit. Batches and force-included mutations now commit after the lock is released.

• Commits stay ordered. Each commit waits for the previous one (a chained deferred, set under the lock). Restart resumes after the highest committed execution index, so a later commit must not land before an earlier one.
• Callers still get a response only after their batch commits.
• A batch enters acceptedBatchesQueue when it executes, so simulate always sees the newest journals. submit waits for the commits it covers after simulating, outside the lock, and before it broadcasts. Nothing settles onchain before it is committed.
• A failed commit stops the program, which fails waiting callers with TypewriterStoppedError.

This lowers latency for work that contends on the lock. It does not raise throughput: the batch loop still runs execute → commit serially.

Test plan

• bun run typecheck, bun run test (packages/typewriter), biome check
• Simulation seeds 0x2, 0x3, 0x4 pass

🤖 Generated with Claude Code

[kyscott18]

Superseded: the persist-fiber work is moving to a new branch off main.