[lavaman131]

Moves the Atomic harness to the stable 0.9.43 release.

Versions. catalog:harness pins @bastani/atomic 0.9.43. The lockfile resolves @bastani/atomic, @bastani/atomic-natives with its eight platform packages, and @bastani/pi-ai at 0.9.43, with no older @bastani package left. Atomic's @earendil-works pi packages stay at 1.4.0.

What changed in Atomic, and what it means here. 0.9.43 is about local workflow worker pools:

• A crashed worker's child is replayed on a new worker instead of failing the root.
• Leases follow worker liveness, with a new workflows.localWorkers.leaseMs setting.
• Workers start without reading the whole journal, with a new startupTimeoutMs setting.
• WorkflowInfrastructureError and isRetryableWorkflowError mark retryable pool faults.
• Each worker's standard error is kept in worker-logs/ beside the journal.

Pools stay off unless the operator's Atomic settings enable them. None of the types the adapter uses changed, so the adapter needs no code change.

The worker logs do affect Chopin. I ran a real Planner workflow with an operator-enabled pool and a child that crashes. Atomic put the logs under Chopin's journal directory, not its default, replayed the child three times, and Chat's run card ended as failed. That directory is private only because the adapter creates it with mode 0700, and nothing tested that. The journal test now checks the mode, and it fails if the adapter drops it. The self-hosting guide names the new version, says where the logs go and that the logs of failed workers are kept, and names the two new settings.

Connector. Not changed; it does not depend on Atomic.

Verification

• bun test: 6734 pass, 22 skip, 0 fail
• bun run types and bun run ci pass
• The new mode check fails without the adapter's 0700 and passes with it
• Server Docker image builds and starts healthy with HARNESS=atomic, and it imports Atomic 0.9.43 and its new workflow exports
• Targeted bun run e2e (harness, run card, smoke, decisions deck): 35 passed