sdk: a stale lease error from the local LLM worker kills the whole run, though the kernel already retried the step

[khaliqgant]

Problem

packages/sdk/src/cli/direct-run.ts:85 treats any error from the local LlmWorker as fatal:

localLlm.on('error', error => { llmFailure = error; client.close(); });

So when the worker hits a lease error on an attempt the kernel has already expired and retried, the whole run ends as protocol_error and can't be resumed. The retry succeeds in the journal anyway. Seen two ways in relayflows 2.0.29:

1. Stale dispatch. The worker receives a dispatch after its lease deadline. withWorkerLease throws Agent lease is already expired for <run>/<step>, and that becomes fatal. The kernel had already journaled lease_expired → retry, and attempt 2 completed success. Journal: 00-job1-attempt2-parallel-lease-expired.txt; the replay is quoted in feat(examples): prompt-lab — the Prompt Lab product brief as one relayflow #559.
2. Renewal racing completion. A step whose child run journaled step.completed success + run.completed success was reported as lease_conflict: attempt has no active worker lease, and the run died. Intermittent: once in about 50 sequential calls. 00-prove-attempt1-lease-conflict-after-success.txt

Scope

• A lease refusal for an attempt the kernel no longer owns shouldn't abort the run. Examples: already expired before execute, lease_conflict / no active worker lease on heartbeat or complete. The kernel owns that attempt's fate (retry, or a completion that already landed). Drop it with a diagnostic (WAITING/warning line); don't close the client.
• Any other worker error stays fatal, exactly as today. Fail closed.
• The agent worker path (AgentWorker) likely has the same shape. Check it and apply the same rule if so.

Acceptance

• A unit test injects a dispatch whose lease_deadline_ms is already past. The run continues, and the kernel's retry completes the step.
• A unit test makes stepHeartbeat / stepComplete reject with lease_conflict after the step's journal is success. The run reports success.
• A non-lease worker error still ends the run as today (regression test).
• Mutation-verified per AGENTS.md.

Related: the root cause of case 1 under concurrency is tracked separately (it's why the dispatch was stale).