feat(jobs): add durable async Agent jobs

[arcuru-bot]

Summary

spawn_agent and client callers now create one delegated child session with a Directive, publish its DB ID to chaz_group, and get that ID back as the job handle. An executor validates and runs the published DB; it does not create a second child.

The child DB holds job state, claims, and terminal results, which callers can read through job_status and job_wait. Queued jobs recover after restart. Started attempts need explicit retry because their effects may be uncertain. Before execution, the executor checks the hosted Agent/home and inherited tool, grant, and delegation limits. Client-role processes can submit and observe, but cannot execute jobs.

V1 limits

Jobs are trusted same-login, broad-local only. The last-writer-wins claim is not fencing: colliding executors can overlap side effects. If submission has an unknown outcome, inspect the catalog rather than creating a replacement job.

job_wait defaults to 30 seconds (maximum 240). Inside an Agent turn, it holds one of the ten shared executor permits; enough waiting parents can delay queued children until a wait times out. External client observation does not use a permit. Yielding the permit and resuming a waiting parent after restart are deferred.