Non-blocking State/Event writes in the concurrent scheduler

[qinhaihong-red]

Parent

#4

What to build

Found by the #4 review (deferred — low severity at v0.1's agent-CLI workload; ADR 0003 points this work toward the failure-semantics/await issues).

State writes (synchronous SQLite commit, fsync-class) and Event-log appends run directly on the asyncio event loop between scheduler wakeups. While the loop blocks in a commit, the loop-resident pipe readers from communicate() cannot drain, so several simultaneously-chatty subprocesses whose output exceeds the OS pipe buffer (~64KB) block on write and the parallelism the scheduler exists to provide degrades toward serial. Make state/event persistence not stall in-flight pipe draining — e.g. batch commits per wakeup, or offload the blocking writes off the loop.

Acceptance criteria

• State/Event persistence does not block in-flight subprocess pipe draining under concurrency
• Durability of recorded attempts/timestamps/events is preserved (no lost records on crash)
• A test demonstrates concurrent chatty branches are not serialized by state writes

Blocked by

• Failure semantics: retries, timeouts, cancellation, and resume #6 (failure semantics: retries, timeouts, cancellation, resume) may reshape state-write batching; coordinate