Summary

Five kernel-side fixes around spawned-task lifecycle, lock recovery,
and per-session task tracking.

• persist_runs_async: stop swallowing the outer JoinError from
  spawn_blocking. Now returns Result<(), String> and both
  execute_run / resume_run propagate persistence failures into the
  run result. Closes persist_runs_async swallows JoinError from spawn_blocking — workflow persistence panics never surface #3753.
• spawn_supervised helper in a new supervised_spawn module:
  wraps a fire-and-forget future in AssertUnwindSafe(...).catch_unwind()
  and emits a tagged error! on panic. Existing spawn_logged
  becomes a thin alias. ~14 critical-path fire-and-forget sites
  migrated (inbox, whatsapp gateway, session stream fan-out, auto_dream,
  cron agent turns, mcp connect/reconnect, a2a discover, approval
  resolution, memory event publish, daemon-loop tickers around audit
  trim / metering cleanup / memory consolidation+decay / GC sweep /
  session retention / upload cleanup / heartbeat / local provider
  probe / staggered background-agent start). Remaining raw
  tokio::spawn calls are in three buckets that don't fit
  fire-and-forget: test code, handles collected for join, and
  closures whose JoinHandle is the contract. Partially closes tokio::spawn at ~30 sites swallows panics — no JoinHandle awaited, supervisor never notified #3740.
• running_tasks overwrite (running_tasks DashMap.insert overwrites previous task — stop_agent_run can't kill the orphaned LLM call #3739): kept the existing atomic
  insert-then-abort-displaced pattern under the DashMap shard lock
  and documented why it's correct. Closes running_tasks DashMap.insert overwrites previous task — stop_agent_run can't kill the orphaned LLM call #3739.
• running_tasks insert-after-spawn race (running_tasks insert-after-spawn race leaves stale abort handle #3445): gave
  RunningTask a unique task_id (Uuid). Cleanup uses
  remove_if(... task_id matches ...) so a slow predecessor can't
  evict a successor, and the registration site skips insert when
  handle.is_finished() so a task that completed before we got to
  register can't leave a stale AbortHandle. Closes running_tasks insert-after-spawn race leaves stale abort handle #3445.
• budget_config poison (budget_config RwLock .unwrap() panics on poison, breaking every LLM turn #3447): the RwLock read/write paths
  already recovered with unwrap_or_else(|p| p.into_inner()), but
  silently. Added tracing::warn so operators can see when an
  upstream panic poisoned the lock. Closes budget_config RwLock .unwrap() panics on poison, breaking every LLM turn #3447.

Test plan

• CI green (clippy, workspace tests)
• Trigger a panic inside one of the migrated background loops
  (e.g. inbox watcher) — verify error! shows up with
  task=inbox_watcher
• Run two concurrent send_message_full calls for the same
  (agent, session) and verify the older AbortHandle is the one
  aborted, and that cleanup of the older task does not evict the
  new entry
• Cause a panic inside a RwLock write critical section (in a
  test) and verify the next budget_config() call logs a warn
  and proceeds with the recovered state