Two related races in flow.usage_metrics aggregation:
1. Paused kickoff dropped its listener. When kickoff_async returned a
HumanFeedbackPending, the finally detached the listener even though
the bus dispatches LLM event handlers on a thread pool that emit
does not wait on. Any pre-pause LLM call whose handler future was
still queued would silently lose its tokens.
Fix: track a paused_for_feedback flag in kickoff_async and skip the
detach when set. The listener stays attached on the instance so late
events continue to accumulate.
2. Stale handlers from a prior kickoff could bleed into a later one.
The handler closure captured flow_ref and wrote into
flow._aggregated_usage_metrics. If a handler from kickoff #1 was
still queued when kickoff #2 reset the accumulator, it would
contaminate the new run's totals.
Fix: snapshot a per-flow _usage_epoch in the handler closure at
attach time. kickoff_async bumps the epoch before attaching, so any
in-flight handler from a prior kickoff sees its stale snapshot and
bails out.
a64b41dd42