Picks up the root-cause fix for the 2026-08-18 partial-deafness incident. zio's poller delivery is destructive -- it clears the token and drops the epoll registration as it emits -- and the scheduler discarded any event whose fiber was not parked. Since `step()` runs the ready batch before it polls, a fiber readied by a cancellation, a close-fd wake or a futex wake is `.ready` at exactly that moment, so its wake was lost permanently: the fd left epoll with the peer's close frame still unread, the worker parked forever, the socket never closed, `relay_connected_inbound` never decremented, and the host never marked down. That is the shape captured in prod -- 365 CLOSE_WAIT sockets, none of them in epoll, 205 still holding an unread close frame. This bump also moves zio off a canary branch and onto main. main had diverged since 2026-08-05 and was missing every fix the 2026-08-04 incident produced, including the ready-queue starvation fix; it is now reconciled. `zio_sched_events_rearmed_total` exports the condition. The soak in zio's bench/deafness.zig does not trigger it with ordinary hangup and cancel churn, so the field trigger is still unnamed -- this counter is what will name it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>