fix(buzz-acp): clear stale 👀 on events delivered via native steer

[tlongwell-block]

Problem

Since goose-native steering landed (#1160), messages absorbed into an in-flight turn via SteerAck::Success keep their 👀 reaction forever. Reported by Jem in #buzz-bugs: switched channels mid-turn, came back, eyes still on the message with no sign of activity.

Root cause

The 👀 is added fire-and-forget at queue-push time. Cleanup is owned by run_prompt_task's ReactionGuard, which collects event ids from the FlushBatch. But a successfully steered event is consumed by queue.remove_event() in the SteerAck::Success arm and never enters any batch — so no guard ever owns its id, and the 👀 is never removed. Every successfully steered message leaks its eyes.

Fix

Contract: eyes go away when the agent turn stops. One cleanup helper (spawn_steered_eyes_cleanup) is shared by every path.

• TaskMeta.steered_event_ids: the SteerAck::Success arm attaches the consumed event id to the in-flight turn's meta.
• Exact-turn binding (review, Wren): pool.send_steer returns the turn's tokio::task::Id; the ack watcher carries it, and record_steered_event matches on the exact task id — a delayed ack for an ended turn can never bind to a successor turn on the same channel. Stale id → immediate 👀 removal instead.
• Turn end — normal completion (handle_prompt_result) or panic (recover_panicked_agent) — drains the ids and removes the 👀.
• Graceful shutdown (review, Wren): the post-loop drain bypasses handle_prompt_result, so tokio_main now drains all steered ids from every in-flight TaskMeta (drain_all_steered_event_ids) as the loop exits, and — after the grace period settles every prompt task — drops its steer_ack_tx and drains steer_ack_rx (drain_crossing_steer_acks) to recover Success acks whose watchers resolved after the loop stopped polling. Both cleanups are bound-awaited before exit; a 2 s per-recv timeout backstops a wedged watcher.
• Known limitation (pre-existing, documented in fn docs): non-Success acks crossing shutdown and queued-but-never-dispatched events share the same stale-👀-at-exit fate; relay-side cleanup owns that class.

Testing

• pool::tests: exact-id binding, false-when-gone, the generation-reuse race (turn A ends → turn B same channel → A's late ack must not bind to B), shutdown drain take-once semantics.
• steered_eyes_lifecycle_tests: wire-level assertions against a relay stub that answers the reaction /query and captures the signed kind:5 (NIP-09) deletion posted to /events — normal completion, panic recovery, the exact shutdown composition, late-ack immediate cleanup, crossing-ack drain (Success kept, non-Success ignored), wedged-watcher timeout (paused time), and the empty no-op path.
• Full cargo test -p buzz-acp: 416 passed, 0 failed. Clippy --all-targets clean, fmt clean.