[Philenst]

poll_finished_renders acquires a read lock on the render's shard and only drops it at the end of the loop body, but then calls senders.own(render_id) for the unsubscribe — a write lock on the same shard while still holding the read lock. This self-deadlocks the poll task as soon as it finds a finished render with a live subscriber (~5 minutes after boot).

The frozen read guard then wedges all future subscribers of that shard, whose queued write locks in turn block the websocket event loop's reads (tokio's RwLock is fair), freezing it outside of its select! — so the watchdog can never fire and no further websocket events are dispatched. Symptoms: Subscribing gets logged but no events ever follow, and no warnings appear.

• drop the read guard before any await by cloning the senders out first (also in the websocket event loop, so a slow subscriber can't stall the loop while holding a shard lock)
• progress updates now use try_send; they're cosmetic so drop them on a full channel instead of blocking
• log the poll's previously silent error paths (failed query / missing render)