oskiewar relay: a claimed room lets go of the dead host's sequence master
Joining and leaving both ends of a versus room verify clean end to end — join opens the fight in ~1.2s, a closed guest reparks the chair 2.6s later, the same link reseats a returning guest, and a dead host's room is claimed by its challenger in 2.7s. But after every claim the room sat wire-silent for ~14 seconds: the successor publishes from its own seq zero while publish()'s stale-frame guard still held the corpse's high-water mark, so one frame was dropped for every frame the old host had ever sent. The same leftover snapshot handed mid-handoff spectators a freeze-frame of the old fight and let the matchmaker offer a chair beside a corpse. addPublisher now drops the previous snapshot when a new publisher takes the chair; a test pins the takeover (fresh seq fans out, no stale frame served). Measured with a two-window CDP harness on blueberry: page enters the lobby 2s after the claim, publisher socket sends 30/s from 2.7s — only the relay was muting the room. Claude-Session: https://claude.ai/code/session_01Nj9XDa5sX89pPk8N1cwh4g