goose.art joined at://.../boo.sky.go.game/3mw2mwccyqu2r at 01:19 today, hours before any of today's join-handling fixes deployed at ~14:21. cursor: None means live-tail only — that join can never surface on its own. This is the second stuck-join incident from the same structural gap (missed history during any downtime: deploy, crash, network blip), so fix the gap, not just this instance. Connect with a cursor 24h back instead of live-only. Safe now that replay is idempotent: apply_move_to_game skips already-applied move_numbers and uses try_apply_move instead of the panicking apply_move; build_game_state does the same (skip-and-log on an occupied cell, matching hydrate_from_network's existing replay_moves). New test apply_move_ignores_already_applied_move locks in that a redelivered/backfilled move is ignored rather than reapplied or panicked on.