Notification bell: the websocket stream was rendered with a plain filter, so
a stream replay after a reconnect surfaced — and counted — the same
notification twice, inflating the unread badge. De-duplicate by
notification_id (newest copy wins, arrival order preserved) and key the list
on the id instead of array index.
PM kanban: the board had no columns for the recovery states (paused,
needs-revision, awaiting-CEO-approval, cancelled), so a human had no way to
drive a wedged task into them from the UI. Add those columns; dragging a card
issues an admin status override, giving the CEO direct recovery from the board.