Each connection used to run its own handler_send loop consuming from
the shared dispatcher queue. With zero clients connected (e.g. an
iPad that silently drops on sleep), nothing drained the queue, so
state updates piled up unbounded and got replayed as a burst on
reconnect. Now one pump task drains the queue continuously regardless
of connection count, and reconnecting clients resync from
global_state instead of a backlog.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TvgC7oy9MxaA4ZQxXqkNdS