Coalesces users.is_online transitions into one write per flush window.

Accounts.set_user_online/1 and set_user_offline/1 already skip the write entirely when the flag is unchanged, so a reconnect or a second socket for an already-online player costs nothing. What is left is the storm: N distinct players connecting at once is N separate transactions, and on single-writer SQLite those serialize behind each other.

This buffers the transitions and flushes them as one SELECT (which rows would actually change) plus one UPDATE per direction, then fires the same cache invalidation, broadcasts and hooks the synchronous path fires — once per user that really transitioned, never for a no-op.

Buffered state is in ETS, not in the GenServer

Callers have to read the buffer to decide whether they are making a real transition: a player who disconnects inside the flush window must see their own pending connect, or the disconnect reads as a no-op and the stale connect is what reaches the database. Routing that read through the process would put a GenServer.call on every socket join — the exact serialization this module exists to remove — so the buffer is a public ETS table and the process only owns the timer.

Configuration

config :gamend_core, Gamend.Accounts.PresenceWriter,
  flush_ms: 200      # 0 writes through synchronously

flush_ms: 0 is exactly the old behaviour and is what the test environment runs, so a test can assert on is_online immediately after the call. In production the flag lags by at most one window — StalePresenceSweeper is the backstop that already tolerates far more drift than that, and realtime subscribers are told over PubSub from the channel, not from this write.

If the process is not running (some test setups start a partial tree), flush_ms/0 reports 0 and callers write through rather than dropping the transition on the floor.