[zmstone]

The acceptor closes some sockets before a connection process starts. It counted most of them as closed_other_reasons. This PR splits out two new accept-result counters, which esockd:get_stats/1 returns:

• closed_early: the socket failed with econnaborted, econnreset, enotconn, einval or closed before the connection process was started. This most likely means the peer closed or reset it. Scanners, health checks and load balancers usually cause these.
• closed_forbidden: the access rules denied the peer address.

closed_other_reasons now only counts ignore, connection start errors and crashes. A rise in this counter now points at a real problem.

Related fixes

• esockd_server:inc_stats/3 and dec_stats/3 no longer crash the caller when the metric row is missing. They pass a default row to ets:update_counter/4. Without this, an acceptor on a listener started before the new counters existed would crash with badarg on its first early close.
• The acceptor no longer counts {error, closed} from an async accept. That error means the listen socket is closed, not a client socket. The sync accept path already skipped the count.

Tests

• esockd_acceptor_fsm_SUITE: t_closed_early, t_forbidden and t_tune_fun_other_error. t_einval now expects closed_early. The listen-socket-closed case asserts that no close counter moves.
• esockd_SUITE: t_closed_forbidden uses a real deny rule. t_closed_early resets the client connection while the acceptor waits in tune_fun.
• esockd_server_SUITE: t_inc_uninitialized_stats.

Follow-up: EMQX only reads esockd:get_shutdown_count/1 today. A separate EMQX PR will expose the accept-result counters in the CLI and in Prometheus.