`LowLoop#start` (lib/low_loop.rb:49-58) only rescues exceptions from `handle_connection`:
loop do
socket = server.accept # not covered by the rescue below
task.async do
handle_connection(socket)
rescue StandardError => e
render_error(e)
ensure
socket&.close
end
end`server.accept` itself is outside that rescue. Real TCP servers hit accept() failures under load — Errno::EMFILE/ENFILE (fd exhaustion), Errno::ECONNABORTED (client resets mid-handshake). Any of these propagates out of `loop`, out of the `Async` block, and up through `LowLoop#start` uncaught. `bin/server` calls `.start` with no surrounding rescue, so one transient accept() failure takes down every in-flight connection on the server, not just the one that triggered it.
Suggested fix: wrap the `accept` call (or the whole loop body) in its own rescue so a single accept() failure logs/backs off rather than killing the reactor.
`LowLoop#start` (lib/low_loop.rb:49-58) only rescues exceptions from `handle_connection`:
`server.accept` itself is outside that rescue. Real TCP servers hit accept() failures under load — Errno::EMFILE/ENFILE (fd exhaustion), Errno::ECONNABORTED (client resets mid-handshake). Any of these propagates out of `loop`, out of the `Async` block, and up through `LowLoop#start` uncaught. `bin/server` calls `.start` with no surrounding rescue, so one transient accept() failure takes down every in-flight connection on the server, not just the one that triggered it.
Suggested fix: wrap the `accept` call (or the whole loop body) in its own rescue so a single accept() failure logs/backs off rather than killing the reactor.