systemctl restart leaves :8888 dark ~5-15s (161 function imports before listen); every bare reverse_proxy failed its first dial (lb_try_duration defaults to 0) and Cloudflare relayed the 502. All 36 proxies now import a shared lith_proxy snippet with lb_try_duration 30s / lb_try_interval 250ms, so requests queue through a deploy instead of erroring. Server also severs idle keep-alives at SIGTERM (closeIdleConnections) so no request dies on a reused dying socket. Trade-off: a genuinely crashed lith now takes up to 30s to surface a 502 instead of failing instantly. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>