On a bare Linux VPS the durable split is nginx owning 80/443 and one PM2 listen bound on loopback. Let PM2 own cluster fan-out on that single port — listing every worker in an nginx upstream doubles fail surfaces and breaks session affinity when a process drains. Keep proxy_http_version 1.1 with Connection cleared so keepalive into Node reuses sockets, and always forward Host, X-Forwarded-For, and X-Forwarded-Proto so Express builds redirects and secure cookies against the public scheme.
After pm2 reload or a cert renew I check three layers in order: pm2 status, a health hit on the private listen, then the same path through the 443 server block. A short 502 window usually means nginx still talked to a draining worker; wait for the new cluster to accept, or tighten max_fails / fail_timeout so a half-dead process leaves the pool quickly. Log $upstream_status beside $request_time to separate TLS or vhost drift from a Node crash. Publish only 80/443 on the firewall; leave the PM2 port local.