tarantool-etcd-fv9 gRPC connections closed after 30s idle; no server keepalive PINGs Lined Up

lib/nghttp2/grpc.lua:394-398 hardcodes 'local readable = self.sock:readable(30)' and breaks the read loop when it returns false — the comment reads 'Idle timeout — close the connection.' So any gRPC connection with no INBOUND bytes for 30s is dropped. etcd does the opposite: --grpc-keepalive-interval (default 2h) PINGs idle connections to keep them alive.

Supporting evidence: lib/nghttp2/session.lua:611-620 defines submit_ping with zero callers; ENHANCE_YOUR_CALM (lib/nghttp2/ffi.lua:892) is declared and never used; no keepalive knob exists in app/roles/etcd.lua.

IMPACT NOTE (deliberately downgraded from the sweep's 'client-breaking'): clientv3 auto-reconnects and resumes watches from the last revision, so the real cost is reconnect churn, dropped idle streams, and latency spikes — not breakage. An idle Tarantool EE config.etcd watch would bounce every 30s. Worth fixing; not an outage.

The conformance suite cannot see this: no test idles for 30s.

Cheapest correct fix is to stop treating 'no inbound bytes' as death — loop on readable() rather than breaking, and only close on EOF/error. Optionally send a PING on idle and expose grpc.keepalive.{interval,timeout} in roles_cfg. Consider whether any deployment relies on the 30s reap to shed dead sockets before removing it outright.

A connection idle >30s stays open and serves the next request; an idle Watch keeps delivering after 60s+; keepalive interval/timeout configurable; a test covers the idle case.