mqp2p heartbeat-ttl presence: review follow-ups #29

Out-of-scope items surfaced during the PR #28 review and quorum verification. None are defects in that change; all are tracked here for later.

1. srflx/public_addr is a one-time snapshot. STUN runs only in Peer::new and the heartbeat re-create reuses the address captured at registration time. If the NAT mapping rebinds mid-session, the advertised WAN address goes stale and nothing re-discovers it. update_peer_addr exists but is never called in the running loop. Pre-existing (STUN always ran once); broader than the re-create path the PR fixed.

2. Re-create resets MQDB _version to 1. A reaped-then-recreated peer reappears at version 1 rather than continuing to increment. Nothing in this feature reads _version; only matters to an external consumer relying on version monotonicity.

3. list_peers fetches all $DB/peers rows and filters by _expires_at client-side rather than using a broker-side filter. Correct at current scale, but a churn-scalability improvement. Related: the broker TTL sweep interval is ~60s while the lease is 15s, so dead rows linger server-side and the client-side filter is load-bearing for timely discovery (which is also why the cross-machine clock-skew caveat matters operationally).