fix(net): broadcast LEFT on disconnect, add keepalive/reaper, cap PLC hiss

Three reported bugs traced to one root cause plus two missing designed features:

1. Stale users + eternal PLC hiss (root cause): ConnSession::close() silently
   erased dropped users without broadcasting UserEvent::LEFT, so peers never
   learned the user left and their audio engines never called remove_stream —
   Opus PLC synthesized comfort noise forever. Fix: broadcast_left() helper
   + close() broadcasts LEFT before erasing.

2. PLC cap (defense-in-depth): on_playback now caps pure PLC at ~2s, then
   emits digital silence so a stale stream can never hiss forever even if
   remove_stream is skipped. Resets automatically on fresh packets.

3. No timeout / no ping: client never sent Ping, server had no last_seen /
   reaper, so half-open connections (NAT timeout, wifi loss, sleep) left
   ghost users forever. Fix: client Ping every 15s with RTT measurement,
   ConnSession::last_seen bumped on every inbound TCP/UDP frame, steady_timer
   reaper sweeps every 15s and drops sessions older than 45s (configurable
   via server::Config).

4. UDP KEEPALIVE: client sends plaintext kFrameKeepalive every 5s; server
   bumps last_seen + echoes back. Keeps NAT bindings alive and lets media
   activity defer the reaper independently of TCP.

5. Graceful client disconnect: vc_disconnect() sends Disconnect{code=0} via
   a flag-based io-thread exit (no double-close race); server handles
   client-sent Disconnect with immediate close() + LEFT broadcast.

3 new tests: disconnect_left, plc_cap, reaper_timeout. 21/21 ctest green.

Docs: protocol.md §6/§7, voice.md §6, architecture.md §5, PROGRESS.md.

docs/architecture.md

@@ -197,6 +197,10 @@ Design notes:
   and presence can be subscribed more broadly. This keeps fan-out bounded on big servers.
 - **Stateless-ish media.** UDP carries no auth per packet beyond the media-AEAD session;
   the 5-tuple→session binding is established once via a token (see protocol.md §4).
+- **Keepalive reaper.** An `asio::steady_timer` sweeps every 15 s and drops any session
+  whose `last_seen` (bumped on every inbound TCP or UDP frame) is older than 45 s. Each
+  drop broadcasts `UserEvent::LEFT` so peers clean up immediately. This catches half-open
+  connections that never produce a TCP EOF. Configurable via `server::Config`.
 - **Single process, scalable later.** v1 is one process, one machine. The session registry
   and router are written behind interfaces so a future build can sit them behind a shared
   bus for multi-node, but that is explicitly out of scope for now.