fix(livekit): single UDP media port + pin node_ip for the NAT'd server

The prod box is behind NAT (private 192.168.88.61 behind public 118.95.33.89), so
LiveKit auto-IP-detection would pick the wrong (outbound) address. Pin
rtc.node_ip=118.95.33.89 and collapse media to one UDP port (50000) + TCP 7881 to
minimize the upstream gateway port-forward the network team must add. Docs updated
with the exact forward table.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

DEPLOY.md

@@ -213,14 +213,21 @@ The app mints join tokens with the secret (server-side only); the same key/secre
 - **Websockets Support: ON**. Request an SSL cert (Let's Encrypt) + Force SSL.
 - NPM reaches `livekit:7880` by container name — both are on `nginx_proxy_manager_default`.
 
-**4. Open the media ports on the VPS firewall** (these are host-published by the compose service,
-bypassing NPM — WebRTC media can't traverse an L7 proxy):
-```bash
-ufw allow 50000:50100/udp   # WebRTC media (must match livekit.yaml port range)
-ufw allow 7881/tcp          # WebRTC-over-TCP fallback for restrictive networks
-```
-If the VPS is behind NAT and LiveKit can't auto-detect its public IP, set `rtc.node_ip` in
-`livekit.yaml` to `118.95.33.89` and redeploy.
+**4. Media ports — NAT port-forward (REQUIRED here).** This box sits **behind NAT**: its only
+interface is a private `192.168.88.61`; the public `118.95.33.89` (DNS) is mapped by an upstream
+gateway. WebRTC media can't traverse NPM (L7), so the gateway/router must forward the media ports
+to the box. To keep the ask minimal, LiveKit is configured for **one** UDP port + one TCP fallback:
+
+Ask whoever controls the network/gateway to forward, from `118.95.33.89` → `192.168.88.61`:
+| Port | Proto | Purpose |
+|------|-------|---------|
+| 50000 | UDP | WebRTC media (all participants mux over this one port) |
+| 7881  | TCP | WebRTC-over-TCP fallback (restrictive client networks) |
+
+`livekit.yaml` already pins `rtc.node_ip: 118.95.33.89` (auto-detect would pick the wrong outbound
+IP behind this NAT). The host's local `ufw` is inactive, so no host-firewall change is needed — the
+only requirement is the upstream port-forward above. Until it exists, signaling connects but media
+won't flow (participants see each other's tiles but no video/audio).
 
 **5. Deploy** (the livekit service is behind a `sfu` compose profile, so it stays dormant on a
 normal deploy — start it explicitly):

docker-compose.yml

@@ -44,7 +44,7 @@ services:
       - ./livekit.yaml:/etc/livekit.yaml:ro
     ports:
       - "7881:7881"          # WebRTC over TCP (fallback)
-      - "50000-50100:50000-50100/udp"  # WebRTC media (UDP) — must match livekit.yaml port range
+      - "50000:50000/udp"    # single WebRTC media UDP port (must match livekit.yaml rtc.udp_port)
     networks:
       - npm
 

livekit.yaml

@@ -7,10 +7,12 @@
 port: 7880                     # signaling (HTTP/WS) — NPM proxies wss://livekit.bizgaze.com -> here
 rtc:
   tcp_port: 7881               # WebRTC-over-TCP fallback (restrictive networks)
-  port_range_start: 50000      # WebRTC media (UDP) — keep in sync with the published range in compose
-  port_range_end: 50100
-  use_external_ip: true        # discover the public IP for ICE candidates (VPS). Or set node_ip below.
-  # node_ip: 118.95.33.89      # uncomment + set if use_external_ip can't detect the public IP
+  udp_port: 50000              # SINGLE UDP media port (all participants mux over it) — minimizes
+                               # the NAT port-forward to one UDP + one TCP port.
+  # This box sits behind NAT (private 192.168.88.61 behind public 118.95.33.89). Auto-detection
+  # would find the wrong (outbound) IP, so pin the inbound public IP clients actually reach.
+  use_external_ip: false
+  node_ip: 118.95.33.89
 
 # Embedded TURN over TLS on 443 helps clients on locked-down networks. Left off by default because
 # NPM already owns 443; enable via a dedicated hostname + NPM stream if you need it (see DEPLOY.md).