Make the TrainingPeaks Hub app connect in seconds: pasta networking + LAN re-advert

[BenA-SA]

Summary

Makes the TrainingPeaks Hub companion app connect to TPV in the container within seconds, for every user and with no manual steps. Hub's In Game remote has Gear Up/Down and Difficulty −/+, which covers the handlebar-control goal in #3.

Problem: Hub discovers TPV via mDNS (_tpvirtual._tcp) and connects in on TCP 7779. TPV advertises an address for every interface Wine reports (172.17.0.1 Docker bridge, 100.89.193.106 Tailscale, 192.168.0.146 Wi-Fi). Hub tries only the first and waits out Android's ~127 s TCP connect timeout. Meanwhile it shows "please start a ride".

Fix:

1. run-tpv.sh runs the container on pasta networking bound to the LAN interface, so TPV only sees that interface.
2. The entrypoint re-advertises TPV through the host avahi-daemon, pointing at the LAN address. TPV's own advert doesn't reach the phone under pasta, so this is required.

Changes

• run-tpv.sh: --network=pasta:-i,<LAN if> -p 7779:7779 --hostname <host>.
  • The LAN interface is TPV_LAN_IF, or by default the default-route interface.
  • TPV_NETWORK=host, or qz mode, keeps host networking.
  • Also passes TPV_HOST_NAME / TPV_HUB_ADVERT / TPV_HUB_ADDR / TPV_HUB_NAME.
• scripts/entrypoint.sh:
  • advertise_hub runs before the TPV launch, and stop_hub_advert after wineserver -w.
  • It publishes <hostname>-tpv-<last octet>.local and <hostname> (LAN)._tpvirtual._tcp port 7779 txtvers=1.
• Containerfile.full: adds avahi-utils (image rebuild required).
• verify.sh: check 6 flags each _tpvirtual._tcp IPv4 entry OK/BAD. Under pasta, checks 3 and 5 run inside the container, because QZ's ports live in its network namespace.
• README: new env vars, the real cause in "Hub app cannot find TPV", and the updated podman flags table.
• docs/adr/0001-…: the full diagnosis, including a red herring (the advert name) and the measurements below.

Testing

Cold starts of Hub, driven over adb: force-stop, wait 20 s for the Android NSD cache to expire, launch. Connection times are taken from the phone's netstat and TPV's Player.log:

Setup	Hub's first attempt	Connected
--network=host + re-advert (Docker + Tailscale present)	172.17.0.1:7779 SYN_SENT	132 / 133 / 133 s
--network=host, docker0 deleted	100.89.193.106 (Tailscale, reachable by luck)	3–4 s ×3
pasta on wlp0s20f3 + re-advert (Docker + Tailscale present)	192.168.0.146:7779	3 / 4 / 3 s

• ✅ Under pasta, Wine's ipconfig shows only wlp0s20f3 and lo.
• ✅ The rebased branch builds (tpv-full:pr8).
• ⏳ Still to do before merging (needs the bike):
  • A ride in both mode with the real trainer under pasta. QZ's DIRCON advert goes through host avahi over D-Bus and QZ shares the container's network namespace, so it should work. It's the main risk.

    cd ~/src/tpv-qz-container && git checkout feat/hub-readvert && git pull
    IMAGE=localhost/tpv-full:pr8 ./run-tpv.sh        # expect "==> network: pasta on wlp0s20f3"
    ./verify.sh                                       # checks 2-6 should be OK

    Pair the trainer and ride a few minutes (power/cadence arrive, ERG responds).
  • Hub Live and the In Game remote during that ride (close and reopen Hub: it should connect within a few seconds).
  • Optional: a 60-minute ride watching for "Lost heartbeat" drops.

Known limitation

Trainers that advertise themselves on the LAN (a Wi-Fi KICKR, a remote DIRCON bridge) probably aren't discovered under pasta, because inbound multicast isn't forwarded. The workaround is TPV_NETWORK=host. Tracked in #13.

Notes

• The manual tpv-hub-addr / tpv-hub-svc workaround units used before this PR are no longer running and aren't needed with it.
• A report for TrainingPeaks (Hub should try all resolved addresses; TPV should advertise only reachable ones) is drafted separately.

Refs #3

🤖 Generated with Claude Code

[BenA-SA]

@sentry review