fix: bound DoH fallback-IP discovery with wall-clock deadline

[rkfshakti]

Problem

The Telegram gateway hangs indefinitely at Connecting to Telegram (attempt 1/8)… after Discovering Telegram API fallback IPs via DNS-over-HTTPS…. No timeout fires, no retry, no error — the process must be killed manually.

The fix from #64370 (_await_with_thread_deadline on initialize()) is confirmed present in the running image, but the issue still reproduces.

Root cause

The discover_fallback_ips() call in the connect path is not wrapped in _await_with_thread_deadline. While the DoH client has its own timeout, the system-DNS leg (asyncio.to_thread(socket.getaddrinfo)) can hang for minutes on broken VPN/DNS configs, blocking the event loop and preventing the "Connecting to Telegram" retry ladder from ever running.

The issue reporter sees the "Discovering Telegram API fallback IPs via DNS-over-HTTPS…" log, but the "Auto-discovered Telegram fallback IPs" log never appears — meaning discover_fallback_ips() itself is hanging, not initialize().

Fix

Wrap discover_fallback_ips() in _await_with_thread_deadline using the same HERMES_TELEGRAM_INIT_TIMEOUT (default 30s). On timeout, fall back to the seed fallback IPs (_SEED_FALLBACK_IPS) instead of hanging indefinitely.

This ensures the entire connect path — DoH discovery + initialize() — is bounded by wall-clock deadlines that don't depend on the event loop being responsive.

Fixes #74140

[teknium1]

Thanks for tracing the fallback-IP discovery path.

Automated hermes-sweeper review found that current main already ships this fix:

• plugins/platforms/telegram/telegram_network.py:242-258 starts system DNS in a worker thread, gathers DoH independently, and bounds the resolver wait with _DOH_TIMEOUT.
• plugins/platforms/telegram/telegram_network.py:280-285 already falls back to _SEED_FALLBACK_IPS when discovery has no usable DoH result.
• tests/gateway/test_telegram_network.py:467-490 covers a hung socket.getaddrinfo and verifies discovery returns without being gated on it.
• This was implemented by e16743b0d5 (fix(telegram): diagnose blocked-loop init hangs, unbind DoH from system DNS) and shipped in v2026.7.20.

Closing as implemented on main.

[Pachnoc]

@teknium1 was this an automated sweep and not a your human review? Not trying to be rude however, I'm running v2026.7.20 — the exact version cited as containing the fix. The code referenced is confirmed present in my image (verified via grep). The hang still reproduces every time. The existing fix wraps initialize() but discover_fallback_ips() hangs before initialize() is ever reached — the 'Auto-discovered Telegram fallback IPs' log line never appears in failing runs, only in the intermittent successes. Requesting another review if possible?