fix(telegram): prevent connect hang with retry watchdog and fresh app per attempt (#67498)

[webtecnica]

Summary

Fixes #67498 — Telegram gateway hangs forever at 'Connecting to Telegram (attempt 1/8)...' with all event loop threads idle in select(). The retry loop silently exits without logging attempt 2/8 or timeout.

Root Cause Analysis

Three bugs in the retry loop inside connect() (plugins/platforms/telegram/adapter.py):

1. Reused self._app across all 8 retry attempts

The Application object was built ONCE before the retry loop. After a failed initialize() the app could be in a partially-initialized state where:

• _shutdown_abandoned_app() had closed the httpx transports (on TimeoutError)
• PTB's _initialized flag might have been set before the failure, causing subsequent initialize() calls to return immediately (skip real init) and hit break

2. CancelledError propagated silently

asyncio.CancelledError is a BaseException, not an Exception. None of the three except handlers caught it, so it propagated through the retry loop with zero logging — the task driving the loop could silently exit with no trace.

3. No total watchdog for the entire retry loop

Only per-attempt timeouts existed (via _await_with_thread_deadline, 30s default each). If the loop itself stalled between attempts (during asyncio.sleep(), cleanup, or scheduling), there was no mechanism to bound the total connect time.

Changes

1. Total watchdog deadline: Computes a total deadline (8 × init_timeout + 120s margin) before the loop. Each attempt checks the wall clock against this deadline and raises OSError immediately if exceeded — ensures the retry ladder always terminates.

2. Fresh Application per retry: On each failed attempt (signal via rebuild_app flag), the finally block rebuilds self._app via builder.build(), re-registers all handlers, and best-effort shuts down the old app. Each retry starts with a clean Application — no stale state leaking across attempts.

3. BaseException logging: Added except BaseException: (placed LAST after Exception, OSError, TimeoutError) that logs CancelledError and other non-Exception signals before propagating. Prevents silent task exit without diagnostics.

4. App rebuild in finally: The finally: clause runs after every failed attempt, ensuring the app is always rebuilt regardless of which exception class caused the failure.

Testing

• Syntax verified with ast.parse
• Existing tests continue to pass
• The fix targets the hang at source: even if the event loop stalls between attempts, the total watchdog catches it; even if CancelledError arrives, it's logged; each retry starts with a clean app

[kshitijk4poor]

Merged via #70884. Your commits cherry-picked with authorship preserved.

Note: the commits were authored under the "Hermes Agent" identity (agent@hermes.dev) — added a contributor email mapping so CI attribution passes. For future PRs, consider setting your git identity before committing so your commits carry your name.

Thanks for the thorough root cause analysis on the connect hang — the three-bug breakdown (reused app, CancelledError propagation, missing total watchdog) was spot on.