fix(telegram): retry polling on transient network errors

[snemesh]

Problem

The Telegram plugin's polling loop exits permanently on any error that isn't a 409 Conflict (line 991-992 in server.ts). A single transient network issue — ECONNRESET, ETIMEDOUT, DNS failure, or a Telegram 5xx — silently kills the bot. Messages stop arriving and the user must restart Claude Code to recover.

This is the root cause behind several open issues:

• Telegram plugin: polling loop exits permanently on network errors (no retry) #963 — polling loop exits permanently on network errors (no retry)
• Likely contributes to Telegram Plugin: Bot receives messages but Claude Code never responds #860, Telegram channel plugin: grammY polling not consuming updates #862, Telegram plugin: MCP server (bun) process silently dies mid-session with no auto-restart #777, Telegram plugin: MCP server silently dies on network interruption with no recovery — workaround included #917 — all describe the bot silently dying mid-session

Fix

• Added isTransientError() helper that classifies errors as retryable: Node.js network codes (ECONNRESET, ECONNREFUSED, ETIMEDOUT, ENETUNREACH, EAI_AGAIN, EPIPE), fetch/TLS failures, and Telegram 5xx server errors.
• The polling loop now retries with exponential backoff (up to 30s) on transient errors, matching the existing 409 Conflict retry behavior.
• Permanent errors (401 Unauthorized, invalid token) still cause immediate exit — no infinite retry on bad credentials.
• The backoff counter resets after a successful onStart callback, so intermittent flakes don't accumulate penalty across reconnections.
• Added shuttingDown guard to avoid retrying during graceful shutdown.

Changes

• external_plugins/telegram/server.ts — 43 insertions, 12 deletions

Test plan

• bun build passes (1.1MB bundle, 0 errors)
• Unit tests: isTransientError() correctly classifies 19/19 error types (ECONNRESET, ETIMEDOUT, EAI_AGAIN, fetch failed, 5xx → retry; 401, 400, 403, 409, unknown → exit)
• Integration tests: simulated polling loop verifies retry+recovery for 7 scenarios (network flakes, DNS failure, Telegram 500, 409 conflict, mixed transient→permanent)
• Live test: bot started with bun server.ts, process suspended with SIGSTOP for 35s to force Telegram to drop the long-poll connection, resumed with SIGCONT — bot recovered via retry loop and resumed polling

Unit test output (19/19 passed)

Testing isTransientError() classification:

  ✅ ECONNRESET           → RETRY
  ✅ ETIMEDOUT            → RETRY
  ✅ ECONNREFUSED         → RETRY
  ✅ EAI_AGAIN            → RETRY
  ✅ ENETUNREACH          → RETRY
  ✅ EPIPE                → RETRY
  ✅ socket hang up       → RETRY
  ✅ fetch failed         → RETRY
  ✅ network error        → RETRY
  ✅ TLS handshake        → RETRY
  ✅ Telegram 500         → RETRY
  ✅ Telegram 502         → RETRY
  ✅ Telegram 429         → EXIT
  ✅ Telegram 401         → EXIT
  ✅ Telegram 400         → EXIT
  ✅ Telegram 403         → EXIT
  ✅ Telegram 409         → EXIT
  ✅ random Error         → EXIT
  ✅ Aborted delay        → EXIT

19/19 tests passed

Integration test output (7/7 passed)

═══ Integration test: polling loop retry behavior ═══

✅ ECONNRESET x3 → recovers on 4th attempt
      0ms  bot.start() threw: read ECONNRESET
      0ms  transient error (attempt 1), retry in 100ms
    101ms  bot.start() threw: read ECONNRESET
    101ms  transient error (attempt 2), retry in 200ms
    302ms  bot.start() threw: read ECONNRESET
    302ms  transient error (attempt 3), retry in 300ms
    603ms  bot.start() succeeded — connected!
    654ms  clean exit (bot.stop called)

✅ DNS EAI_AGAIN x2 → recovers on 3rd attempt
      0ms  bot.start() threw: getaddrinfo EAI_AGAIN api.telegram.org
      0ms  transient error (attempt 1), retry in 100ms
    102ms  bot.start() threw: getaddrinfo EAI_AGAIN api.telegram.org
    102ms  transient error (attempt 2), retry in 200ms
    303ms  bot.start() succeeded — connected!
    354ms  clean exit (bot.stop called)

✅ Telegram 500 x1 → recovers on 2nd attempt
      0ms  bot.start() threw: ISE (500: Internal Server Error)
      0ms  transient error (attempt 1), retry in 100ms
    101ms  bot.start() succeeded — connected!
    152ms  clean exit (bot.stop called)

✅ 409 Conflict x2 → recovers on 3rd attempt (existing behavior preserved)

✅ 401 Unauthorized → exits permanently (no retry)
      0ms  bot.start() threw: Unauthorized (401: Unauthorized)
      0ms  exit: permanent error

✅ ECONNRESET → 401 Unauthorized → exits after transient retry

✅ fetch failed x2 → recovers

═══ Result: ALL PASSED ✅ ═══

Fixes #963