New skill for exposing an mm-backed agent over Telegram: in-process
long-polling (not an LLM-reasoned relay - that path doesn't guarantee
determinism, learned the hard way building a reference implementation),
a fail-closed allow-list, and an explicit callout for deciding whether
Telegram-triggered actions bypass the agent's own action-authority gate.

- SKILL.md: overview + the two callouts (deterministic-relay rationale,
  access-control decision)
- workflows/telegram-bridge.md: the concrete flow, gate-bypass code
  pattern, edge cases
- scripts/telegram_poller.py: runnable skeleton, call_agent() left as the
  integration point for whatever agent runner the caller uses

Distilled from a real, mainnet-verified deployment (not included here -
generic pattern only).