fix: ignore BrokenPipeError from console prints in gateway mode

[Krlosimc]

Summary

• prevent gateway/Telegram conversations from failing when internal console prints raise BrokenPipeError
• this surfaced to users as the generic 'unexpected error' response

Root cause

In gateway mode, run_agent.py can still emit debug/progress print() calls even with quiet_mode=True. If stdout/stderr is detached or its pipe closes, those prints raise BrokenPipeError, abort the turn, and gateway/run.py returns the generic error message to Telegram.

Fix

• wrap module-level print() in run_agent.py
• delegate to builtins.print()
• swallow BrokenPipeError so best-effort console output cannot abort the conversation

Validation

• ./venv/bin/python -m py_compile run_agent.py
• restarted the local gateway and verified Telegram reconnects cleanly
• reproduced from logs that the previous failure path came from BrokenPipeError in run_agent.py

[teknium1]

Thanks for the report and patch here. Current main already landed a broader _SafeWriter-based fix for broken stdout during run_conversation, so I pulled the remaining init-time edge case you surfaced into PR #1271 instead of merging this stale branch directly. That follow-up fixes the Honcho init-time print path on current main and keeps the existing _SafeWriter approach as the single mechanism. Credit for surfacing the remaining gap.