Summary

• add a lazy conversation_loop wrapper for run_agent._pool_may_recover_from_rate_limit
• preserve run_agent patchability after the run_conversation extraction
• cover the gateway/Discord fallback NameError path with a regression test

Reference

Observed in local Discord gateway logs when Codex primary hit HTTP 429 and fallback_model was configured:

agent.conversation_loop: API call failed ... provider=openai-codex ... HTTP 429
File "agent/conversation_loop.py", line 2254, in run_conversation
    pool_may_recover = _pool_may_recover_from_rate_limit(...)
NameError: name '_pool_may_recover_from_rate_limit' is not defined

The helper lives in run_agent.py, but the extracted conversation loop referenced it as a module-local global. Gateway sessions call the extracted loop directly, so the eager fallback branch crashed before activating the configured fallback provider.

Test Plan

• /usr/local/lib/hermes-agent/venv/bin/python -m pytest tests/run_agent/test_provider_fallback.py tests/agent/test_gemini_fast_fallback.py -q -o 'addopts='
• /usr/local/lib/hermes-agent/venv/bin/python -m py_compile agent/conversation_loop.py run_agent.py
• git diff --check