fix: restart mcp loop during reload

[itsflownium]

Summary

• Add a serialized MCP reload helper that fully shuts down existing servers before rediscovery.
• Wait for MCP loop startup readiness and for loop-thread shutdown before closing/reusing loop state.
• Route TUI, CLI, and gateway reload paths through the shared helper.

Validation

• scripts/run_tests.sh tests/tools/test_mcp_stability.py tests/gateway/test_mcp_reload_refreshes_cached_agents.py
• scripts/run_tests.sh tests/tools/test_mcp_tool.py tests/tools/test_mcp_cancelled_error_propagation.py
• scripts/run_tests.sh tests/cli/test_cli_loading_indicator.py tests/hermes_cli/test_mcp_reload_confirm_gate.py
• $HOME/.hermes/hermes-agent/venv/bin/python -m py_compile tools/mcp_tool.py tui_gateway/server.py cli.py gateway/run.py tests/tools/test_mcp_stability.py tests/gateway/test_mcp_reload_refreshes_cached_agents.py
• $HOME/.hermes/hermes-agent/venv/bin/ruff check tools/mcp_tool.py tui_gateway/server.py cli.py gateway/run.py tests/tools/test_mcp_stability.py tests/gateway/test_mcp_reload_refreshes_cached_agents.py

Refs #46512

[liuhao1024]

Verification: looks good

Reviewed the full diff. The serialization approach is well-structured:

• _mcp_lifecycle_lock (RLock) guards loop lifecycle; _mcp_reload_lock (Lock) serializes full reload cycles — no deadlock since they are separate lock objects
• _ensure_mcp_loop now uses threading.Event for ready signaling with a bounded 2s timeout — good fail-safe
• _stop_mcp_loop returns bool and uses a bounded 15s join — the old 5s hardcoded timeout was tight for stdio transports with anyio cleanup
• reload_mcp_servers() raises RuntimeError if the old loop did not stop cleanly — prevents discovery on a partially shut down loop
• All three UI entry points (CLI, gateway, TUI) updated to use reload_mcp_servers() instead of manually stitching shutdown+discover

One minor note: the test test_reload_mcp_servers_discovers_on_fresh_loop verifies the old loop thread is joined and a new loop is created, which is the critical path. Test coverage looks solid.

[teknium1]

Thanks for consolidating the reload lifecycle behind one helper. The premise remains present on current main: tools/mcp_tool.py:5611-5622 clears loop references and joins for five seconds without checking whether the old thread is still alive, while cli.py:10757-10760, gateway/run.py:14080-14083, and tui_gateway/server.py:11560-11561 immediately rediscover afterward.

Problems

• The added success-path test in tests/tools/test_mcp_stability.py does not exercise the helper's new refusal path: shutdown_mcp_servers() -> False must raise and must not call discovery.

Suggested changes

• Add a unit test for that failed-stop branch, asserting RuntimeError and that discover_mcp_tools() was not called.

Automated hermes-sweeper review.