fix(mcp): drain connecting tasks before loop shutdown

[nuch1011]

Summary

• drain in-flight MCP tasks on their dedicated event loop before closing it
• make MCP task submission atomic with loop shutdown and clear stale connecting state
• add regression coverage for connecting-server cancellation and the schedule/shutdown race

Root cause

shutdown_mcp_servers() only knew about fully registered servers. A fast one-shot command could exit while background discovery still had an MCPServerTask.run() coroutine connecting. The loop was closed first, so later coroutine finalization tried to cancel work on a closed event loop and printed RuntimeError: Event loop is closed despite exit code 0.

Verification

• scripts/run_tests.sh tests/tools/test_mcp_tool.py tests/tools/test_mcp_stdio_watchdog.py — 219 passed
• scripts/run_tests.sh tests/hermes_cli/test_tui_resume_flow.py — 71 passed
• .venv/bin/ruff check tools/mcp_tool.py tests/tools/test_mcp_tool.py — passed
• real configured-MCP smoke: hermes chat -q returned the expected response and exited 0 with no traceback, pending-task warning, or event-loop-closed error
• same smoke against origin/main reproduced Exception ignored in: <coroutine object MCPServerTask.run ...> / RuntimeError: Event loop is closed

No credentials or secret values are included in this PR.

[alt-glitch]

This was generated by AI during triage.

Related to #60104 and #61466. This branch covers the connecting-task/scheduling race; #60104 has a broader pre-close pending-task drain, while #61466 guards cancellation after closure. Maintainer choice is needed.

[GottZ]

This was generated by AI during triage.

• fix(mcp): drain connecting tasks before loop shutdown #67956 — fixes: drains in-flight connecting MCP tasks before closing the dedicated loop and blocks late loop restart on one-shot teardown, resolving the RuntimeError: Event loop is closed cleanup traceback.
• fix(mcp): shut down one-shot MCP runtime cleanly #67961 — fixes: byte-for-byte identical diff to fix(mcp): drain connecting tasks before loop shutdown #67956 (same author, same 5 files, same 95/31 + 120/0 changes) addressing the same issue with the same code.

Recommend consolidating on #67956 because it is the older of the two identical PRs (opened ~21 min earlier); #67961 can be closed as a duplicate. Neither has CI checks configured, both are mergeable. Reviewers: note that the one-shot drain in _stop_mcp_loop runs outside _lock and is timeout-bounded (5s drain + 5s join), so worst case is a bounded teardown delay on a hung loop thread, not a deadlock.

[nuch1011]

Closed at the repository owner's direction. Superseded patch remains private; no change should be proposed against the official NousResearch branch.