fix(mcp): silence Event loop is closed on reload parked teardown

[buckster123]

Summary

• /reload-mcp (and any full MCP teardown path) could print a noisy ignored exception when tearing down parked MCP servers:

Exception ignored in: <coroutine object MCPServerTask.run ...>
  File tools/mcp_tool.py, in _wait_for_reconnect_or_shutdown
    t.cancel()
RuntimeError: Event loop is closed

• Happens when run() parks after a permanent/initial connect failure: start() raises, the server is not added to _servers, then reload closes the MCP loop while the parked wait helper's finally still calls Future.cancel().

Changes

1. _safe_cancel_await — cancel + await wait-helper futures without raising if the loop is already closed. Used by:
   • _wait_for_lifecycle_event
   • _wait_for_reconnect_or_shutdown
   • _wait_for_lazy_reconnect
2. _parked_failed_servers — if start() fails but leaves run() alive (parked for revival), track it so full MCP teardown still calls server.shutdown().
3. _stop_mcp_loop — cancel leftover tasks on the loop before loop.stop(), so any remaining orphans exit while the loop is still alive.

Test plan

• uv run --extra dev python -m pytest tests/tools/test_mcp_stability.py -q → 30 passed
• Manual: configure a failing MCP stdio server (bad command), let it park, run /reload-mcp — no Event loop is closed traceback; healthy servers reconnect.

Context

Observed while hot-swapping Hermes MCP config (dropping failed servers + swapping Cerebro to a Rust binary). Reload still succeeded (Reconnected: blender, cerebro, sonus); the traceback was cosmetic but scary.

[alt-glitch]

This was generated by AI during triage.

Related: overlaps #61466 (safe cancellation) and #60104 (drain before loop close), but this patch also tracks failed-start parked servers for explicit teardown. These are competing lifecycle mechanisms; maintainer choice is needed.

[teknium1]

Thanks for the careful MCP lifecycle investigation. This is now implemented on current main through a stronger combined lifecycle solution.

• c00a1d58d505123a64fef7e2f0500f6e483e8193 retains recoverably parked failed-start tasks in _servers for shutdown/revival at tools/mcp_tool.py:5861-5873.
• ab0d3fac3d4cde14333d4107d8b5ae1475399a04 drains pending MCP-loop tasks on their owning loop before closing it at tools/mcp_tool.py:6781-6822.
• tests/tools/test_mcp_stability.py:526-612 covers an unregistered parked task and verifies its cleanup completes before loop.stop().

This is an automated hermes-sweeper review.