fix(mcp): drain the MCP loop and retain parked startup tasks for clean shutdown

[kshitijk4poor]

Summary

Ends the RuntimeError: Event loop is closed shutdown noise from parked MCP tasks (#60197) by combining the two complementary community fixes: the MCP loop owner now drains pending tasks before closing the loop, and a server parked after a recoverable initial connect failure is retained in _servers so shutdown owns it and /mcp-refresh revival actually works.

Root cause (per @shady2k's analysis on the issue): a server that exhausts _MAX_INITIAL_CONNECT_RETRIES parks its run() task to self-probe, but start() raises before _servers[name] is written — so shutdown_mcp_servers() never reaps the task, _stop_mcp_loop() closes the loop with the coroutine still suspended, and GC finalization drives t.cancel() into the closed loop. Under systemd this turns a clean SIGTERM stop into exit code 1 / failed unit state (see issue thread).

What's salvaged

• fix(mcp): drain parked tasks before closing loop #60104 by @seppegadeyne (+ drain commit by @shady2k, salvaged from fix(mcp): drain pending tasks before closing the MCP loop #66143) — cherry-picked as-is: _stop_mcp_loop() schedules a single loop-owned _drain_and_stop_mcp_loop() that cancels and boundedly reaps pending tasks, then stops the loop from its own thread, so an outer timeout can't let loop.stop() overtake the drain. Direct, blocked-loop, and end-to-end regressions included.
• fix(mcp): retain parked startup tasks for clean shutdown #62026 by @mrzlab630 — cherry-picked and reconciled onto current main: _connect_server claim contextvar + adoption of recoverable parked startup tasks into _servers; standalone probe failures still reap locally. The _register_discovered_tools_if_needed() hunks were reconciled with main's 106d1822e (kept main's _servers.get(self.name) is self ownership check, added the stale connect-error clear on successful publication).
• Follow-up commit (ours) — folds in the CancelledError rule from fix(mcp): reap orphaned parked server task in _connect_server #72054 by @adurham (start() reaps its own task on external cancellation; don't await a redundant shutdown() inside a cancelled context), logs orphan-reap failures, and replaces the nonlocal claim rebind with a list-based claim to keep the ty diff clean.

Validation

Check	Result
E2E reproducer (deterministic failing stdio server, real imports, subprocess) on upstream/main	dirty: RuntimeError: Event loop is closed ×2, Task was destroyed ×3, server absent from _servers
Same reproducer on this branch	clean: 0 noise lines, broken-test retained in _servers, exit 0
MCP test suites (tests/tools/ -k mcp, 812 tests)	all pass
New regressions (ownership/revival/terminal/probe + drain direct/bounded/blocked-loop + e2e shutdown)	all pass
ruff / git diff --check	clean
ty check tools/mcp_tool.py	10 diagnostics, identical to upstream/main baseline

Credit

• Loop drain: fix(mcp): drain parked tasks before closing loop #60104 by @seppegadeyne, incorporating fix(mcp): drain pending tasks before closing the MCP loop #66143 by @shady2k — commits cherry-picked, authorship preserved.
• Parked-task ownership: fix(mcp): retain parked startup tasks for clean shutdown #62026 by @mrzlab630 — commit cherry-picked, authorship preserved; conflict reconciliation against main's 106d1822e done here.
• CancelledError propagation rule: fix(mcp): reap orphaned parked server task in _connect_server #72054 by @adurham — folded into our follow-up commit with Co-authored-by.

Closes #60197