Drain MCP NIO response tasks before event-loop shutdown

[lynnswap]

Problem

The MCP HTTP suite intermittently logs:

ERROR: Cannot schedule tasks on an EventLoop that has already shut down. This will be upgraded to a forced crash in future SwiftNIO versions.

The warning appears around stopDrainsAdmittedHandlerBeforeClosingItsSession. It does not fail the current test, but SwiftNIO explicitly says it will become a crash.

Evidence

• Clean detached PR checkpoint 22874f3: reproduced in 1 of 3 full CodexReviewMCPHTTPServerTests runs; all 27 tests otherwise passed.
• Current Wave 3B replacement worktree: reproduced in 2 of 3 full MCP runs; all tests otherwise passed.
• Five isolated runs of stopDrainsAdmittedHandlerBeforeClosingItsSession did not reproduce, so this is suite-order/lifecycle timing rather than a deterministic assertion failure.

Required investigation

Trace NIO handler response/stream completion tasks and event-loop callbacks relative to listener/session close and shutdownGracefully(). The MCP lifecycle owner must await every Task that can call eventLoop.execute/write after shutdown. Do not suppress the warning or add delays.

Acceptance

• No scheduling occurs after event-loop shutdown.
• Repeated full MCP suite and full package runs emit no warning.
• Admission close -> handler/response flush drain -> session/channel close -> event-loop shutdown remains explicit.
• Deterministic task/flush ownership tests; no sleeps, polling, or warning filtering.