Make long-running MCP tool calls respond to cancellation

Goal

When a user cancels a pending tool operation, have the server stop promptly or finish safely.

Steps

1. Identify long I/O or compute stages and define a safe post-cancellation state. [mcp-cancel]
2. In an async @mcp.tool() handler, put cleanup in finally; wrap awaited cleanup in anyio.move_on_after(5, shield=True). [mcp-cancel]
3. For a plain def tool running in a thread, call anyio.from_thread.check_cancelled() between work units; clients cancel the awaiting task or let its read timeout expire. [mcp-cancel]
4. Do not enable json_response=True on modern Streamable HTTP or stateless_http=True on legacy connections; both suppress handler cancellation. [mcp-cancel]
5. Cancel a slow test task and verify the handler stops, resources are released, and the result is marked cancelled. [mcp-cancel]

How to verify

After cancellation, async cleanup runs in `finally` within the five-second shielded bound, sync tools exit at a poll, and HTTP settings allow cancellation.

Troubleshooting

If cancellation is ignored, check whether the handler propagates the cancellation token to downstream I/O.