fix(streamable_http): handle ClientDisconnect during POST at WARNING level

[wiggzz]

Problem

When a client disconnects mid-POST (network timeout, cancel, LB drop), the MCP SDK's _handle_post_request catches ClientDisconnect in its generic except Exception handler, which:

1. Logs an ERROR ("Error handling POST request") to Datadog — this is pattern 1, 67% of all prod ERRORs in ai-codegen-api
2. Synthesizes a 500 response and sends it to the already-closed socket
3. Does not notify the inner session writer, so stateful session tasks can hang

Solution

Catch ClientDisconnect before the generic except Exception handler in _handle_post_request:

• Log at WARNING — honest level: we aborted handling a request, but it wasn't our fault
• Notify the session writer with ClientDisconnect() so inner session tasks unblock cleanly (stateful sessions)
• Skip sending a response — the socket is closed; sending to it is just noise
• Scope: POST only — Datadog confirms 100% of pattern 1 stacks originate from _handle_post_request

Tests

4 new unit tests in tests/server/streamable_http/test_client_disconnect_post.py:

• test_client_disconnect_logs_warning_not_error — verifies WARNING, not ERROR
• test_client_disconnect_does_not_send_response — verifies no ASGI sends to closed socket
• test_client_disconnect_notifies_writer — verifies writer receives ClientDisconnect
• test_client_disconnect_writer_suppresses_errors — verifies broken writer doesn't crash us

All existing streamable_http tests (49 + 18) pass.

Upstream context

This patch combines the approach of two open upstream PRs (neither merged):

• modelcontextprotocol/python-sdk#1647 — right scope (POST only) but wrong semantics (still sends response, doesn't notify writer)
• modelcontextprotocol/python-sdk#1947 — right semantics (notify writer, skip response) but broader scope than needed (covers GET/replay paths)

This is a tight ~10-line change taking the spirit of modelcontextprotocol#1647 + semantics of modelcontextprotocol#1947. Tracking issue: modelcontextprotocol/python-sdk#1648.

Follow-up

After this lands in the fork and is verified in prod via ai-codegen-api bump:

• Open mirror PR upstream
• Consider dropping ClientDisconnectHandlerMiddleware in ai-codegen-api (no longer load-bearing for POST path)