fix: use blocking stdin read instead of anyio.wrap_file() in stdio_server

[celstnblacc]

Summary

Replaces anyio.wrap_file() wrapping of stdin in stdio_server() with a blocking run_sync(raw_stdin.readline) loop.

Problem

anyio.wrap_file() wraps sys.stdin.buffer as an async iterator. When the MCP client closes its end of stdin between connection cycles, the async iterator sees EOF and exits the async for loop, killing the read stream and the server.

This manifests as -32000: Connection closed errors in MCP clients like Claude Code, OpenCode, and Codex CLI, requiring a client restart every time the connection resets.

Fix

The default stdin path now uses anyio.to_thread.run_sync(raw_stdin.readline) in a while True loop. Blocking readline() waits until the OS delivers more data, surviving transient client disconnects. An empty read (real process EOF) still cleanly exits the loop.

The existing custom-stdin path (used by tests and when callers pass an anyio.AsyncFile) is preserved unchanged.

Changes

• src/mcp/server/stdio.py: Default stdin path → blocking run_sync(readline) loop (37 insertions, 11 deletions)
• tests/server/test_stdio.py: Added test_stdio_server_survives_stdin_eof regression test (46 insertions)

Testing

All existing and new tests pass:

uv run --frozen pytest tests/server/test_stdio.py -v
tests/server/test_stdio.py::test_stdio_server PASSED
tests/server/test_stdio.py::test_stdio_server_invalid_utf8 PASSED
tests/server/test_stdio.py::test_stdio_server_survives_stdin_eof PASSED

Ruff lint: clean.