fix(stdio): use os.dup to avoid closing real stdin/stdout on server exit

[jrlprost]

Fixes #1933.

Root cause

stdio_server() wraps sys.stdin.buffer in a TextIOWrapper. When that wrapper is garbage-collected after the context manager exits, Python's finalizer calls close() on it, which propagates down and closes sys.stdin.buffer (fd 0). Any subsequent print() or sys.stdin.read() then raises ValueError: I/O operation on closed file.

The existing comment ("Purposely not using context managers for these, as we don't want to close standard process handles") shows the intent was correct, but the wrapper's __del__ closes the buffer regardless.

Fix

Duplicate the file descriptor with os.dup() before wrapping. Our TextIOWrapper now owns a private copy of the fd; closing it leaves the original sys.stdin/sys.stdout untouched.

Falls back to the original buffer when the stream is not backed by a real fd (e.g. BytesIO in tests), matching prior behaviour in that case. The dup'd wrappers are explicitly closed in a finally block after the task group exits.

Changes

• src/mcp/server/stdio.py: os.dup() the fd; track ownership with stdin_created/stdout_created; close in finally.
• tests/server/test_stdio.py: regression test using real OS pipes — asserts sys.stdin.buffer and sys.stdout.buffer are still open after stdio_server() exits.

Testing

uv run --frozen pytest tests/server/test_stdio.py -v  # 3/3 pass
uv run --frozen ruff check .                           # no issues
uv run --frozen pyright src/mcp/server/stdio.py tests/server/test_stdio.py  # 0 errors

Branch coverage for src/mcp/server/stdio.py: 100 %.

The except io.UnsupportedOperation branch is covered by the existing test_stdio_server_invalid_utf8 test (BytesIO stdin). The new test covers the try success path with real pipe fds.

[maxisbey]

closing as there's already an existing PR for this issue