It's not a bug in Python's ssl module. If I understand David's approach correctly, then he is using the SNI callback the wrong way. By using it the wrong way he has discovered a threading bug in OpenSSL. There is some kind of race condition going on in which two threads free and replace the RSA private key at the same time.

I'm -1 to try to make the SSLContext object magically read-only.

David, which examples did you read? The documentation https://docs.python.org/3/library/ssl.html#ssl.SSLContext.set_servername_callback is pretty clear:

A typical use of this callback is to change the ssl.SSLSocket‘s SSLSocket.context attribute to a new object of type SSLContext representing a certificate chain that matches the server name.

Apache mod_ssl does it correctly, the first hit on stack overflow, too. https://stackoverflow.com/questions/5113333/how-to-implement-server-name-indication-sni