control-connection fallback: preserve session keyspace

[dkropachev]

Problem

A session created with cluster.connect("ks") stores its keyspace:

python-driver/cassandra/cluster.py

Lines 1744 to 1749 in b666e5c

	def connect(self, keyspace=None, wait_for_all_pools=False):
	"""
	Creates and returns a new :class:`~.Session` object.
	If `keyspace` is specified, that keyspace will be the default keyspace for
	operations on the ``Session``.

python-driver/cassandra/cluster.py

Lines 2619 to 2622 in b666e5c

	def __init__(self, cluster, hosts, keyspace=None):
	self.cluster = cluster
	self.hosts = hosts
	self.keyspace = keyspace

When no node pool is available, fallback sends directly through the shared control connection without applying that keyspace:

python-driver/cassandra/cluster.py

Lines 5065 to 5096 in b666e5c

	def _query_control_connection(self, message=None, cb=None, connection=None, host=None):
	self._control_connection_query_attempted = True
	if message is None:
	message = self.message
	if connection is None:
	control_connection = self.session.cluster.control_connection
	connection = control_connection._connection if control_connection else None
	if not connection:
	self._errors['control connection'] = ConnectionException("Control connection is not connected")
	return None
	if host is None:
	host = self.session.cluster.get_control_connection_host() or connection.endpoint
	self._current_host = host
	request_id = None
	request_sent = False
	try:
	request_id = self._borrow_control_connection(connection)
	self._connection = connection
	result_meta = self._bound_result_metadata
	if cb is None:
	cb = partial(self._set_result, host, connection, None)
	cb = partial(self._handle_control_connection_response, connection, cb)
	log.debug("No usable node pools; falling back to control connection for host %s", host)
	self.request_encoded_size = connection.send_msg(message, request_id, cb=cb,
	encoder=self._protocol_handler.encode_message,
	decoder=self._protocol_handler.decode_message,
	result_metadata=result_meta)

Unqualified queries and prepares can therefore fail or execute against stale keyspace state on the shared connection.

The control connection is multiplexed between concurrent requests, while USE changes connection-wide state. It cannot safely serve fallback requests from sessions using different keyspaces without serializing complete request lifetimes or introducing separate connections.

Expected behavior

Keep control-connection fallback deliberately single-keyspace:

• The first session that uses fallback binds application use of the shared control connection to its keyspace, including no keyspace.
• The binding lasts for the lifetime of the Cluster and survives physical control-connection replacement.
• Other sessions may use fallback only when their keyspace matches the binding. A session using a different keyspace is rejected immediately with InvalidRequest rather than risking execution in the wrong keyspace.
• Before the first application request on a fresh or reconnected physical control connection, select the bound keyspace when one is configured.
• Reject an explicit USE statement on the fallback path, including the statement issued by Session.set_keyspace(), because it would mutate connection-wide state beneath every session sharing the fallback.
• Requests served by ordinary node pools keep their existing behavior.

Add coverage for same-keyspace sharing, conflicting keyspaces, keyspace-less sessions, explicit USE, prepared statements, and control-connection replacement. No fallback request may execute using keyspace state established by another session.