I would like to suggest wrapping internal I/O exceptions in a custom exception (i.e. MongoCommunicationException).
Imagine the following scenario:

• I connect to a replica set and begin perform queries
• Suddenly, the server I was querying goes down and the NetworkStream.Read() throws IOException, which gets to my code
• I have no knowledge which exceptions could be thrown from the below code and if I should retry the operation (catching all types of exceptions, especially in a hosted environment is very risky, think ThreadAbortException for example).

But, if the connection management layer of the driver would wrap all I/O-related exceptions into one type of exceptions (a new one, MongoCommunicationException or even the existing one MongoConnectionException), the driver user would instantly know that there was a problem with MongoDB connection, not some fluke in his code, and retrying/reporting failure becomes much easier.