1. there may be some environments where there is a formal customer care agent downstream in the workflow perhaps who doesn't understand (or need to understand) how Social Contacts are filtered; therefore the training UI is not relevant important too them the tool allows it to be "hidden" from this type of user by going out of training mode.
 
For example, we have expect some users to only view the Campaign Results gadget with a specific tag.  they would only get work that was traiged by someone else.  Perhaps there is a PR/Marketing person that is looking at the full stream and they are tagging work as "support" and there is a support person who only see's work show up when it's tagged support.  That support person would not need the complexity of the trianing mode information.
 
2. Yes, ultimately this makes sense.  We implemented training as a distinct operation as a means to first "experiment" with the bayesian filter technology.  Once we're comfortable that it's working well, we plan to offer a "auto-train" mode which would train without the up/down arrows.  For example, we'd train-in and SC's that marked handled and we would train-out any SC's that are discarded.
 
3. Yes, the trained out message is a response that indicates success of training.  We do not permanently mark the SC as "trained out".  We have had other users ask that we perserve this as a state so it is an enhancement we have on our backlog.