Thanks for your reply @kyrofa! Good to know that I was testing it as expected.
Regarding the decision of using “classic” confinement, that’s basically because of the hardware we are using. Our robot uses an in-house NVIDIA TK1 running on the top a L4T based on Ubuntu Trusty with a custom kernel, using CUDA and its own modified OpenCV libraries. Our options are really limited. 
We managed to create a snap containing the whole ROS workspace, including the necessary modified OpenCV libraries and successfully compiling against them. The problem comes when we try to launch the snap, as it seems that supporting CUDA in a snap is not that easy task. I found some posts on the topic, specifically this one.
Considering what you indicate there and that we are mainly interested in snap for the automated updates, rollbacks, channels and so on, but not that much in the confinement/isolation, we thought that giving it a try in “classic” could be a fast and clean way of solving our problem. Are we on the right track?