You are correct. The iterations that occurs is for training the classifier but it will only produce one classified output per image. Since your 10 images are contained in a single file, it will apply the trainer on the full 10-band data set and produce one classification. ENVI does not have a tool like you describe (ensemble SVM) but you might be able to split up your ROI samples using something like Generate Random Sample: http://www.harrisgeospati...ingRandomSample.html
Then use the different sets of ROIs in the trainer and then classify. It may require a significantly large ROI to generate an adequate number of samples for training.