What is it called to cluster some inputs, then classify other inputs into those clusters? I am learning about the problem of whole-book recognition, which is tangential to optical character recognition. Some of the strategies used to identify printed characters rely on first unsupervised clustering of letters by their visual characteristics, then solving the resulting cryptogram to create a training set, then performing a supervised classification for the rest of the data.
As far as I can tell, this kind of bootstrapping is related to but distinct from semi-supervised learning, in which a subset of training labels are available from the get-go.
Does such a two-step learning process, unsupervised and then supervised, have a specific name?
 A: I haven't seen a common name for this practice, but I typically call it "cluster classification". The idea is to perform unsupervised clustering on one set of data, build a classifier to identify those clusters in that data, and then apply the classifier to another set of data as a means to identify the original clusters in new data. 
This allows you to find a consistent set of clusters across datasets, as performing unsupervised clustering on each dataset individually (or combined) might yield different clusters than the original. The method of clustering followed by classifier building sidesteps this problem by fixing the clusters in the first step.
Note that there isn't really an unbiased estimator of classifier performance in this case. When applying the classifier to unseen data, you have no ground truth labels to compare against. If you include data in the clustering, but then try to exclude it in the classifier training step, your performance metrics will be overoptimistic, since you've used the test data in the clustering in the first place, and your classifier will just be recovering those differences.
