How can we ``scale down'' an N-node network G to a smaller network G', with K << N nodes, so that G' (approximately) maintains the important structural properties of G? There is a voluminous literature on many versions of this problem if K is given in advance, but one's tolerance for approximation (and the resulting value of K) will vary. Here, then, we formulate a ``rescalable'' version of this approximation task for complex networks. Specifically, we propose a \emph{node ordering} version of graph summarization: permute the nodes of G so that the subgraph induced by the first K nodes is a good size-K approximation of G, averaged over the full range of possible sizes K. We consider as a case study the phonological network of English words, and discover two natural word orders (word frequency and age of acquisition) that do a surprisingly good job of rescalably summarizing the lexicon.