I agree with the assessment in the "dumps" section. The idea is to use "blob addresses" in the place where currently "text ids" are used. Instead of looking up text in the text table, blob addresses are looked up using a BlobStore.
For Wikimedia sites, using ExternalStore, the access pattern of fetchText.php was: select row from text table, load from External store. This behavior will stay exactly the same when using BlobStore.
The dumpTextPass script will have to join three tables (revision, slots, and content) instead of using just the revision table, but that should not have much of an impact.