Sure, but what I'm hearing in the parent post is a question about why we don't use linguistically motivated subword units (of similar length/vocabulary size and

Sure, but what I'm hearing in the parent post is a question about why we don't use linguistically motivated subword units (of similar length/vocabulary size and thus memory usage) e.g. cutting across morpheme boundaries instead of whatever an algorithm like BPE caclulates.