I'd like to learn how CouchDB's replication protocol works. Not at the  
implementation level (I don't read Erlang well enough for that!) but  
are there any design docs that describe what's happening  
architecturally?

Conceptually, I can imagine it being based on an "_all_docs_by_seq"  
query, where the pulling server remembers the last sequence ID it's  
pulled in the past, grabs all the newer docs, and puts them to its  
local database. Is there more to it than that?

—Jens