... currently we're using a count that is never actually flushed out, except on reconfig (in 2.1, potentially in 2.0).

Also we may want to rethink our logic of waiting until the majority of slaves (# slaves / 2 + 1) is caught up. Why not majority of the set, for example?