It's interesting to hear about Future LZ. It sounds promising. Maybe I missed something, but it sounds like forward references use uncompressed bytes, which doesn't seem as efficient as using compressed symbols to count. Perhaps there are benefits I am not aware of to using offsets based on uncompressed data rather than compressed data.
While algorithms such as bpe, diqqs, krc and tree are all offline, there doesn't seem to be anything preventing online recursive matching algorithms from being developed other than a loss of effectiveness. Maybe it's a silly point, but for that reason I am hesistant to include offline in the category name for these types of algorithms.