Hmm, so it seems the situation has at least somewhat improved. That’s good to know, most of the references I found were not looking promising initially, and that particular issue discussion had strong WONTFIX undertones to it.
As a side note, it’s kind of disappointing that the state of compression options in DNG without doing something non-standardized has kind of stagnated. Unless you are Bayer-mosaiced you’re basically limited to 8-bit lossy JPEG. I’m curious what feeding the lj92 predictors into a more modern entropy coder (such as the underlying ones used by zstd) would result in performance-wise.