You are right: While GeertJohan is concentrating on a new diff algo for NLG which is a big task and takes time, I was digging into the code we use now for a better understanding how it reacts and how it could be modified as a quick solution with low risk until the new algo is ready to go live.
Analysing the DGW3 code I found 2 smaller mistakes leading to a max factor of ~3.13 instead of 3.0 (not a big deal) but the set of 50 blocks thsminer posted earlier smelled like I needed more data. Ask and you will be given thsminer was sending me data of 2000! blocks a few days ago.
On the first view it looked like it was basically in line with what I was expecting based on the code.
Running some calculations on this big pile of records to compare the diff of the block against what I was expecting from the DGW3 code was the easiest part of the work. Reading through all those lines of numbers took a lot of time and coffee, but finally I found a row of blocks where the diff changes still seems far too late and far slower than they should be; right now I am trying to understand why and how this happened. Understanding those problems seems necessary to me to prevent them showing up again in the future.