Planned investigations
I would also suggest, to relabel old variant and put it in rumble. I feel that old rumble stats had a lot of disabled bots. It might appear that old version is better but it is actually not. Plus we have at least 3 new above top 100 bots which might pull your APS down.
That is true, it is also my intention to do so. I always compare against the older version via Botdetails, it let you see the differences betweeen the common pairings, so that really is the difference, even if only 1000 bots are in common. Just remove the 'd' from versionnumber and press Compare.
My main worry here is about difference in liteclient performance. Most of the old rankings were done with ancient version of robocode. It is possible that your old bot was tuned to exploit a glitch without you even knowing about.
I do know that some bots don't work very well in Robocode 1.9.2.x and/or Java 8. See f.e. Xiongan.xiongan and tcf.Drifter, they get scores of 0 against GrubbmThree, while on my system (1.9.2.5 and Java 7) they just work ok.
There is else but Java 7 vs 8. I ran tcf.Drifter vs. GrubbmThree 0.9d in robocode GUI with Java 8. Drifter constantly wins the match.
So I cannot attribute 100% loss which we see in rumble just to the Java version.
I did read somewhere that tcf.Drifter had some problem with the 1.9.2.5_beta versions, maybe someone is running the rumble with a non-official 1.9.2.5 version ?