Your benchmark gives as input a list in sorted order. As you mentioned, mergesort is extremely good on sorted or nearly-sorted lists. You should shuffle the list first to avoid giving the builtin a slam-dunk.

I tried it, and in a couple of cases limbic gets comparable times and even beats baseline:

Looking for top 5 in 10000 (running for 9.5 CPU secs)
            Rate aristotle browseruk    limbic  baseline
aristotle 13.7/s        --      -46%      -89%      -92%
browseruk 25.2/s       84%        --      -81%      -85%
limbic     130/s      847%      414%        --      -23%
baseline   169/s     1134%      571%       30%        --
 
Looking for top 5 in 100000 (running for 15 CPU secs)

            Rate aristotle browseruk  baseline    limbic
aristotle 1.38/s        --      -44%      -85%      -90%
browseruk 2.48/s       79%        --      -73%      -82%
baseline  9.14/s      560%      268%        --      -35%
limbic    14.0/s      913%      465%       53%        --
[download]