As a sanity check, I scaled the size of my subproblems down to 1, and compared the performance of the BinInt and MaxOnes problems with identical parameters. Ta dah! Exactly the same results, as you would expect.
Now I'm dialing up the BinInt size gradually. The effect of scaling is obvious, as many bits converge too early, carried along on the coattails of the high value bits. This series of experiments is with a 0.0 mutation rate, to test the powers of a selectorecombinative GA in isolation. In a problem l bits in size, with population and number of generations also held at l for l^2 function evaluations in total, the best fitness over a run is not even reaching half the known optimum.
I'm also still learning lot's of basic things about ECJ, such as how to chart statistics using the gui console. Woot!