Alas, the small percentage is the case.
The alkanes in the ether set use the new aliphatic VDW radii, while the current lipid parameters do not. The new and improved lipid parameters, in final testing, will also use the new VDW radii. That said, the carbonyl parameters used may be more a more significant variation for the cyclic ketone. I'd probably build a couple of different cyclohexanone models, perhaps using CGenFF and "all32_lipid" sets, and compare normal mode results (VIBRAN) to each other and to expt.