I investigated the performance of the task on a small dataset by inspecting the cProfile statistics. The bulk of the 26 second runtime is taken by convolve, warpExposure, and persist. Sorting by cumulative time (which includes the time in that function and all called functions), the MaxPsfWcsSelectImagesTask.runDataRef call takes only 0.3 seconds.
I therefore conclude that any potential savings from changing the code covered by this ticket (including either of the two fixes identified above) would give only ~1% speedup, so I am marking this as "Won't Fix."
(There is some possibility that performance hot spots in this module would appear with a larger number of input images, but it's not obvious to me how or if it would be dominant. I will check the profiles on larger runs as needed.)