Your plan is the right shape - subset the WGS down to the array backbone, impute everything through one pipeline and one reference panel, then match. Two additions.

The thing that will bite you isn't technical: platform is perfectly confounded with group, since one group is WGS, the other array, and your UKB controls are array. Any PRS difference you see could be platform rather than biology, and no amount of harmonising removes a confound sitting exactly on top of your comparison.

Which is where the useful bit comes in - UK Biobank has both WGS and array data on the same participants. Take a few thousand of them, run one copy through your WGS arm and one through your array arm, and you get a direct empirical estimate of how much the pipeline itself shifts the PRS. That's a calibration you can actually report, and it costs you nothing.

Also compute imputation R2 within each batch separately and keep only variants that pass in all of them, and report what fraction of your PRS weights survived - differential shrinkage between the groups would invalidate the comparison.