I think the biggest effect here is that of step size, as per your previous question. Upsampling means the nominal voxel size provided to tckgen is half what it was, so you’ll be using half the step size on essentially the same data. This will reduce the dispersion during tracking, and probably makes it less likely to branch, etc. This might explain the algorithm’s reduced ability to follow all the hard to track subcortical projections.
I had a brief look into the effect of upsampling on tractography in my 2012 MRtrix paper, and found it made very little difference - when using the same step size (the default in MRtrix 0.2 was a hard 0.2mm, independent of the voxel size). So I reckon you should try setting the step size to the same as would have been used with the original data. You may also need to modify the angle constraint to match (should be 45° by default in iFOD2). That would be a fairer comparison if you’re trying to evaluate the effects of upsampling alone.
BTW, all this relates to the use of probabilistic algorithms. With deterministic algorithms, you may well find that upsampling reduces interpolation errors. Although I’m not entirely convinced that the results will necessarily be noticeably better in practice…