... which command should I use to threshold my group results?

Presumably you are going to threshold your group results in standardized 2x2x2 mm space that SPM uses, and inside the group mask "mask.hdr" (that is, voxels outside this mask will be ignored, even if above threshold). In which case, command A is the reasonable choice.

The other commands suffer from (i) not having a mask, so that the entire 3D volume is used rather than just a brain-shaped subset of it, and (ii) in command C, not using the voxel size at which the group results were calculated. 3dClustSim works by simulating noise-only random fields and then thresholding+clusterizing them, and so for the cluster-size threshold to make sense, the simulation should correspond to the situation where you will apply this result. Otherwise you are computing apples and applying them to oranges.

Finally, I'll add that 3dClustSim (and its companion, 3dFWHMx) are currently in a state of flux, to address the issues publicized here:
http://blogs.discovermagazine.com/neuroskeptic/2015/12/07/false-positive-fmri-revisited/#.Vp5C8VMrK-Y
At the present time, I have to recommend that people use the new "-acf" options to both programs to carry out the smoothness estimation and cluster simulation. I'm still working on this approach, and adding some wrinkles beyond that I commented on at Neuroskeptic, but those are not ready for general use -- they need to be finalized and then productized. You can find my abstract submitted for the 2016 HBM meeting at
https://goo.gl/XQeGdA
This document reflects my results as of the deadline, 11 Jan 2016. My thinking has evolved somewhat since then, based on thinking about Figures 3 and 4.