Hi Phil,
No definitive answers here, you should use what seems appropriate for testing your hypothesis... Some comments:
about 2-
You can transform IRFs to TLRC, you can even do your linear regression (deconvolution) on TLRCed time series, just don't use 1x1x1 for time series data or you'll end up with ginormous data sets (2x2x2 or 3x3x3 makes plenty more sense given the resolution of your EPI).
about 3-
You can always edit the clusters by hand to cut out some of the large clusters that appear to link across two regions of interest. You can also use the -1erode and -1dilate options of 3dmerge to separate sprawling clusters
about all-
For voxel-wise group stats, you are better off running the stats first on all voxels in the dataset (at least those inside the brain). Then focusing on the results in a particular ROI. Rather than apply the ROIs then perform the group statistics.
cheers,
-ziad