Thanks again for your help, Tamas! I could easily be misunderstanding something but we've done VBM with SPM in our lab per the tutorials online by John Ashburner et al and they all discuss how the use of the smoothing kernel at the first level helps to meet the assumptions of GRF to allow for correction for multiple comparisons at the second level. (e.g here: https://www.fil.ion.ucl.ac.uk/~john/misc/VBMclass10.pdf and here: https://www.tnu.ethz.ch/fileadmin/user_upload/teaching/SPM2016/Ridgway2016_SPMZurich_VBM.pdf)
And to answer your question, for this analysis the FWER p<0.05 T-score threshold for the spmT image is 4.54.
I'm doing a run of randomise now with the -x option and will see what the results look like tomorrow. I did explore the user guide further to find that the recommended number of permutations is 10000 whereas we were using 500 based on the example at the top of the page. We can see how too few permutations might lead to inaccurate p-values but wouldn't expect it to be particularly biased in the direction of over-boosting, right?