1) I understood that I can increase the significance of cluster based thresholding with the -c option, for t contrasts. This should break apart the large clusters. Can I apply this -c or -C option on the TFCE results?

I'm afraid this won't work. TFCE doesn't use either of these thresholds, i.e., it is "threshold free".

If yes should -c applied on the tbss_tfce_corrp_tstat2 so on the corrected results thresholded at 0.95 or on the uncorrected results tbss_tstat2?

Is this correct:

randomise -i all_FA_skeletonised -o tbss -m tbss_tstat2_masked_th95 -d design.mat -t design.con -n 5000 -c -V ?

Hmm, this isn't correct because there is an argument missing for the -c option (i.e. the cluster-forming threshold). But even if the threshold were there, it wouldn't behave in the way you want with respect to TFCE. The -c option will execute a separate inference at the cluster level and report the results. They are different things.

2)What is the best value to be used for -C? Is 4 OK?

Sadly there is no rule for what the best cluster forming threshold would be. I don't believe, however, that a reviewer would complain that 4 would be too low.
Also, note that -c (lower case) isn't the same as -C (upper case). One is for inference on cluster extent (size), the other on cluster mass (sum of the statistics within the cluster).

3) You mentiond that in case of large clusters I could just report corrected *voxelwise* p-values . Are this values the one from tbss_tfce_corrp_tstat2 or from randomise with option -x or -c?

Voxelwise results are produced with the option -x. Type 'randomise' without arguments to see a list of options.

4) How I can get the histogram of the raw t-stat image?

In FSLview, menu Tools -> Image histogram.

Hope this helps!

All the best,

Anderson