Dear Mike,

Well, there's no ultimate solution as in most instances, one can think of several reasonable approaches with specific pros and cons, and usually it's much easier to critize a model than to come up with a solution. Concerning your approch, in my opinion the main drawbacks are that
a) you rely on voxel selections for further analyses, which are then based on the same data set though. In your constellation this seems to introduce bias, as described in the Kriegeskorte paper. It doesn't mean the findings are nonsense, but they would be biased to some extent, which one could criticize.
b) inconsistent use of small volume correction. If your hypothesis is about MCC, why don't you use SVC for (1) and (2) as well, as this might result in more sig. voxels (in case SVC is conducted on voxel level) or additional sig. clusters (in case of cluster level). Thus, (1) and (2) seem to serve as a localizer, which brings us back to a). 
c) As you look at (1), (2), <Result 3-1>, <Result 3-2>, thus sort of a stepwise regression (if you really want to conduct all these analyses, possibly they are just different alternatives), it would be interesting to receive some measures like (adjusted) R^2 for each of the models. Alternatively, you might start with (3) right from the beginning.
d) Analyses (5) and (6) would be double-dipping. With (1) - (3) or  the combination (1), (2), (4) you detect some significant clusters/voxels on a voxel-by-voxel basis, then you extract some composite score and conduct the same analyses (which is circular, as the voxel selection is based on those that have already shown up sig.). If you go with (1), (2) followed by (5), or (3) followed by (6), then this is not circular but biased due to a) (and also see b), instead of SVC vs. whole-brain it would be ROI vs. whole-brain).
e) if you wanted to report effect sizes they would often be based on clusters that have been detected to be sig. in a previous step, resulting in bias.

To me, any of the combinations would be imperfect. Given that you have a-priori hypotheses about the MCC (and thus probably, given some citeable literature) it's also too much about clusters showing up sig. or not. I mean, what would you have done if (1) or <Result 3-1> had not resulted in a sig. cluster? Thus I would turn to a proper ROI analysis right from the beginning. Define your ROIs e.g. based on anatomy. There are different parcellation schemes for the cingulate cortex and various options,
a) the Hammersmith n30r83 brain atlas might be worth a try (the ACC seems to include the MCC), it provides a label for ACC, subcallosal area, subgenual frontal cortex (please check yourself whether it could be regarded as CC instead of FC)
b) if these atlases are not precise enough for your a-priori regions you can adjust labels yourself. Brent Vogt is an expert on the CC, he and his colleagues have released a series of papers on subregions. You could try to start with an appropriate CC label e.g. from the n30r83 and then edit the labels manually based on y and z coordinates to get reasonable boundaries for subsections (which should work quite well for CC). One example is Yu et al. (2011, Neuroimage, https://dx.doi.org/10.1016/j.neuroimage.2010.11.018 ), who relied on Vogt and a "four-region model with 7 specified subregions". Of course there are also papers and parcellations by other authors, e.g.  McCormick et al. (2006, Neuroimage, https://dx.doi.org/10.1016/j.neuroimage.2006.04.227 ) 
c) you could go with coordinates reported in papers and build your own functionally defined ROIs
d) turn to connectivity-based parcellation, e.g. that by Beckmann et al. (2009, J Neursci, https://dx.doi.org/10.1523/jneurosci.3328-08.2009 ) looks like many subregions. I'm not sure whether you can obtain exactly these labels, but there are certainly several freeely available resting-state parcellations with many subregions, likely also for the CC
e) if this is not possible and/or if you don't want to rely on previously published parcellation schemes, as they might not reflect the neural processes investigated in your study, you can divide a large CC label into arbitrary small subregions along e.g. an anterior-posterior gradient like in Torta & Cauda (2011, Neuroimage, https://dx.doi.org/10.1016/j.neuroimage.2011.03.066 ).

Based on a parcellation scheme as proposed in a) - e) you obtain independently defined ROIs (of course you should not try to optimize them to catch the already detected clusters), which should also be interpretable in a very meaningful way (e.g. reflecting a certain anatomical scheme). For each of the ROIs, you can extract a composite score, which should increase sensitivity due to reduced noise, and you can rely on a more liberal threshold as multiple testing is much less of an issue, which should again increase sensitivity. You can report average beta estimates and statistics and present scatter plots for these ROIs without having to state that this is just for illustrative purpose due to selection bias via sig. voxels, you can easily contrast subregion A with subregion B, conduct various types of regression without having to stick with the limitations in SPM and so on and so forth. Overall, it would be much more elegant. And of course, you can still report a GLM as in (3) to look at whole-brain effects to test whether non-CC regions show certain patterns as well.

Best

Helmut





 , unneccesarily too