A way to work with this, but avoid the potential bias as to which region is seeded from, would effectively be to do this both ways, and combine the results. Although in your example, this would still not get you what you were looking for: even seeding from a region doesn’t guarantee it to only track in one direction; for that you’d need the unidirectional option and probably also indicate a starting direction. But even then, streamlines would still be present in your seeding regions, while being not present in your inclusion/stopping region, as they stop upon hitting the latter one. So it’s again not exactly “between” 2 ROIs… depending on how you read/interpret the word “between” there… 
A workaround of sorts at the tckedit stage may be to use the -mask option, and essentially mask out the bits that are beyond the regions… but depending on how messy and extensive they are, this may still be a very manual approach. Of course, if the actual bundle itself is relatively limited in size, you could draw a mask that simply includes the bundle, but doesn’t include the bits you want to get rid of. You could even potentially go with a track count map (tckmap) and use a clever threshold to get an initial mask, and then edit that a bit further to your liking to get an appropriate mask for tckedit. This may work decently if your bundle itself is quite compact, but the streamlines start dispersing a lot the moment they’ve hit your “end” regions, due to no further constraints applying beyond that point. The track count would inherently drop quite a bit at those points. Actually, now thinking of that, you could go with the example that I gave in the post you linked to, where I used the -stop option to tckgen in an “unsuccessful” way, but I did end up with a strong track count drop beyond my intended “endpoint” regions…
@jdtournier, @rsmith: any other bright ideas how to succeed at the use case @tbilliet described? It’s actually one that I hear quite often from users… if it can’t be done in a straight forward manner with the current functionality, this may be a candidate for improvement of either tckgen and/or tckedit…