It could be that the source format played a big role in my case... it was MadCap Flare XML/HTML, so the segmentation is pretty much defined by the file type, rather than the TM-defined rules.
Of course I cannot be expected to know how Studio works internally... all I know is that I:
- created new empty TM where I changed the segmentation to Paragraph based for both source- and target language
- used this TM for running the alignment
That's all. I don't (and can't) know exactly which "magic" (or coincidence) made it to align just as one would expect ;-). Perhaps is the internal "reversed" TM created by reversing the actual TM (similarly to what AnyTM does)? It would quite make sense...
I didn't explore it any deeper as we ended up not going further with paragraph-based segmentation and went the harder way of sentence-based segmentation.