Good question. This issue had already crossed my mind, but somehow i never cared to find out exactly.
Right now, any glyph marked as 'junk' will be tagged as 'junk', except for "take with next" glyphs (as it seems). This is not ideal.
The best solution here would be to manually decide for every single glyph/part if it's junk-or-not, to have full control over all glyphs, but...
If you mark only the first part as 'junk' (which is not possible right now) and then add the next part as regular (non-junk) glyph, the matrix would contain only the complete glyph, but not the left part of it. If you save/re-load the matrix, that first part would be gone and if detected again, the OCR engine would pop up and ask, you would answer 'junk' and 'extend-right', and then the next glyph (the complete one with parts 1+2) would pop up again, you'd enter it again and then it would be added to the matrix a second time (as a 'double', i.e., completely identical glyph), which is also not ideal (the OCR engine does not OCR the next part after a click on "extend-right").
Jeez, and i thought this latest bugfix release would be the last one for a while..
Problem is, this "take with next" thing is another pretty complicated part of the OCR code. Frankly, i don't see any quick solution for this. I'm taking a note, but i'm not sure if i can come up with an easy fix for this. I also have no plans for any new releases in the very near future.
So right now, my advice would be to mark the right-most part as 'junk' and then manually delete the "take with next" glyphs afterwards.