As far as I see, the first chunk is scaled (S) but not referenced (R), so its external orientation cannot be considered fixed even if you have set this chunk as anchor. So there is no reason for photoscan to keep chunk's current external matrix for the model.
But apparently the align chunks operation requires the anchor chunk to be referenced (R), and it seems that the temporary orientation that you have set manually is dismissed and the original omega phi kappa of the anchor chunk become fixed as reference. I would agree with you that it would be more desirable that happened what you expect, but maybe someone has a diferent opinion.
Alexey might want to put some light here.
Would't be nice to have the option to turn scaled (S) model into referenced (R) accepting the eventually weird coordinates it yielded to?