I was testing this by creating a new file. I drag n dropped from a small file twice into the new file. Then I used AutoMerge to merge everyone. I found that two sources did not merge. When I did automerge on sources they still didn't merge. They are attached on the person and fact levels so that's not the issue. I can't figure out why these sources are not merging. They appear exactly the same in every way. This has been added to our tracking system for development to look at further.
The following is a wild theory, and I'm even going to shoot down my own theory.
I've never tested this, but what if what I will call "notes in a source" - Master Source Text, Detail Text, and that sort of thing - have the same bug as general notes, family notes, and event notes where a trailing carriage return line feed is dropped in a GEDCOM export. Then the source or citation would not look the same after being dragged and dropped as it did before being dragged and dropped.
Here is where I shoot down my theory. If you drag and drop the same data twice and if the same carriage return and line feed in a "note in a source" is dropped both times, then then neither dragged and dropped source will match the original but they should match each other and therefore they should merge.
But what if you created the new database first by copying the old database to the new database (thereby preserving everything) and second by dragging and dropping from the old database to the new database. In this case (and if "notes in a source" have the same GEDCOM bug as regular notes), then it's possible the otherwise identical sources could differ only by the presence or absence of a trailing carriage return linefeed in a "note in a source". And if they differ, then they won't merge.
Still a shot in the dark!
Edited: oops, I didn't reread the whole thread carefully enough. Tom already identified a problem where any non-blank "notes in a source" prevent merging, even if the non-blank "notes in a source" are equal. That very much supercedes my wild theory.