(I apologize for the crooked translation of specific explanations, it is not my native language).
That is, it is clear that in fact any merge must be marked as a merge and checked whether the license allows it, but in essence for me as a user it is much more important how the merge did or did not affect the final checkpoint, and it is not always obvious. Here are a few examples.
Let's take 32 checkpoints and each one has a 0.5 merge with the other. It turns out that the first checkpoint is present in the last one at 0.015625.
Let's take the Supermerger extension and merge only the first and last five layers. That is, the merged model will only affect the random noise distribution, not the visible picture.
Let's take the Model converter extension and open additional settings, leave convert only for text encoder, while the rest should be deleted, and merge obtained couple hundred megabytes with another checkpoint. In other words, we expanded the conceptual base of the checkpoint without affecting what is responsible for the visible picture.
The point is that I encounter each of these three cases one way or another in my various checkpoints, sometimes even all at once, and I wonder whether I should mark the checkpoints as merged and whether I should check the licenses of the other checkpoints.