Hi David,
I have not spent much time on this aspect of the text yet, so I am quite uncertain about the meaning of the above observations. My provisional opinion is that the phenomenon discussed here is well distinct from labels, which possibly is in agreement with your statement that "not all images and text overlaps can be handled the same".
An obvious hypothesis is that what happens in readable manuscripts also happens here: an image break can be coincident with a normal space (I believe this could be the case in You are not allowed to view links. Register or Login to view.) or it can split a word into two parts (this could explain the frequent stand-alone s). A (minor?) problem with the second option is that it does not seem to be confirmed by the average length of words after the image-break: they appear to be whole words.
A high rate of -m and -g before the break would have supported the idea that these are abbreviations (a pet theory of mine), but I could not observe any such behaviour. In principle, it could also be that words before the image break are sometimes truncated.
I think that some of this speculation could be supported/ rejected by further analysis.
I have not spent much time on this aspect of the text yet, so I am quite uncertain about the meaning of the above observations. My provisional opinion is that the phenomenon discussed here is well distinct from labels, which possibly is in agreement with your statement that "not all images and text overlaps can be handled the same".
An obvious hypothesis is that what happens in readable manuscripts also happens here: an image break can be coincident with a normal space (I believe this could be the case in You are not allowed to view links. Register or Login to view.) or it can split a word into two parts (this could explain the frequent stand-alone s). A (minor?) problem with the second option is that it does not seem to be confirmed by the average length of words after the image-break: they appear to be whole words.
A high rate of -m and -g before the break would have supported the idea that these are abbreviations (a pet theory of mine), but I could not observe any such behaviour. In principle, it could also be that words before the image break are sometimes truncated.
I think that some of this speculation could be supported/ rejected by further analysis.