Thanks for that question. The relation between collections of TEI-encoded text and the uses of that encoding for computational stylistics seems to deserve more attention. There was a panel about this at the DH Conference in Hamburg in 2012. A misconception, I think, is the idea that all text analysis people ever do with TEI markup is get rid of it.
I've used TEI marked-up plays for computational stylistics and, although I have been getting rid of the markup before running the actual analysis, I still used it and would not have wanted to do without it. For instance, in a collection of around 600 French plays, I was able to use a very simple XSLT to select just the character speeches (without the prefaces, but also without the stage directions and speaker names) and then analyse my plays based on this selection of each play. Thanks to TEI (or rather, thanks to the fine people encoding texts in TEI), this is a matter of minutes. A bit of this research has been published here:http://journalofdigitalhumanities.org/2-3/big-smart-clean-messy-data-in-the-humanities/
Similarly, I have used crime fiction novels encoded in TEI with chapters marked up, for topic modeling. In this case, I would have been able to split the novels into chapter chunks very easily, with a little bit of help from XSLT-savvy people around me. I ended up splitting the texts into even smaller, arbitrary chunks, but still I think this is a realistic use case. A first go at such research is documented here: http://dragonfly.hypotheses.org/530
These are very simple examples and others may have much nicer ones, but I thought I'll share them anyway. There is a nice tool for text analysis that imports TEI text and saves structural information from it (such as div, p, and much more) which you can later use when making queries over the texts. The tool is called "TXM" and is probably well-known on this list anyway. More information about it can be found here: http://textometrie.ens-lyon.fr/?lang=en