I think this is being over thought a bit. Even in it's most primitive form, it could still be quite useful in practice.
It's pretty easy to construct cases where it fails, but all such systems will fail either due to changes in content or changes in structure, and many will fail due to ambiguity as well. It's only really useful to compare how this might fare against other approaches. An example would be producing an XPath expression. Which may be more robust to ambiguity but weaker to structural changes.
This system will fail if the content is different enough that the target text is gone, which I believe is very reasonable. And while you could construct ambiguous text that couldn't be disambiguated by picking a larger string, in practice the value added with the complexity of trying to solve that problem is probably not very good. For the purpose of linking to sentences and paragraphs, this system seems sufficient and probably even better than many alternatives. I don't think many articles have a lot of repeating sentences.
Other use cases than linking to actual text would probably have their issues. Like, trying to link to a cell of a table would be quite problematic. That case may be better handled with something like an XPath expression, but in any case it's a fairly different use case imo.
All in all, my initial opinion is that this design is useful and easy to understand. It gives you the equivalent of Ctrl+F which is not bad.
(Disclosure: I work at Google, but this is the first I've heard of this and I don't work on Chrome.)