Mass rename so instance fields all have underscore
The codebase is split half way between two schemes, and not having them clash with locals is more convenient
The benefit over converting just the selected characters is that we're converting what's hopefully a well formed document, so the rest of the document can be used for context, and returning something that represents a node, rather than trying to convert the text "tring", when someone slightly mis-selects.
If two layers of the node hierarachy fail to go through the trivia converter, then the whole document will be converted, which seems like an OK failure case.
We'd probably get a better conversion out of converting a whole file then picking a subset, but since I don't actually understand the use case, the effort isn't really worth it here.