I don't mind downgrading the plain text export to plain ASCII if someone
would like to provide a patch to configure Sphinx to do that, and also a
diff to prove that it doesn't spoil any examples anywhere.
Spoilage risk is implicitly ~none, since the majority of the quoting/etc used is already valid ASCII, with the errors added later as inconsistencies to that - i.e. we know the system can and will produce correct output, if provided with a suitable source - but the problem is really that we obviously can't rely on it remaining as such, because whatever tool is now being used for at least some edits to that master has changed. That being the case, the only practical solution is to fix the export itself, as you say, or it'll just regress again in the future.
Yet Another Sphinx made things very confusing for a few minutes, but assuming I have a correct view of the process now this clearly needs far more involvement than I have. Crucially, it also appears to be pulling at least data and potentially even code (in an HLL sense) from a non-Debian site, which makes the splash radius of any such fix unacceptably large, especially for an outsider. I could provide a patch for libxslt easily enough, but it would never be accepted upstream; and this isn't worth forking over because of the maintenance burden.
So, would have been nice, but isn't happening. Many thanks for the quick response: I'd have ended up wasting the weekend if you'd left it til Monday, so that's very much appreciated.