I turns out I picked up on the wrong origin for the HTML in this package. It's not from `pkgdown`; that is just used to generate their upstream website and isn't shipped in the CRAN tarball. Mea culpa.
The HTML in this case (`inst/doc/*.html`) appears in fact to be generated by RMarkdown when upstream built this package, from the corresponding files in `vignettes/*.Rmd`. Two of these documents (`extending-ggplot2` and `ggplot2-specs`) contain only a short code highlighting script (inserted by Pandoc) and a generated link to Mathjax. These links are likely to be common to any RMarkdown documentation prebuilt by upstream. They could be fixed by regex, but these HTML files are in any case not source and should be removed (and ideally rebuilt). `profiling` contains a huge pile of minified javascript, which appears to be generated by the `profvis` library. This generates an HTML widget in the RMarkdown output including all the necessary inlined libraries. Rebuilding this vignette would be hard, but I think is probably of pretty marginal interest. `dh-r` already supports rebuilding vignettes (`dh $@ --buildsystem R --with vignette`), but this functionality has never been much tested or used, I think. It means adding a bunch of extra build dependencies (either pandoc or texlive, depending on the output format, in addition to any extra packages needed for the vignette). I also doesn't currently have any mechanism to only build a subset of vignettes if some are known to break, require exotic dependencies of take too long to run. I'm a bit concerned that trying to enable this more broadly is a bit of a rabbit hole which risks making a lot of packages slower and more flaky to build.
