Well, generally it is possible to produce a PDF where each page is one big vector (or even one big raster image for that matter). It would visually look the same as the one created from separate objects and text, but from my understanding with this approach it would be tricky or impossible to create a properly tagged accessible PDF.

That aside, using Inkscape for parts of the doc looked like an interesting idea. I checked and unfortunately the API seems limited. There is a Python plugin system but I don’t think it is possible to create a solution that works entirely headless on CI boxes without having to invoke Inkscape GUI.