Also, the text styling has a lot more complexity than you would expect.
As I recall, regions were essentially something akin to an array of run-length-encodings for each scan line (where, additionally, there was a sentinel or some way to indicate the next n rows are the same as the previous). The fun part is then writing a set of routines to union, difference, etc. pairs of these RLE region objects (and here I use the word 'object' loosely).
I’ve gone on a lot of rabbit holes digging into original source code and dev manuals for this project, hopefully one day I’ll write some of it up somewhere. I think what’s been interesting is iterating towards the most elegant solutions has often meant ending up with pretty similar approaches to what Atkinson originally used.