Since you are a wrapper on top of OpenCascade, I assume you have the ability to export STEP solids, and so one potential test you could run is have a series of standard objects that are exported and then re-imported with the basic opencascade tools and then measure the volume of the solid, verifying a solid volume to some small epsilon. This type of volume/mass verification is used in a lot of CAD modeling competitions to determine if a part has been correctly modeled; if the part is of sufficient complexity (especially things involving fillets), the odds that the solid is computed with the correct volume, down to several decimal places, while something else in the modeling is wrong, are astronomically small.
Nice to see other open source CAD development being explored. BRL-CAD has had the pixdiff, bwdiff, and pixcmp tools for a very long time, for very similar purposes:
however, if you have SVGs already, compare those without rasterizing them first, then if that fails rasterize & odiff the baseline and new on-demand. this way you dont waste a ton of space storing a bunch of binary [possibly evolving] pngs in git.
When it comes to LLM's doing adversarial training, I wonder if there is any value to letting them rapidly design and simulate circuits or mechanical models.
pixdiff: https://manpage.me/index.cgi?q=pixdiff&sektion=1&apropos=0&m...
bwdiff: https://manpage.me/index.cgi?q=bwdiff&sektion=1&apropos=0&ma...
pixcmp: https://manpage.me/index.cgi?apropos=0&q=pixcmp&sektion=1&ma...
however, if you have SVGs already, compare those without rasterizing them first, then if that fails rasterize & odiff the baseline and new on-demand. this way you dont waste a ton of space storing a bunch of binary [possibly evolving] pngs in git.