The underlying shapes are defined by geometries (so vectors) and then rasterized to canvas. So the SVG export could happen before the rasterization.

There are other libs that are native SVG though, so why even use Konva? The benefit of Canvas is that it can be a lot, LOT faster than having multiple complex SVGs. SVG is especially slow on Firefox IIRC.

I make a canvas-based diagramming library (GoJS) and implemented SVG export, and then a whole SVG renderer if you want the same scene in real-time. Generally if you're using HTML Canvas commands, you can engineer a perfect SVG equivalent. There are some really really really annoying edge cases with transforms, clipping, gradients, etc. But nothing is impossible :D

We used Kinetic a lot for a specific type of visual analysis we had, but when it came to creating publish-worthy figures, working with PNG screenshots was less than desirable.

Would it be so difficult to crawl over a snapshot of the object hierarchy and render to a cairo-like library?

At one point I started implementing that myself using my own object representation that I had running parallel to Kinetic's internal one, but never finished it.