There might be a turnkey solution for making inline canvases as a Jupyter extension. Good luck identifying logical objects, but there must be some advantages over drag-and-drop. I'd use this for my current prototype but that my target domain's existing DSL is not (yet) graphical. For any other paper tool, diagram, nonlinear algebraic notation, what have you, there is a lot of potential.

brainstorming: • a Lisper might imagine that a "canvas" thingie can contain: (1) pixels, (2) canvas thingies • some shape recognition is certainly possible, my Mac can recognize some hand-drawn shapes, my neurons want me to say Ink-and-Switch