Triangulate

The canvas starts as the mean colour of the target. For each new triangle the program generates thousands of random candidates, sets each one’s colour to the mean of the target underneath it, and keeps whichever lowers the squared error most. Coverage comes from the edge function, so either winding order works without a special case.

Recomputing the loss over the whole image is far too slow to try six thousand candidates per triangle, so scoring touches only the bounding box and accumulates a difference instead. That is fast, but it can drift. The program recomputes the error from scratch at the end and prints it beside the running total, and if the two disagree the incremental scoring is wrong somewhere.

Placement is still uniformly random, when it should start where the residual is large. Colour is fitted as a plain mean, but with the canvas and the opacity held fixed the best colour has a closed form, and so does the best opacity for a fixed colour. Alternating those exact updates should beat random search on both quality and speed.