A highly colored image became a small vector program.

Image decomposition became a compiler: structure discovered in the raster is translated into paths, colors, and local models rather than discarded before export.

The raster became a program before it became an SVG.

This did not begin as an SVG maker. It began by asking for a smaller structured explanation of an image. Once regions had owners, topology, and local models, the image could be compiled into a vector program and simplified according to the damage each saved byte would cause.

Everyone assumed
Vectorization is chiefly a matter of tracing boundaries and simplifying the resulting paths.
I asked
What if an image must first acquire owners, hierarchy, texture, and an explicit account of which errors matter?
I changed
Build a structural image model, preserve topology, and spend a reconstruction-error budget only where merging actually buys a smaller program.
Now there is
A local raster-to-SVG instrument that emits editable, deterministic vector output under explicit error control.