That doesn't seem like a very sharp line (hah). "Abstract geometrical or statistical features like edges" have semantic meaning. An "edge" is just an area of the image with a high derivative, but as soon as you use that information to process it in some way, you are making some implicit assumption about what this statistical feature means - the boundary of an object, a discontinuity in depth, etc.

There is no dividing line between syntax and semantics - semantics is just syntax scaled. Godel proved it, LLMs exploit it.

If I run an unsharp mask filter I'm not making any decision about what individual edges mean. I might look at the result and decide to apply more or less sharpening, but that's still not "semantic editing", because the edit itself only sees derivatives.