I’m not sure why more programmers don’t use LaTeX — it seems like a natural fit for that segment of the population, looks a lot nicer when typeset, and has the aforementioned benefit of being CVS/SVNable, not to mention segmented across multiple files, just like regular source code. Perhaps the word processor is such a ubiquitous cultural element that by the time most people realize there are alternatives, they’re unwilling to put the effort into switching.

Storing things in a sparse, structural markup like LaTeX doesn’t completely solve the problem you pose regarding semantically insignificant edits, but barring some sort of semantic meta-markup to be superimposed on top of everything we do, the problem seems rather intractable. I look forward to someone much smarter than myself proving me wrong.

So, for your table example, the source files must first be parsed into a DOM tree and then re-serialized with a standard indentation and attribute order before diff will work, but after that it works very well at highlighting the semantic change, since the normalization ensures that the meaningful unit of change (a line) is tokenized identically between the two versions.

SSDDiff — a diff for semistructured data
ssddiff.alioth.debian.org

The author of the above tool also wrote this interesting thesis: “Stucture-Preserving Difference Search in Semistructured Data”. It has comparisons against Logilab xmldiff (written in Python!) which shows how his algorithm improves the result.