I think the problem can be defined equally as: we can't invest in something more abstract than "plain text" at this time. When we try, it gets downgraded to a plain text projection of the syntax.

The plain text encoding itself exists in a process of incremental, path-dependent development from Morse Code signals to Unicode resulting in a "Gigantic Lookup Table" (GLUT, my coining) approach to symbolic comprehension. The assumption is useful - lots of features can "just work" by knowing that a particular bit pattern is always a particular symbol.

If we push up the abstraction level, we get a different set of symbols that are better suited to the app, but not equivalent GLUT tooling. Instead we usually get parsing of plain text as a transport. For example, CSV parsing. It is sloppy; it is also good enough.

Edit: XML is also a key example. It goes out of its way to respect the text transport approach. There are dedicated XML editors. But people want to edit it as plain text and they can't quite get there because funny-business with character encodings gets in the way, adding a bunch of ampersands and semicolons onto the symbols they want to edit. Thus we have ended up with "the CSV of hypertext documents", Markdown.