Does this allow for semantic, non-text diffing? This would allow for code formats without syntax errors (i.e. storing the valid AST in binary as separate, likely structured in binary, from the editor representation).

Diffs would then be truly semantic, representing semantic refactorings (e.g. "extract method to function", "rename symbol") at the patch level. you could easily then query "when did this variable show up" even post-refactor where the variable moves modules. Sure this has its limits but it'd push the tooling to levels virtually impossible to automate with text-based code now.

A man can dream!