Maybe I am missing what they meant here, but isn’t the entire point of tree-sitter that you can reuse old trees to make edit updates faster? It is seems quite probable that all the performance gains came from optimising for fixed files with no error recovery, at the cost of how ts is actually used.
I've used tree-sitter for a few projects, but I never touched the tree-reuse stuff. The value of tree-sitter came from the fact that grammars compile to a .so and can be easily distributed through package managers. Adding support for a new grammar in ast-grep might just mean pointing to the .so at runtime (idk, I haven't checked). So this optimization might make sense for this project.