Is tree-sitter really slower than TextMate grammars? Some benchmarks indicate that this isn't really the case [1]. On the other hand, breaking parse trees is a real issue, because the error-recovery in tree-sitter is pretty rudimentary [2][3], but as you said, it's not an issue for Shiki.
Several TextMate grammars suffer from inaccuracy bugs, and issues of maintainability. Perhaps the biggest hindrance in the adoption of tree-sitter, is that the most popular editor, VSCode, still doesn't support it.