Another approach would be for the tool to accept doing structured merging even if there are error nodes in the parsed tree. If those error span the parts of the file where the extended language is used, then the tool could still help with merging the other parts, treating the errors as atomic blocks. I'd be a bit reluctant to do that, because there could be errors for all sorts of other reasons.

- for diffing, the matching of the leaves is what matters the most, for merging the internal nodes are more important,

- for diffing, it feels more acceptable to restrict the matching to be monotonous on the leaves since it's difficult to visually represent moves if you can detect them. For merging, supporting moves is more interesting as it lets you replay changes on the moved element,

- diffing needs to be faster than merging, so the accuracy/speed tradeoffs can be different.

Packaging parsers into separate executables seems like hard work indeed! I assume you also considered fixing the tree-sitter grammars (vendoring them as needed, if the fixes can't be upstreamed)? Tree-sitter parsers are being used for a lot more than syntax highlighting these days (for instance GitHub's "Symbols" panel) so I would imagine maintainers should be open to making grammars more faithful to the official specs. I'm not particularly looking forward to maintaining dozens of forked grammars but it still feels a lot easier than writing parsers in different languages. I guess you have different distribution constraints also.

As to mismatches: yes, they are bound to happen in some cases. Even for line-based diffing, Git uses rather convoluted heuristics to avoid them (with the "histogram" diff algorithm), but they can't be completely ruled out there either. I hope that with enough safeguards (helper to review merges, downstream consistency checks with local fall-back to line-based diffing) they can be lived with. I'm happy to try other matching algorithms if they are more promising though (there isn't much coupling with the rest of the pipeline).

Concerning tree-sitter, I have noticed some small issues, but nothing that was a show-stopper so far. I actually like it that it's designed for syntax highlighting, because it's really helpful that the representations it gives stay faithful to the original source, to avoid introducing reformatting noise in the merging process. Parsers written for a specific language can sometimes be too zealous (stripping comments out, doing some normalizations behind your back). That's a problem in Spork (which uses Spoon, a pretty advanced Java parser). And the uniform API tree-sitter offers over all those parsers is just too good to give up, in my opinion.