Looks good! 👍
One observation: the indexed-tuple approach optimizes for compactness and performance, at the cost of relying on shape-sniffing (isinstance, len, nested type checks) to distinguish formats in __setstate__. That's perfectly reasonable for this and Version use cases, just noting that there's no explicit format tag or version marker, so future format changes will need another heuristic branch rather than a clean dispatch. Not a current risk, and if it does become a problem in the future, a format tag can then be introduced at that point...

Yes, I usually start the tuple with an "format version" integer. But I don't think this will be likely to change much, and simply adding a value is safe. Longer term, we could probably use pattern matching here if it's not too much slower.

Indeed, sounds good

The copilot review boils down to one thing: protect against malformed pickles. I've added it (generated tests, but threw away the model-implemented solution and added my own), but I'm not sure it's really needed to always check this (auto-generated pickle code doesn't have such checks), but I think it's probably a good idea for the "old" pickles. I've added it for everything, but I'd also be fine to drop it for new-style pickles (outer tuple). It's not too expensive, though, and it does make typing happy.

Pattern matching would be so nice here. :)

Also, I've followed #1163 in generating old pickles and storing them, but we could also just store the intermediate state, which wouldn't be a somewhat opaque binary string, and would unit test the same. I think the current one is a bit better, but if there's a desire to avoid binary strings, we could make something work.

Cheers, thank you! 🍻