Wow, that makes so much sense - thank you! It sounds like it essentially "vibed up" the parts it didn't understand, which is amazing.
Just as an experiment, I fed the resource fork to GPT-5.2 to see whether it could render the windows/dialogs in the resource fork - it did a fairly okay job. I think the fundamental limit it ran against (and acknowledges) is that lot of Mac's classic look and feel was defined programmatically, literally, down to calls to RoundRect(...).
https://chatgpt.com/s/t_694dddb290308191babcb07a72367e97
Thanks for posting your experience.
When a filter is implemented half in Pascal (setup and loop over the rows) and half in assembly (each row) Claude did it all in Go, but the structure in Go is the same: one entry point for setup and iterating on rows, and one function (ported from the assembly) to process each row.
(As for the resource fork, it just reimplemented the UI in HTML. There's not enough info in the transcript of it's thinking to know if it read the resource file and understood it, or if it used a general understanding of what was in Photoshop, from training data, to do it.)
My mind is blown. I keep trying to find evidence that it just copied this from someplace, but I can't see how.