The gameplay is always the most difficult part of any game to RE. For most of the code, you can see on the screen what it's doing, or otherwise know what it is supposed to do. But for game logic, you don't know the algorithm that was implemented - it may not even be anything sensible - so there's no point of reference. I've had a few projects where there were still parts of the original code that I couldn't quite understand, but it didn't prevent a transcode because all that matters is that the same logic is implemented.
I'm not surprised that AI is starting to get to the point where it can RE software. Recently a core was released on MiSTer that was written by AI, and I know from experience that it's no trivial process. I suspect a transcode by AI would be fairly straightforward... I'd like to try one on a simple program one day. RE would be more difficult but I can see how even now it would do quite a bit of it.
I just had to educate myself on Godot. Looks pretty cool. I also had to Google Torvak; despite having an Amiga back in the day, and hundreds of games, I'm not familiar with this one. Interestingly my 11yo son has to develop and showcase a 'game' next week at school as a special project, and I have no idea what to advise him to do. Maybe Godot is the answer...
Yes Godot is a good start. At beginning I would not like to use any SDK and made 2 games using javascript, but collision between elements was a nightmare. You have a lot of tuto for Godot on YT, collision are built-in using mask index, tiles can be set as solid just by a click, and the most interesting are the nodes: graphics elements inheritance using relative or globals positions. Then you code in python-like (or C++ if you prefer) You can compile in a lot of devices, including smartphones. I wish I can compile/export for Amiga or Neo-geo but for now, I'm using a Godot script to export animation for making a game using ngdevkit.