"given that AI can now supposedly engineer a clean-room reimplementation of any published program whatsoever"

I'm missing something there, that's precisely what I'm arguing again. How can it do a clean-room reimplementation when the open source code is most likely in the training data? That only works if you would train on everything BUT the implementation you want. It's definitely feasible but wouldn't that be prohibitively expensive for most, if not all, projects?