> They're not under any obligation to include their own code in the training data. Why would they?

Because these models work better with more data and presumably this a lot of high quality data that they already have lying around anyway? Because there no downside according to their own reasoning? Because it would shut up a lot of these criticisms right away? Because marketing would be so much easier with that kind of dogfooding?

In short: because according to their own story there would be only upsides, no downsides.