What you see here is a summary of thinking tokens written by some other smaller model (e.g. old sonnet). The actual thinking sometimes (rarely) leaks and is not easy to parse.

There are various justifications on this, but it's mostly to make distillation and fine tuning off their model outputs a bit harder for their competitors

Well, what's the difference? If it's pretending to think and its thoughts correlate to its final output, then I'd say that really is thinking.

In some cases, an LLM may truly "consider the architecture" internally, within its latent representations, and in others, it can output a similar phrase simply because it's "expected" of it.

"Where" is pretty clear. There aren't that many places within an LLM, and hidden state is the main culprit. How to read that space is another matter entirely.