Your definition of "reasoning" is doing a lot of the heavy lifting. No one is claiming this reasoning is analogous to human reasoning. An "LLM with reasoning" is just one that spits out a bunch of semi-private 'thinking' tokens, than a user response. No one is trying to claim "it reasons and understands like a human". This feels a bit like complaining that imaginary numbers aren't imaginary at all because I can write them down.