Agreed, having text as a reasoning layer impedes performance while giving a false sense of security.

It could be encoding anything in its adjective choices etc. The same thing that model providers are using for watermarking can used by the model itself for private thought.