Pomilon-Intelligence-Lab/CRSM ? reverse-engineered prompt

Build me a Python research prototype for an experimental reasoning model that can think in the background while it generates output.

I want a small Mamba based model with an asynchronous planner loop, where the planner uses MCTS to explore better future states and then gently updates the model’s hidden state instead of writing out chain of thought. Please make it easy to run a simple demo, train on toy tasks, and benchmark the reasoning behavior on ARC style grid problems. I also want clear tests for the state mixing and stability parts, plus a simple config driven setup so I can tweak hidden size, number of layers, and injection strength without touching much code.

Please keep it easy to inspect and debug, since this is still a prototype and I need to see what is happening step by step. If you need current docs for Mamba or PyTorch APIs, look them up online first.