rahulrao9/Agent-Arena-Neural-MMO — reverse-engineered prompt

Build me a clean CPU only training project for Neural MMO 2.0 that can run on a headless Linux server, even if the machine has old or broken CUDA drivers.

I want it to train a basic reinforcement learning agent using PufferLib, save the trained PyTorch policy weights, and generate a replay file that I can download and open in the Neural MMO web viewer. It should also write logs that I can monitor with TensorBoard while training is running.

Please make the setup reliable for a conda environment with Python 3.10 and avoid anything that tries to compile or use CUDA. If PufferLib tries to detect nvcc and build GPU code, patch or work around that so the install still works on CPU only.

Also include the fixes needed for Neural MMO replay generation and multi head action sampling, so it doesn’t crash when saving replays. Add clear run instructions for setting up the environment, starting training, stopping safely, viewing logs, and downloading the replay.