danijar/dreamerv3 ? reverse-engineered prompt

Build me a Python project that recreates DreamerV3, a reinforcement learning system that learns a world model from experience and then trains a policy by imagining future trajectories.

I want it to run from a simple command line entry point, support different task configs like Crafter and Atari, and let me change settings with command line flags without editing code. It should be able to train, save checkpoints, resume from the same log folder, and write metrics to the terminal, JSONL files, and a simple results viewer. Please make it work on Linux and Mac, and assume Python 3.11 or newer.

Use the existing README as the source of truth for how training should look, and keep the defaults strong enough that I can run it without tuning. If there are any environment specific extras needed, handle them cleanly and make the setup easy to follow. If you need current details for JAX or related docs, look them up online.