q-rz/ICLR26-GRAPHITE — reverse-engineered prompt

Build me a usable research repo for the GRAPHITE paper, the graph homophily booster for heterophilic graph learning. The current repo looks like a placeholder, so please turn it into something I can actually run to understand and test the method from the paper.

I want a clean implementation of the model described in the paper, plus the normal research workflow around it, like loading the benchmark graph datasets used in the paper, training, validation, test evaluation, and a simple way to reproduce the main results as closely as possible. Please include sensible defaults, saved checkpoints, result logging, and a short script or notebook that shows the full pipeline from data download to final metrics. If the method depends on discrete node features or special preprocessing, make that easy to follow and well documented.

Also write a solid README with setup steps, example commands, expected outputs, and any caveats. You can look up the paper and current docs online if needed.