The cache filename format was inconsistent:
- get_cached_model_path used: {hash}{ext}
- load_model download used: {hash}_{filename}

This caused cache lookups to always fail. Now both use {hash}_{filename}
format to ensure cached models are properly found.