Added NumPy, Arrow IPC, Feather, JSONL, Safetensors, HDF5 in an afternoon. 6 formats. Should've been a victory lap. Instead every single one broke something about ID handling. NumPy doesn't have row IDs at all. Arrow IPC has them but in a different column convention. JSONL keys can be anything. Safetensors has an explicit ID tensor I didn't notice until later. HDF5 has dataset names that may or may not be meaningful.
The feature work and the bug work were the same work. I couldn't add a format without discovering a new assumption I'd baked into the ID pipeline. Every format is a different opinion about what a "row" is and whether it has a name.