What does LoRA fine-tuning actually train?

"LoRA trains about 1% of the model" is the usual one-line explanation. It's true, but it hides the...