It is not. That is the point here.
It's not independent.
If an LLM were able to generate code that can access ext4, then that LLM has to have ingested the original ext4 source code as part of its corpus.
LLMs cannot count. They cannot add 2 + 2. If it can emit code to do something then it contains in its model code to do that thing.
Therefore it is not an independent implementation.