Function compile_to_rdu 

pub fn compile_to_rdu(
    source: &str,
    chip: RduChip,
) -> Result<RduArtifact, ParseError>

Compile to the SambaNova RDU — ONE front door, three lanes chosen by program shape: the tensor lane (a tensor program lifted through the spatial IR, placed on the chip’s descriptor and certified), the loop lane (a counted-range-loop function through the Plasticine metapipeline model on the RDU mesh descriptor), and the whole-program lane (any construct the fabric’s decision datapath covers, as Scalar stages). Every lane ends in the same certified neutral envelope, honestly SDK-gated for the vendor config bytes. A tensor program the chip cannot host is refused BY NAME on the tensor lane, never handed on; a program no lane hosts gets a rdu backend: refusal naming every lane’s verdict.