Function preflight_device_arch 

pub fn preflight_device_arch(
    ordinal: usize,
    compiled_arch: Option<&str>,
) -> Result<()>

Fail fast if this binary’s kernels cannot run on GPU ordinal.

Call this BEFORE constructing the backend: AtlasCudaBackend::new loads every PTX module, and the point is to answer before the driver does.

compiled_arch is None when the build recorded no architecture — the ATLAS_SKIP_BUILD=1 stub registry compiles nothing and can attest to nothing. That is warned and skipped, never treated as a pass: a check with no input has no opinion, and inventing one would make the stub build claim hardware compatibility it never tested.