Function tap_highway 

pub fn tap_highway(
    gpu: &dyn GpuBackend,
    streams: DevicePtr,
    layer: usize,
    tag: &str,
    num_tokens: usize,
    hc_dim: usize,
    stream: u64,
)

Tap the FP32 highway. No-op unless the dump directory is set.

Synchronizes before reading, so it must never run inside CUDA-graph capture — which is already true of this model’s path (ATLAS_DEBUG_NO_GRAPH).