Function w4a16_v3_kernel 

pub fn w4a16_v3_kernel(gpu: &dyn GpuBackend) -> KernelHandle

Resolve the W4A16 m128 v3 GEMM. Opt-in ONLY, same contract as w4a16_v2_kernel: ATLAS_W4A16_VARIANT=v3 selects it, anything else resolves to a ZERO handle WITHOUT issuing a lookup.

Not issuing the lookup is the point. prefill_weights dispatches on v == 3 && handle != 0, so on the default (v1) the probe could never be used — it only ever added a permanently-failing row to the boot audit on every target that does not ship w4a16_v3. Requesting the variant on such a target is a HARD error, not a silent fallback discovered in a perf report.