9cf2d94a75
#658 saturates the float16 stores of the mHC pre/post kernels; the fused hc_head kernel still wrote its float32 weighted sum straight into the float16 output. For an attention-sink row, whose four streams add up past 65504, the head input became inf. Clamp the store to the float16 range under use_fp16, the same as the other mHC stores. The new test builds a sink row whose exact hc_head output exceeds float16 and checks that the kernel stays finite, saturates that row and matches a float32 reference elsewhere; on main it fails with inf in the sink row. Signed-off-by: Peuqui <peuqui@github.com> Co-authored-by: Peuqui <peuqui@github.com> Co-authored-by: Claude <noreply@anthropic.com>