Decode is 5–7% faster on the 35B. The MoE block — the largest remaining
inefficiency in the per-token decode step — now runs as two custom Metal kernels
(JIT-compiled at load, no wheel rebuild): one dispatch computes gate|up for all 8
routed experts plus the shared expert with the SiLU epilogue in-register, the other
folds every down-projection, the routing weights, the shared-expert gate, and the
residual add into a single pass. Routing itself is bit-identical to stock; measured
on the shipped checkpoint: 70.7 → 74.3 tok/s at 8k context, decode-path perplexity
a hair better (the fused combine accumulates in fp32), prefill unchanged
(756 tok/s @7k vs 723 baseline), peak memory unchanged. Engages only on the exact
35B geometry, falls back to the stock graph anywhere else; opt out with
CHAD_NO_MOE_FUSED=1.
