moe.miner_topology (default full) replaces the unread partial_moe: the miner
trains with every routed expert present and native top-6 routing, the model
validators score, with frozen experts in fp8. The per-step hash and the
single-group save no longer dequantize the whole model. Validators unchanged. (#283)