- configuration_corm.py: pass `tie_word_embeddings` to `PretrainedConfig.__init__`. It was reset to True, so recent transformers tied `lm_head` to `embed_tokens` and discarded the trained output layer.
- config.json: regenerated from the original Megatron checkpoint; `use_dynamic_routing=false`, `top_p=1.0` (as trained), `has_shared_expert=false`; inert shared-expert schedule keys removed.
- modeling_corm.py: `routing_temperature` has shape [1], matching the checkpoint.
Weights are unchanged. Verified: re-conversion reproduces model.safetensors tensor-for-tensor, and the loss on Megatron's validation tokens matches the paper-evaluation code.
