Skip eval, disable MoE aux loss, disable dropout to fix OOM

Reduce context 4096->3072, add expandable_segments for OOM fix

Remove max_prompt_length (not in TRL 1.9 DPOConfig)

Use DPOConfig for TRL 1.9 compatibility

Fix DPOTrainer: tokenizer -> processing_class (TRL 1.9)

Fix target modules: remove gate (custom MoE router)

Add full plan doc

Genesis-2.0 RLHF Phase 0 artifacts

DPO training script

initial commit