Add tailored from-scratch training stack (tokenizer + random-init LM) for math conjecture solving

Retarget model to math-conjecture solver profile; remove DeepSeek configs and legacy sync snapshot

Keep training stack parity: CPU fallback path and mixed-precision safeguards.

Upgrade SOTA curriculum: 4-stage training, post-eval metrics, and quality-gated promotion.

Add SOTA advancement pipeline: multi-stage weighted curriculum + pass@k eval harness.

Rename model repo target to math-conjecture-model and upload pipeline.