Train-ready SFT mix for a non-Thai-base DeepSeek/Qwen-style <=14B candidate.
Delta from v44:
reuses all v44 normalized MCQ replay and identity rows unchanged
adds locked, verified synthetic OpenThaiEval-style rows
normalizes those bridge rows so explanation precedes the final answer
keeps single-model training only; no BoN, self-consistency, routing, ensemble, or benchmark-label leakage