Segment-Level Training/Inference Alignment for Efficient Long-Context LLMs (PoRigyDOcC)