low-rank-reconstruction

Proposes an SRC pipeline that uses entropy-based selection and low-rank reconstruction to summarize KV cache instead of pruning tokens, reducing VRAM for million-token LLM contexts while avoiding catastrophic attention errors.