follow dev73_1, 1. 完善C++ CUDA Clover多网格V-cycle求解器(扩展lattice_clover_multigrid.h至约1100行,优化multigrid.cu/apply_clover_multigrid.cu的restrict/prolong/coarse-dslash kernel与5流同步); 2. 重构并新增多网格测试/基准脚本(重构conftest.clover.multigrid.py,新增bench_cpp_mg/mg_dev_*/v3_*/quick_test/verify_coarse/diag_coarse等); 3. 输出多网格调试/优化/v2/v3报告与收敛曲线(logs/mg-*); 4. 重构qcu.pyi类型存根,修正Cython桥接的C API类型声明; 5. 补全全库CLAUDE.md文档体系,每个父目录嵌入子目录完整skill并新增.claude/skills、cpp/cuda/qcu/logs的CLAUDE.md; [Claude Code].