Produce and defend a number: run perftest and nccl-tests correctly, prove GPUDirect is actually on, and translate busbw into a statement about which piece of hardware is the bottleneck.

1. 3.1 What ib_write_bw actually measuresThe 180 that was never 180Ap
2. 3.2 NUMA, data validation, and doca_perftestThe first word on the lineAn
3. 3.3 GPUDirect RDMA: proving it is actually onWhichever slot was easiest to reachAn
4. 3.4 NCCL environment variables on a RoCE fabricThe design was correct; the traffic never entered itAp
5. 3.5 algbw, busbw, and reading a nccl-tests resultThe 1.75x he called marketingAn
6. Checkpoint