Reproduce a published result

Every published record names its method, configuration and fingerprints, so a published result can be reproduced on any GPUs that fit the model.

1. Install the GPU environment with PyTorch 2.9.1 and prepare the data; the trainer verifies the tokenizer and validation shard against the pinned hashes.
2. Check that the method still matches its proof: python -m nexttransformer.protocol.fingerprint --methods <method>.
3. Train with the admitted configuration, --formal and a fresh output directory, as in the quick start. Any registered layout (8, 16 or 32 GPUs) keeps the data order.
4. Audit the run with python -m nexttransformer.audit_run, then compare its validation BPB and CORE at update 46,875 with results/runs/<method>.json.

Runs are single-seed; expect agreement in data, compute and evaluation schedule, and numerical rather than bitwise agreement across different GPUs, kernels and rank counts. A few published runs were trained from a frozen source revision; their records say so in comparisonNote, and the run notes explain what differs. A reproduction that confirms or contradicts a record is welcome as a run-notes contribution.