GPU run provenance

Raw evidence for the real-GPU numbers quoted in ../real-gpu-results.md, so each figure traces back to a command, a date, and captured output instead of being transcribed by hand. This mirrors the convention already used for the local mock-vLLM evidence (see ../../local/runs/2026-08-27-scale-out/README.md for an example), applied to the physical two-GPU cluster.

Layout

One directory per run, named <date>-<run-name>/ (e.g. 2026-08-27-3b-kv-saturation/), containing:

File	What it is
command.txt	The exact gpu-loadtest.py invocation — concurrency, duration, --base-path, model — plus the vLLM server flags (--gpu-memory-utilization, --max-model-len, --enforce-eager, etc.) it was run against.
stdout.log	Raw stdout of that gpu-loadtest.py run (client-side TTFT percentiles and throughput).
nvidia-smi.log	nvidia-smi output captured on the serving node while the load was running (power draw, utilization, memory).
vllm-metrics.prom	The raw curl <node>:8000/metrics scrape at peak load — the server-side cross-check (vllm:gpu_cache_usage_perc, vllm:num_requests_waiting, vllm:generation_tokens_total, …) that the Method note in real-gpu-results.md refers to.
image-digest.txt	The vLLM image digest the run used (docker inspect --format='{{index .RepoDigests 0}}' <image> or the equivalent crictl/kubectl lookup), so the exact engine build is pinned.
README.md	Short human summary of the run — what it shows, any caveats — same as ../../local/runs/*/README.md.

capture-run.sh automates writing the first four files for a new run; see its --help for usage.

Status

This convention did not exist when Runs 1–4 in real-gpu-results.md were captured, and the raw nvidia-smi/vLLM-scrape/stdout output from those sessions was not retained outside the derived numbers already written into that file — there is nothing genuine to commit here for them without re-running the benchmarks on the physical two-GPU cluster. Backfilling those four runs, and adding the date + run-directory link to each table caption in real-gpu-results.md as called for in #5, is left as follow-up work to do from the GPU nodes themselves. Every run captured from here on must land in this directory before its numbers go into real-gpu-results.md.