AUS-1012: Publish the benchmarks from the cpu-performance job, regenerated from data

[Astn]

Linear: AUS-1012 (parent AUS-1001; supersedes AUS-985). Built on #171 (AUS-1011), now merged.

The published benchmarks now come from one Hugging Face Jobs cpu-performance run (AMD EPYC 7R13, 32 cores, whole host) instead of the desktop, and republishing them is a pipeline rather than a transcription.

Pipeline

• Every benchmark mode writes a JSON file when JSONRPC_BENCH_RESULTS is set (TestServer_Console/BenchResults.cs; benchmarks/Baseline has its own writer). Rows carry a stable id, throughput and the CPU usage from AUS-1011: CPU-utilisation axis per benchmark row, derived machine description, Hugging Face Jobs runner #171.
• benchmarks/hf/run.sh <ref> publish runs the README's policy on the job host: the scaling gate once, two rounds of --sync, --async 3 1, --async 3 N, --kestrel, --kestrel async and --compare, three legacy t runs, two baseline runs and five --sweep 2 runs, every run and every JSON file framed on stdout. fetch.py extracts them.
• benchmarks/charts/ingest.py <dir> folds the JSON into benchmarks.json (measured fields only; ids, labels, notes and groups stay hand-maintained; idempotent) and copies the sweep files.
• render.py renders the SVGs, the explorer and the README tables between <!-- benchmarks:<set> --> markers (README.md and Json-Rpc/README.md); render.py --check fails on a stale block, alt text or chart.
• The sweep and sync ladders reach the core count (powers of two up to Environment.ProcessorCount).

Republishing is: launch.sh <ref> cpu-performance publish, fetch.py, ingest.py, render.py, commit.

Data

Job 6ab9b3936b030d633f69b0fb (https://huggingface.co/jobs/astnh/6ab9b3936b030d633f69b0fb), commit 6c3c272, 2026-09-28. Gate 12.2 / 12.1 / 13.1 at 32 workers, pass. Sync 44.2 M to 45.7 M at 32 threads; Kestrel TCP 18.5 M at 32 clients; the 2.0 string API 9.8 M to 11.1 M against 1.2.3 at 1.29 M to 2.41 M (the two baseline runs differ widely; the ratios use the medians). The WebAssembly table stays a browser measurement on the desktop and says so.

Checks

render.py --check, the 43 chart tests, the request-path checker, the site build and the solution build and tests (1220 on net8.0 and net10.0) pass; a second ingest.py on the job directory reports no change.