Correct the scale claim and publish the evidence behind it

[godofecht]

The correction

The README claimed CI wins 34 of 34 scaled rows. Main's own run says 32 of 34, and the two rows it lost are not the two the earlier runs lost. Across five consecutive runs the counts are 32, 32, 34, 34, 32.

row	run 1	run 2	run 3	run 4	run 5
KernelSVC_RBF 1000x8	0.68	0.66	1.29	1.30	1.16
KernelSVC_RBF 1000x32	0.88	0.83	1.29	1.34	1.36
Lasso 1000x32	3.83	4.07	1.51	1.47	0.92
LinearRegression 1000x32	2.18	1.79	1.80	1.87	0.90

All four are at 1000 samples, where a fit finishes in well under a millisecond. The bottom two swing by 4x on code that differs in nothing touching them, so that is the runner. The top two show a step change at the support-vector commit and have won every run since, so that is the code.

The published claim is now the spread: 30 of the 34 rows win in all five runs, and the ten rows at 10000 samples, the only sizes large enough to measure without fighting the noise, win in all five.

What backs it

summarize_scaled_ci.py folds each run's scaled_comparison.json into scaled_ci_history.json, keyed by run id so a new run extends the history and a repeated one replaces it. publish_headline_v2.py validates it and copies it into the Pages tree. The benchmarks page renders it as a table sorted by median, with runs-won and the observed range beside each row.

Drift the same shape, found on the way

The check that already self-heals the win counts now covers the substrate means too. It immediately caught real drift in both READMEs: 20.99x / 6.84x / 3.96x against an artifact holding 19.04x / 7.39x / 4.29x, and 26.55x against 19.04x.

The site was worse. It still said Flow wins 75% of Python-bound rows, 45% of mixed and none of the external-native-bound ones, and the architecture page said it loses every external-native-bound comparison. Both predate winning all 19. index.html now builds that sentence from the artifact it already fetches, so it cannot go stale again, and its hardcoded 8 / 19 fallback is gone.

Removed

docs/benchmarks.js and docs/benchmark-corrections.js. No page has loaded either for some time, but benchmarks.js still shipped 600 lines of hand-written old timings to a public URL, and two CI steps existed only to sanitize it. Those steps and the two scripts behind them go with it.

Also

• benchmarks.html had a duplicated unterminated <section> opener, leaving the markup one tag out of balance.
• Site copy brought in line with the house rules. The commit hook found these, including some I introduced.
• The generated docs/*.json evidence copies are now gitignored; their sources live in benchmarks/.

Verification

Every page script parses under node --check and was rendered against the real artifacts in a DOM harness; the pages were also loaded in a browser. All four evidence validators pass.