Say the numbers the demo prints, not the ones it used to print

[PNX89]

The page disagreed with its own committed evidence file about three headline figures.

	README said	docs/evidence/demo.txt prints
trial-corrected pass mark	23 of 56, rather than 21	25 of 56, not 24
random-baseline score	14.4 of 56	16.4 of 56

A fresh capture_evidence.py run reproduces the demo's figures exactly, so the evidence was right and the prose had been left behind.

This is the defect this repository exists to detect, in this repository. QUIZZ answers a question as of a knowing-time rather than as of now, because a number that was true when it was written and is false today is the failure that costs people money. Three of its own headline numbers were precisely that: computed once, typed into prose, and left there while the thing producing them moved. A reader who runs the demo, as the page invites them to, saw a different number from the sentence beside the invitation.

The test matters more than the correction, and it took three attempts:

1. Search the whole page for each digit from the demo's sentence. Too weak. Putting the wrong sentence back left it green, because a page this long contains a "25" somewhere regardless. Searching a document for a digit is not a comparison.
2. Require the demo's sentence to appear verbatim. Too strong. The page legitimately rephrases the baseline claim in its own words, so a correct page failed.
3. Compare the figure in the demo's sentence with the figure in the page's corresponding sentence. The page may say it however it likes, and it may not say a different number.

Both wrong turns are recorded beside the code, because the first one is the same weak assertion this portfolio's audit keeps finding elsewhere.

Mutations: restoring 23 fails it; restoring 14.4 fails it with demo prints 16.4 for the random baseline and the page says 14.4.