Counted claims on the demo site are still unguarded

Follow-up to #55, which bound the READMEs' 31 counted claims to version.json.

The same numbers appear 26 more times outside the READMEs and nothing checks them:

All 26 currently agree with version.json. The gap is that nothing keeps them agreeing.

Why #55 stopped at the READMEs

Running #55's patterns over the site finds two matches that are correct English but not claims, so extending the check as written would fail the build on valid copy:

• demo-pricing/app/gallery/page.tsx:55 — Button under 3 skins. A live preview that shows three of the seven, not a count of what ships.
• docs/NEXT-RUNTIME-IMPLEMENTATION-PLAN.md:48 — 8 recipes inside a validate-palettes transcript, where recipes means palette recipes.

Both are shapes, not one-offs: a count scoped to a page's own sample, and a keyword whose meaning depends on surrounding context. Any extension has to handle them rather than special-case these two lines.

Suggested direction

• Read the counts from version.json in the site components instead of typing them, which removes the class of drift rather than detecting it. The value is already published at /version.json and the site already reads that file.
• Where prose has to stay prose, mark it — an explicit opt-out comment beats an allowlist keyed to line numbers, which rots on the next edit.
• Keep the docs transcripts out of scope. They record what a command printed on a date, and rewriting them to match a later count would make them false.

Not urgent: everything agrees today, and the README surface that evaluators read first is now gated.

🤖 Generated with Claude Code

https://claude.ai/code/session_016b5vcXw4iawXAZo4rHuuEf

Start with the existing README checks from #55 and the site's /version.json reader, then inspect the listed demo-pricing/app pages and docs files for the remaining claims. Determine which site counts can read published values and how explicit opt-outs should work; keep dated docs transcripts unchanged. Done means the listed live claims stay synchronized without false positives for sample counts or transcript text.