Capture accuracy on five real sites

Finding

On 2026-10-30, capture was audited on five public sites: ritesh-firodiya.github.io, tailwindcss.com, vite.dev, demo.vercel.store and news.ycombinator.com.

Proof

There is no snapshot in raw/. These facts come from the task text and commits.

• A walk spent its limit on one template: Tailwind gave 20 docs pages of 25. Fixed by spreading over the site (dd5774e6).
• Unnamed links were grouped by how they were built: Tailwind's 501 "LinkButton"s were sidebar, nav and footer links (ee00116d, 8c223a7c).
• Pale and dark tints were filed as grey, and a nested footer was taken for the Footer (3b1b3902, 3797f004).
• A sandboxed frame hung the accessibility check for 90 minutes, so each page now has a deadline (12c19608).
• Brand colours already came through computed values, so reading CSS variables was dropped (3b1b3902).

Verified by

The sessions that ran the audit and made the fixes, as their commits record. Signed-in capture was later tried on staging on saucedemo.com and the-internet.herokuapp.com (12c19608). Nobody re-ran all five sites as a final check, as far as git shows.

What it means

Test capture on real sites, not only on the seed. Rules cannot be right everywhere, so the person needs to correct a capture: corrections are in 89eb2347 and 5ca27d22. The problem page is PROB-0003.

Evidence

• Commit ee00116d, the audit list in .context/tasks.md.
• The commits named above.