Three comments and a check

Each was one line to fix. Each was invisible until somebody happened to look at the right thing, and nobody looks at reachability on purpose — you look at traffic, or at a page, and reachability is the thing that explains the answer rather than the thing you asked about.

After the first, a note beside the fix is the obvious response. It explains the failure to the next person, it costs nothing, and it sits exactly where the mistake would be repeated.

The trouble is that it only works on somebody already editing that line. The fourth page was added by somebody who had no reason to open the file where the first three comments live — which is the normal case, not a lapse. A comment is read by whoever visits it. The fault happens somewhere else.

There is a point where repeating an explanation stops being documentation and starts being evidence. Three notes saying the same thing are a measurement: this will happen again, and the thing that catches it cannot be prose.

The check that replaced them is unremarkable. It reads the pages the site publishes, looks for anything that links to each one — as a path, or through the constant that holds it — and exits with an error naming any page nothing points at. The files that list every page by construction are excluded, because declaring a page is not linking to it, and a sitemap containing everything is the exact condition being tested for.

It was verified by removing the new link and watching it fail with the page named, then restoring the link and watching it pass. That step is not ceremony. A guard written after the fix is written against a codebase where the fault is already gone, so the only state it has ever observed is the passing one.

Plenty of checks in plenty of repositories have never once failed and never could. They are indistinguishable from working ones until the day they are needed, which is the same property the bug had.

Not whether a link exists — whether a page can be reached. Those sound alike and are not. A crawler finds pages by following links; a sitemap tells it where to look but is a hint, and on a site with little authority the hint is the weaker half by a long way. A reader has no sitemap at all.

So the page that nothing pointed at was the one page a search engine had never fetched, and could not have. What a sitemap cannot do has the numbers from that: forty-three pages offered, twenty indexed, eighteen never crawled at all.

Drawn from this site's own history on the date above. The four pages, the three comments and the check are all in the repository; the coverage figures it refers to were measured per address against a search engine's own record for this property rather than inferred from traffic.