/status/benchmarks/ 503s while every sibling serves

Measured 2026-09-05 ~18:30 UTC, from plexus, with plain curl and again with tools/doclinks.py — two instruments, and repeated across several hours:

https://pxxc.org/                     200
https://pxxc.org/status/              200
https://pxxc.org/status/tests/        200
https://pxxc.org/status/benchmarks/   503   (twice in a row, 25s timeout)

Only the leaf fails. The siblings responding rules out DNS, TLS, routing and the host being down, which is why this is filed as a page rather than reported as "the site is broken".

It is a regression against a verified baseline, not an unchecked link

[[task-d-verify-the-published-status-urls-docs-now-delegates-all-numbers-to]] fetched all eight external URLs on 2026-08-30 and recorded:

  ok   https://pxxc.org/status/benchmarks/   [fib sieve]

— a 200 whose body contained the expected content markers. So the page existed, served, and carried per--O-level timings against FPC six days ago.

Why it matters more than a broken link usually would

docs/reference/status.md carries no figures at all by design (ce89ff14b) and delegates every number to these URLs. Line 16 tells the reader that benchmarks is where the timings are. The page therefore promises current numbers and delivers a 503 — which is exactly the failure mode the prior ticket was filed to defend against, arriving for the first time.

Docs deliberately NOT changed

docs/reference/status.md:16 cites the page correctly; the link is not the defect. Removing or rerouting it would take Track D's external-link gate to green by deleting the only thing currently pointing at the outage — a green bought by removing the observation, which is the same move as fixing a false RED by deleting the guard.

So Track D's doclinks.py will keep reporting exactly one BROKEN row until this is fixed or the page is deliberately retired. That row is doing its job.

Routing

Filed to W because the fix is the publish path or the host, not this repo's prose. Noted for whoever takes it: ~/pxx-website was last committed 2026-08-29, and a 503 on one deployed page while its siblings serve looks more like generation or infra than a repo change. frank-coordinator independently reproduced the four rows above and escalated to the owner the same evening, since nobody was on W that night.

What this is evidence FOR, beyond itself

ce89ff14b deliberately traded "a number goes stale (visible, self-correcting)" for "a URL fails (invisible from this repo)", and [[task-d-verify-the-published-status-urls-docs-now-delegates-all-numbers-to]] recorded that the residual had never once occurred. This is the first time it has, six days into the new regime — so the rate is one in six days rather than an unknown, and doclinks.py caught it, which is the mitigation the trade assumed rather than merely hoped for. Both facts are written into that ticket too, because that is where someone revisiting the decision will look. The trade stands, and now for a measured reason.

If this is still red with no movement, say so again rather than absorbing it. A standing red everyone has learned to expect is how a real outage becomes wallpaper — and this one is invisible from inside the checkout by construction.