Show the Core Web Vitals evidence instead of asserting it

[alexburan]

Why

The whole pitch is that locale pages keep the source language's performance. Until now that was a claim with nothing behind it.

The evidence

doctranslator.com/fr is a large Astro site whose French pages this pipeline built. Against the English original the markup is byte-identical:

2,743	tags, in identical sequence
2,031	class attributes, all matching
+0.74%	page weight — only because French is longer

Nothing structural moved, so there is nothing for the browser to lay out differently. Ten seconds to check:

curl -s https://doctranslator.com/   | grep -o 'class="[^"]*"' > en.txt
curl -s https://doctranslator.com/fr | grep -o 'class="[^"]*"' > fr.txt
diff en.txt fr.txt && echo "markup identical — only the text changed"

PageSpeed performance, measured 2026-08-25: desktop 100 vs 100, mobile 98 vs 98. Accessibility, best practices and SEO were 100 on every run of both pages, both strategies.

Published as medians, with the spread stated

Five consecutive runs of the same English mobile page returned 98, 98, 88, 98, 98, LCP swinging 1.8s → 3.2s. That is network and CDN variance, not page quality.

A reader who measures once and sees 91 should have been told why in advance, so the caveat is on the page rather than buried. SKILL.md tells the agent to quote the markup identity first, because that is the part that reproduces.

New failure mode: locale pages built from a stale source build

Found live while gathering the above, and worth the entry in references/failure-modes.md.

The two DocTranslator pages disagree about their inlined critical CSS despite identical markup — because the source was rebuilt three hours after the locale pages were generated, and critical-CSS extraction is not stable across builds. The later build gained .bg-sky-500, .bg-slate-100, .object-cover, .scale-150 and lost .aspect-video; neither build's CSS correctly matched its own markup.

That is the same instability that makes per-locale re-rendering a bad idea, surfacing somewhere else. Fix: re-run build-locales.mjs after any rebuild that changes markup — nearly free, since the memory is keyed by source text and nothing gets re-translated.

How to spot it: substitution never touches <style> contents, so if those differ between a page and its locale twin, the twin came from a different snapshot.

Verification

• Every figure re-verified against the live pages immediately before writing it down
• npm run check and all 32 contract tests pass