fix(seo): publish the URLs the host actually serves

[Batuis]

Stacked on #171.

The problem

Pages are directories on disk, and GitHub Pages answers a directory request with a 301 to the slashed form. That cannot be turned off. Every URL the site published named the unslashed form, so all eighteen were addresses it redirects rather than addresses it serves:

sitemap says      https://stabileo.com/en/blog
the host answers  301 → https://stabileo.com/en/blog/
and that page     <link rel="canonical" href=".../en/blog">

A canonical pointing at a URL that redirects back to the page declaring it, and an hreflang set built the same way.

Why it matters, stated honestly

This is not a ranking lever. Google resolves the chain in practice, and nobody moves up a position because of it.

It is worth doing for two narrower reasons:

1. The language linkage. Google is explicit that the canonical and the hreflang have to agree, and that if they do not it may ignore the hreflang — which is the three-language connection the whole prefix scheme was built to express.
2. Timing. The site went live days ago and is still being crawled. Changing the canonical form after the URLs are indexed costs a reindex and a period of ambiguity. Doing it now costs nothing.

The Search Console reports we are about to rely on will also be readable: sitemap URLs will appear as indexed pages rather than as "page with redirect".

What changed

publicHref ends every path in a slash; the sitemap, the canonical, the hreflang set and every internal link follow from it. parsePublicPath already read both forms, so links people already hold keep working — they take the one redirect they always did.

The root language handoff now sends /es/ instead of /es, which saves a Spanish or Portuguese visitor a second redirect that was ours to avoid.

The gate

A test that asserts the artifact rather than the intention. For every <loc> in the sitemap it requires a real file at that exact path, that the page declares itself canonical at the address it is served from, and that every hreflang target is itself served rather than redirected.

Checked on disk, not over HTTP, on purpose: vite preview serves both forms happily, so an HTTP test would have passed on the very defect that reached production.

The first draft used existsSync, which answers true for a directory — precisely what gets redirected — so it passed when the fix was reverted. Caught by running that control; it now requires a file.

Considered and rejected

Keeping the unslashed URLs by emitting flat files (en/blog.html rather than en/blog/index.html). It works for posts — GitHub Pages serves /404 from 404.html with no redirect, verified — but not for /en, where an en.html file and an en/ directory would have to coexist and the winner is up to the host. Too fragile to save a slash.

Gates

test:unit 7,132 · typecheck at baseline 479 · check:gate · build with prerender · @smoke + @landing.