fix: make the redirect gate actually gate, and stop fighting the canonical

Both from CodeRabbit's review, both real.

The coverage check called pytest.skip when the mkdocs build failed. A skip
exits pytest 0, so the workflow would have read it as a pass and deployed
redirects that had never been compared against the site they replace — on the
one deploy that cannot be redone afterwards. The skip bought a quieter local
run without mkdocs installed and paid for it with the entire point of the
check. It fails now; verified by moving mkdocs.yml aside, which turns both
tests red instead of green.

The stubs also carried robots: noindex, which contradicts the rel=canonical
sitting two lines above it — one asks a crawler to consolidate the page onto
the new URL, the other asks it to drop the page, and the second can stop the
first from being acted on or carry across to the target. The comment in this
file claimed the canonical was what moved search ranking while the markup was
undermining it. Removed, with a note on why it is not an oversight.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01U9zFxq8V4qxG4aMzJhGBFn

Author: claude

scripts/build_redirects.py

@@ -10,6 +10,12 @@
 only mechanism available. The canonical link is what transfers search ranking
 to the new URL; the meta refresh and the script are what move a reader.
 
+Deliberately no ``robots: noindex`` on these stubs. It reads like the tidy thing
+to do, but it contradicts the canonical: one says "consolidate this page onto
+that URL", the other says "drop this page from the index", and a crawler that
+honours the second may never act on the first — or carry the noindex across to
+the target. A migration wants the canonical to be believed.
+
 Why this runs instead of ``mkdocs build``: once this repository is archived,
 Actions stop running and the last deployed artifact is what Pages serves
 forever. That artifact needs to be the redirects, so the redirects have to be
@@ -62,7 +68,6 @@
 <meta name="viewport" content="width=device-width, initial-scale=1">
 <title>Moved to commit-check.com</title>
 <link rel="canonical" href="{target}">
-<meta name="robots" content="noindex, follow">
 <meta http-equiv="refresh" content="0; url={target}">
 <script>location.replace("{target}" + location.hash);</script>
 <style>

tests/redirects_test.py

@@ -36,7 +36,10 @@ def _urls_the_old_site_served() -> set[str]:
             text=True,
         )
         if result.returncode != 0:
-            pytest.skip(f"mkdocs build unavailable: {result.stderr.strip()[:200]}")
+            # Not a skip. A skip exits pytest 0, the workflow reads that as a
+            # pass and deploys redirects that were never checked against the
+            # site they replace — on the one deploy that cannot be redone.
+            pytest.fail(f"mkdocs build failed: {result.stderr.strip()[:200]}")
         site = Path(tmp)
         urls = set()
         for page in site.rglob("index.html"):