chore(pages): replace the gh-pages docs mirror with a branded link stub

[uipreliga]

Why

uipath.github.io/coder_eval/ was serving a second live copy of docs/, duplicating coder-eval.com/docs. Verified before this change:

• Both copies self-canonicalized — the mirror declared <link rel="canonical" href="https://uipath.github.io/coder_eval/"> rather than deferring to coder-eval.com. Google elects one canonical per duplicate cluster, and nothing told it which to pick.
• uipath.github.io/robots.txt is 404, so nothing blocked crawling, and the mirror shipped its own sitemap.xml actively advertising all 14 URLs for indexing.
• The mirror had drifted: it still served /BYOD/ and /CODEX_AGENT_GUIDE/, pages renamed upstream to DATASETS.md and agents/CODEX.md. So the risk wasn't only duplication — the copy competing with us was out of date.

There's no penalty for duplicate content, but the mechanical costs are real: split link equity on a domain whose measured problem is a shortage of inbound links, and Search Console reporting "Duplicate, Google chose different canonical" for pages we're verified for — optimizing without feedback.

What

Replaces the MkDocs deploy with a single branded stub page.

• .github/pages-stub/index.html — Coder Eval wordmark, a one-sentence description of what the product is, a link to the docs, and a brand-orange bar linking to this repo with its live star/fork counts.
• .github/workflows/docs.yml — publishes that stub instead of running mkdocs gh-deploy, force-replacing gh-pages so none of the old output (sitemap included) survives. 404.html is a copy of the stub, so Pages serves it for every old deep URL without needing a file per route.

Decisions worth reviewing

noindex, follow rather than a canonical to /docs. The stub's content differs from the target; Google discourages combining noindex with rel=canonical and ignores canonicals across substantially different pages. This takes the duplicate cluster to zero indexed pages rather than one.

The mkdocs deploy is deleted, not commented out. It had already been commented out once, and that is exactly how the duplicate got published — one uncomment away. mkdocs.yml itself is untouched and still required: CE028 generates the index surfaces from its nav:, the website derives its sidebar from it, and mkdocs build still verifies anchors locally. Only the publishing of a second site is gone.

One logo copy, not two. The wordmark is inlined from the website's src/assets/logo-light.svg. The light and dark variants differ in only four navy fills (three icon bars and the "Coder" text group — "Eval" and the arc are always brand orange), so those became currentColor and one 3.5 KB copy serves both colour schemes. The substitution was scripted with assertions that it recoloured exactly four fills and left #FA4616 intact — no path data was hand-copied. The page also moved onto the real brand palette (#FA4616, #10233A, #F5F7FA) from my earlier approximations.

Counts are never hardcoded. A committed star count is wrong within days, so the markup ships spinners and the script fills them from api.github.com. Anything that isn't a live number stays a spinner — offline, non-2xx, missing field, or scripting disabled. Resolution is per cell, so a partial payload fills what it has. Accepted tradeoff: with JS off the spinner never resolves, which is the price of never publishing a stale number on a page nobody should land on. Unauthenticated API is 60 req/hour/IP.

An earlier revision carried a snapshot value plus restore-on-failure, a 4s AbortController deadline and a race guard; dropping the hardcoded values made all of it dead weight, so the script is now 10 lines.

The page says what Coder Eval is. It previously named the product and said the docs had moved, so anyone arriving from a stale link had to click through to learn what it was. It now leads with a one-sentence description, worded to match coder-eval.com and the package metadata so the product is described identically everywhere. There is a matching <title> and meta description — but note those are for link previews, not ranking: the page is noindex, so search engines will not surface it. A comment above them says so, so they aren't later mistaken for SEO copy doing work it isn't.

No text is repeated. The wordmark already renders "Coder Eval", so the <h1> beneath it was the same two words twice; the heading is now visually hidden, keeping document structure and serving as the accessible name (the logo is aria-hidden, which also stops a screen reader announcing the name twice). Linking coder-eval.com/docs inside the notice then made the "Documentation" list row a second copy of the same URL two lines apart, so that row is gone. Each destination appears exactly once: docs in the notice, the project site in the list, the repo in the bar.

Verification

• Workflow YAML parses; the assemble step dry-run produces index.html, 404.html, .nojekyll, llms.txt.
• All three outbound links return 200.
• Renders in both colour schemes (confirmed for light by deleting the dark @media rule at runtime, so the wordmark's currentColor was actually exercised).
• Counts: live load fills the real values; network reject, a request that never settles, and a payload missing forks_count each leave the spinner in place. Markup contains no numeric literals for the counts.
• make lint passes (166).

After merge

This needs one manual workflow_dispatch run of the Pages stub workflow from the Actions tab — merging alone does not replace the live mirror.

Not in scope

• The sharp HIGH Dependabot alert on main (transitive in evalboard, fixed in 0.35.0) is pre-existing and unrelated.
• /llms.txt on the website was pointing three links at the uipath.github.io mirror; that is fixed separately in the website repo, along with adding /llms-full.txt.

[github-actions]

Claude finished @uipreliga's task in 1m 24s —— View job

---

Code Review In Progress

• Read .github/code_review.md for review standards
• Read CLAUDE.md for project conventions
• Analyze git diff origin/main...HEAD
• Read full files for context
• Review .github/pages-stub/index.html
• Review .github/workflows/docs.yml changes
• Check cross-file consistency
• Perform design-level scrutiny
• Submit structured review feedback

[bai-uipath]

LGTM. Deleting the mirror is right, and 404.html-as-stub is the correct answer to 15 dead routes. All follow-up below.

• Pull the wordmark out into a committed .svg. 5.1 KB of inline SVG is 35% of the page, 3.1 KB of it font outlines for two words the hidden <h1> already renders. Fix: reference it root-absolute (/coder_eval/logo.svg), since 404.html serves at arbitrary depths, and use a <picture> for light/dark because an external <img> can't inherit currentColor; website/src/assets/ already has both variants.

• The multi-repo split is the real exposure, not the stub. docs/ is the source of truth here but publishing now lives entirely in coder_eval_uipath/website/ via scripts/sync-docs.mjs. After this merges nothing in this repo publishes docs anywhere, and mkdocs-material[imaging]>=9.5,<10 is pinned nowhere while CLAUDE.md:208 still says to verify anchors with mkdocs build.

• coder-eval.com is served from an unmerged branch. coder_eval_uipath 62 has been open since Jul 22 at 73 commits and the live site tracks its head. The stub hardcodes /llms-full.txt and copies logo-light.svg, both of which exist only there. Not blocking (site is live, all links 200). Fix: land internal 62, and note the logo's source commit so it reads as a copy.

• Merging doesn't take effect. workflow_dispatch-only leaves the stale mirror live until someone remembers the Actions tab. Fix: push: on main with paths: ['.github/pages-stub/**', '.github/workflows/docs.yml'].

• Minor: #FA4616 as text fails WCAG AA (3.53:1 on white, same for white-on-orange in the bar; rest of the palette is 5.5:1+); with JS off or rate-limited both counts spin forever, so consider dropping them; noindex is a no-op for the 15 deep paths, which now 404 on status.