chore: exclude platform-side stale zh URLs from link checker

[github-actions]

Summary

The scheduled link checker keeps re-firing on Chinese-locale URLs that are valid in source but broken in the deployed Chinese rendering. This PR lands Option 2 from the original analysis: temporary Lychee excludes for the affected URLs so the checker stops re-scaffolding this PR until the underlying bug in fern-api/fern-platform is fixed.

What this PR does

• Adds two Lychee excludes (with inline TODO + comment explaining why) in .github/workflows/check-links.yml:
  • buildwithfern.com/products/sdks/custom-code.mdx — flagged on the rendered Chinese page /learn/zh/sdks/deep-dives/readme even though the source link was corrected in Fix broken links (Devin) #5348 (May 5). Chinese translation S3 blob is stale.
  • buildwithfern.com/learn/zh/docs/ai-features/ask-fern/api-reference/code/get-code-records — listed in sitemap-zh.xml, .md content returns 200, but the HTML page returns 404 (English equivalent serves 200).
• Removes the .github/broken-links/broken-links.md scaffold (no-op vs main; net effect is keeping the working tree clean).

Why no source change

Source files for both URLs are already correct. The English equivalents serve correctly. The bug is in how fern-api/fern-platform renders/publishes the Chinese-locale variants:

• For the readme page: the FDR registerTranslation path is leaving stale S3 content despite repeated successful Publish Docs runs.
• For the ask-fern API reference page: HTML routing returns 404 even though the underlying content is generated and served correctly via the .md URL.

The 5xx URLs from this run (enum-descriptions-and-names, multipart) were transient flakes — both now return 200 — so they are not excluded.

Why excludes (vs. close PR / scaffold-deletion-only)

The previous revision listed three options and deferred to @davidkonigsberg. Without the excludes, the M/W/F link-check job keeps re-creating .github/broken-links/broken-links.md and re-pinging Devin in Slack on every run. The excludes are the smallest change that stops the noise; the inline TODO points back to this PR so the excludes can be removed after the platform fix lands.

Review & Testing Checklist for Human

Risk: green (workflow YAML only, no source/runtime impact)

• Confirm you're OK temporarily masking these two URLs from the link checker until the fern-platform fix ships. Inline TODO references this PR for cleanup.
• Sanity-check the regex shape against the existing entries above them (escaped ., no leading anchor — matches Lychee's substring semantics for excludes already used for _next/image and pdf-ui.png).
• Track a follow-up to remove these excludes once fern-api/fern-platform ships the translation rendering fix (FDR registerTranslation / docs bundle locale routing).

Notes

• Source PR that fixed the underlying source link: Fix broken links (Devin) #5348
• Recent translation-area committers in fern-platform: Deep Singhvi ([email protected]), Sandeep Dinesh ([email protected]), Sarah Bawabe
• Reverted no-op nudge commit on this branch: c1ac544e
• Live verification at the time of this revision:
  • https://buildwithfern.com/learn/zh/sdks/deep-dives/readme.md still ends with (/products/sdks/custom-code.mdx) (broken)
  • https://buildwithfern.com/learn/sdks/deep-dives/readme.md ends with (/learn/sdks/overview/custom-code) (correct)
  • /learn/zh/docs/ai-features/ask-fern/api-reference/code/get-code-records HTML → 404; .md → 200; English HTML → 200

Link to Devin session: https://app.devin.ai/sessions/4421ec58c97f48ea9fa26c32a7660770

[github-actions]

🌿 Preview your docs: https://fern-preview-devin-fix-broken-links.docs.buildwithfern.com/learn

[devin-ai-integration]

Updated PR description with the full analysis. TL;DR:

• The newline change was a no-op — reverted in c1ac544e. The CLI doesn't dedup, so changing file content doesn't change what gets uploaded.
• The bug is platform-side. Source has been correct since Fix broken links (Devin) #5348 (May 5); 7+ successful Publish Docs runs since haven't updated the Chinese translation S3 blob, even though they correctly updated the English base blob.
• Manually invalidating + revalidating with ?locale=zh doesn't help — the response is uncached (x-vercel-cache: MISS, cache-control: no-store), so the stale content is coming straight from FDR/S3.

Most likely candidates inside servers/fdr/src/controllers/docs/v2/getDocsWriteV2Service.ts:registerTranslation:

1. Stale base response from app.docsDefinitionService.getDocsForUrl(...) during the merge step
2. S3 write going to a different key than the read path uses (basepath/trailing-slash mismatch)
3. pageId encoding mismatch in mergedPages so submittedPages[pageId] silently misses for this one page

Need someone with FDR CloudWatch / S3 access to confirm which one. Recent translation-area committers: Deep Singhvi, Sandeep Dinesh, Sarah Bawabe.

[devin-ai-integration]

Tested on the preview deploy — the Lychee exclude approach is the right call here.

Preview verification (fern-preview-devin-fix-broken-links.docs.buildwithfern.com):

URL	Preview	Production
/learn/zh/sdks/deep-dives/readme HTML	200	200
/learn/zh/sdks/deep-dives/readme.md content	ends with (/learn/sdks/overview/custom-code) ✅	ends with (/products/sdks/custom-code.mdx) ❌
/learn/sdks/deep-dives/readme.md content	ends with (/learn/sdks/overview/custom-code) ✅	ends with (/learn/sdks/overview/custom-code) ✅

The preview renders the corrected source faithfully on both locales. Production only renders it correctly for English. So:

• The source (and the docs build pipeline) are correct.
• The bug is downstream of the build — somewhere in the production publish path that writes/serves the Chinese translation blob (most likely fern-api/fern-platform's registerTranslation in FDR, per the analysis in the PR description).
• A second instance of the same class of bug surfaced this run: /learn/zh/docs/ai-features/ask-fern/api-reference/code/get-code-records is in sitemap-zh.xml and serves the correct .md, but its HTML returns 404 in production while the English equivalent returns 200.

The 5xx URLs in this run's report (enum-descriptions-and-names, multipart) were transient — both now serve 200 — so they're intentionally not excluded.

@davidkonigsberg requesting review on the workflow YAML change. The two new excludes have an inline TODO referencing this PR so they're easy to remove once fern-api/fern-platform ships the translation rendering fix.