Stop the generator dropping public surfaces from llms.txt and the h1

[sarthakagrawal927]

Found during a fleet-wide audit of AI-discovery surface drift. Both losses described here are already live, not pending.

What was happening

Running scripts/generate-public-curriculum.mjs produced output smaller than what was committed, in two places, so any build silently discarded content.

1. llms.txt lost the entire Developer docs section

The committed file carried a ## Developer docs section linking /openapi.json, /api/ai, and /sitemap.xml. The generator never emitted it.

Confirmed live rather than inferred — fetching the deployed file shows 18 lines against the committed 24, and the deployed content matches the regenerated file, not the committed one:

$ curl -s https://learn.significanthobbies.com/llms.txt | wc -l
18
$ git show HEAD:public/llms.txt | wc -l
24

All three URLs return 200. /openapi.json and /api/ai are served by functions/_middleware.ts, and /sitemap.xml by the sitemap writer — none of them by this generator. So nothing else breaks when the generator forgets them; they simply stop being discoverable by the agents this file exists to serve.

Restored as a ## Developer docs section built from absolute(), covering /openapi.json and /sitemap.xml. /api/ai is deliberately not repeated — it is already listed under Start here, and the original block duplicated it.

2. index.html lost the product name from its only <h1>

The committed markup wraps the product name in a visually-hidden span before the visible tagline:

<h1 …><span style="…clip:rect(0,0,0,0)…">SWE Interview Prep</span>A complete software engineering learning map</h1>

The generator emitted the tagline alone, so a build stripped "SWE Interview Prep" out of the single <h1> on the page — costing exactly the brand findability the GEO/SEO work is meant to improve. Restored; the generator now reproduces the committed index.html byte for byte, so that drift is zero.

Verification

• Full suite: 483 passed (481 before, 2 new)
• tsc --noEmit → exit 0
• All 11 links in the generated llms.txt return 200 (checked individually)
• git diff against main for index.html is now empty

Two regression tests cover both losses, and I verified they actually catch the bug rather than assuming it: swapped in the previous generator, regenerated, and the suite failed 2 of 14 on that file; restored the fix and it passes 14/14.

Why these went unnoticed

Neither loss breaks anything that fails loudly. The links keep working, the page keeps rendering, and the tests kept passing — the only symptom is that a generated file is quietly smaller than the one in git. That is worth knowing beyond this repo: the same pattern is present in ~12 fleet repos, with anime-list losing an OpenAPI/CLI section the same way and about ten more carrying uncommitted enrichment that has never landed.

🤖 Generated with Claude Code