fix: decode HTML entities in scraped trending descriptions

[DamengRandom]

Closes #38

parseTrendingHtml stripped tags out of the scraped <p> but left HTML entities encoded, so a repo description GitHub renders as Octadesk & Intercom was stored and prompted as Octadesk & Intercom. Three of the 32 repos on today's trending page were affected.

Change

src/tools/trending-scrape.tool.ts — decode entities after the tags come out, so the tag-strip is not re-applied to a decoded <Suspense>. Named entities GitHub emits (& < > " '  ) plus decimal and hex numeric references; anything unrecognised is left exactly as it was, so AT&T and &notanentity; survive intact. parseTrendingHtml is now exported so the parser can be tested directly, matching how fitRepos, toAiNewsItems and mergeSummaries are already exposed for their tests.

Test

src/tools/trending-scrape.test.ts — new, built on an article block trimmed from a real github.com/trending/typescript?since=daily response. Three of the five tests fail without the fix:

not ok 1 - decodes the escaped ampersand GitHub emits in repo descriptions
  expected: 'Open alternative to Kommo, Octadesk & Intercom'
  actual: 'Open alternative to Kommo, Octadesk & Intercom'
not ok 2 - decodes the rest of the entities a description can carry
  expected: 'Type-safe <div> helpers'
  actual: 'Type-safe &lt;div&gt; helpers'
not ok 3 - a decoded angle bracket is text, not markup to strip a second time
  expected: 'Render <Suspense> on the server'
  actual: 'Render &lt;Suspense&gt; on the server'
ok 4 - leaves something that only looks like an entity untouched
ok 5 - still reads the name, language and both star counts
# tests 5
# pass 2
# fail 3

The other two pass either way on purpose — they guard against the fix over-reaching or disturbing the rest of the parse.

With the fix, the full suite, the type-check and the format check all pass:

# tests 84
# pass 84
# fail 0

Observed vs reasoned

• Observed by execution: the corrupted descriptions, the raw page HTML they came from, the curator prompt line built from them, the three failing assertions above, and pnpm test / pnpm tsc / pnpm format:check after the fix.
• Reasoned only: that the corrupted value reaches Postgres. There is no .env in this environment, so no database was written to — saveTrendingRepos passes repo.description through as a bind parameter with no transformation, which is where the claim comes from.

---

Generated by Claude Code