Run LLM cleanup on original article content before storing

[mircealungu]

Summary

• Adds cleanup_article_content() on SimplificationService — a Haiku pass that strips leftover scraper noise (audio/share widgets, "Listen to the article" buttons, related-article link lists, newsletter blurbs, ad text, cookie banners) from the body returned by the readability server.
• Wires it into Article.find_or_create() between language detection and Source.find_or_create(), so both the upload-with-text path and the readability path get cleaned. The cleaned versions are what land in Source.content and Article.htmlContent, so the reader sees a clean article on first open.
• Synchronous on the send-to-Zeeguu path. Latency is acceptable; junk on first read isn't.

Why

With the reading-flow flip toward send-to-Zeeguu as the default and full-article-externally being torn out, the original-content view becomes the primary read path — so the leftover noise that readability + the regex cleanup_non_content_bits() miss is now what every learner sees. A real example that prompted this: an article opened with "Listen to the article" appearing inline in the body because the source page had a TTS-button caption that readability kept.

Design notes

• Fail-soft: any LLM error returns None and the original content is kept. The ingestion path never breaks on a flaky API.
• Guardrails to avoid hallucination/over-deletion:
  • skip bodies under 500 chars (not worth the latency, rarely have junk)
  • skip if ANTHROPIC_TEXT_SIMPLIFICATION_KEY is unset
  • reject the result if cleaned text is >110% or <50% of original length
• Prompt is conservative: model is told to remove only clearly non-article fragments, not to translate / simplify / summarize / reorder.
• Not addressed in this PR (deliberately, to keep scope small):
  • French crawl-path bundling (cleanup-then-simplify in one LLM call for the existing crawl-time pipeline)
  • Completeness signal + filter-from-recommendations for articles that rely on missing interactive content

Test plan

• Send a noisy article (e.g. one with a "Listen to the article" button visible in the body) via the extension/share flow and confirm the reader view no longer shows the widget text
• Send an article on a domain that already cleans well — confirm no over-deletion (no missing paragraphs) and that the guardrail keeps the original when cleanup goes wrong
• Send a very short article (<500 chars body) and confirm cleanup is skipped (no extra latency)
• With ANTHROPIC_TEXT_SIMPLIFICATION_KEY unset, confirm ingestion still works and the original content is kept
• Watch logs for Anthropic cleanup ... lines on staging before merging

🤖 Generated with Claude Code

[github-actions]

ArchLens detected architectural changes in the following views: