Broaden disturbing-content filter to catch death/burial themes

[mircealungu]

Why

A Romanian "coffin therapy" article — a lifestyle/novelty piece about being buried alive in a sealed coffin for relaxation (tagged Business, from adevarul.ro) — passed the "Avoid disturbing news (violence, death, disasters)" filter and showed up for a user who had it enabled.

Root cause: the filter's notion of "death" is death-as-a-news-event (killings, war, disasters, fatal accidents), not death-as-a-topic-or-image. The article is morbid by subject but not a violent/tragic event, so it tripped neither detection gate:

• Keyword pass — the per-language lists are all hard-news death/violence vocabulary (ucis, cadavru, atentat, prăbușire…). None of the article's actual words (sicriu, coșciug, îngropați, terapie) were present.
• LLM pass — the prompt asked whether the primary focus is a current disturbing event, so a novelty wellness story reasonably came back NO.

Plus a structural gap: the keyword flag was only persisted inside the post-simplification block, so a capped/skipped/errored simplification silently dropped it.

Changes

1. Keyword detection (disturbing_content_detection.py) — new "Death imagery & morbid themes" category across all 9 languages (coffin / burial / corpse / morgue / exhume / buried-alive). Intentionally broader than the event-based categories and accepts some false positives — the checkbox promises to hide "death", not only tragedies. Romanian now matches sicri(u) / coșciug, which catches the reported article by title alone.

2. LLM prompt (article_simplification.py) — the DISTURBING_CONTENT question now flags death/burial imagery and themes (corpses, coffins, being buried, funerary/mortuary practices, medical gore) in addition to tragic events, with an explicit example that a novelty/wellness piece whose subject is death/burial is YES. Incidental mentions and purely historical/educational treatment remain NO.

3. Decouple classification from simplification (article_downloader.py) — persist the keyword-based DISTURBING flag immediately after detection, before the daily simplification-cap early-return, so the flag is set even when simplification is skipped/capped/errored. Extracted a _save_classifications helper; find_or_create keeps the later LLM save idempotent.

Verification

RO article:        (True, 'Disturbing content keywords: sicri, coșciug')
EN coffin:         (True, 'Disturbing content keywords: buried alive, coffin')
Neutral RO control:(False, '')
Neutral EN control:(False, '')

Notes

• Affects newly crawled articles only — already-indexed articles won't be reflagged without a backfill. Happy to add a one-off tool to re-run keyword detection over recent articles if we want the fix applied retroactively.
• No existing tests cover this detection; can add a unit test for the keyword lists in a follow-up if desired.

🤖 Generated with Claude Code

[github-actions]

ArchLens - No architecturally relevant changes to the existing views