Stop filtering discovery by CEFR level

[mircealungu]

Why

A learner reported a feed with too few articles. It is not the recommender — it is the level filter, which on-demand simplification (#698) turned into a starvation filter.

available_cefr_levels lists the levels for which a version of an article was already generated: the original's own level, plus its simplified children. Once crawl-time simplification was retired there are no children, so the field collapsed to a single value — and a learner saw only the articles a publisher happened to write at exactly their band.

Measured on the live index, Danish, last 7 days:

candidate pool for an A2 reader	documents
with the filter (A2, A1/A2)	33
without it	479

And 266 of those articles already carry an A2 ArticleLevelSummary — content generated for precisely that reader, which the filter guaranteed they would never see. The level-summary overlay and the level filter have been contradicting each other since #698.

What

The premise behind the filter no longer holds: an article above the learner's level is not something to hide when the card carries a level-appropriate summary and the body simplifies on request. So the filter is dropped, not widened — from all three discovery paths (recommender feed, video recommendations, search). Above-level readers are the mirror case and get the same treatment once we complexify.

Two further empty-feed causes go with it:

• Articles whose LLM assessment failed have no cefr_level, so they matched nothing. Every assessment outage — the Aug-13 article_type bug, the Anthropic usage-cap outages — emptied feeds outright. Feeds no longer depend on assessment liveness. (During the Aug 16–17 cap outage, every article crawled for ~12h was invisible in every language.)
• document_from_video never writes an available_cefr_levels field, so requiring a level in it excluded every video from every leveled user.

feed_delivery_health_check reached its per-segment inventory through the recommender builder, so this change would have silently degraded it into six copies of one language-wide query. It now appends the CEFR clause itself, and the docstrings state what that clause measures now: assessment coverage, not feed delivery.

Notes

• No reindex needed. The field stays indexed; we just stop filtering on it. Effective on deploy.
• UserPreference.SHOW_EASIER_ARTICLES is now inert — there is nothing left to widen. Left in place and flagged; the web toggle in Settings → Feed Preferences should be removed or repurposed in a follow-up.
• Videos will now appear in feeds for the first time. Easy to revert separately if that should wait.
• Separately spotted, not fixed here: User.cefr_level_for_language does [...][cefr_level - 1], so a stored cefr_level of 0 maps to C2. Many users have 0. Moot for the feed now, but wrong wherever else it is called.

Tests

• New zeeguu/core/test/test_elastic_query_builder.py pins the no-filter behavior on all three queries so it cannot creep back.
• zeeguu/core/test: 226 passed.
• zeeguu/api/test: 16 failures, all pre-existing on clean master (test_bookmark_positions, test_verbal_flashcards) — verified by stashing.

🤖 Generated with Claude Code

[github-actions]

ArchLens detected architectural changes in the following views: