Attempting to delete another user's key returns 404 (not 403) to avoid leaking existence.

paulgraham.com is in the HN top domains list; domain_score should be > 0.

Test domain scoring and URL management in Redis

Results are indexed at the end and the count is reported in 'done'.

Test edge cases with empty strings and malformed URLs in the list

Test error recovery and system resilience

Rust and Python feature names should be in the same order.

Rust and Python feature vectors should agree to within ATOL.

The final ranking applies MMR diversity (as standard search does): a second same-domain result is demoted below a lower-scored fresh-domain result

A 'results' event with the full ranked list is emitted before 'done'.

A followed link with a genuine crawled title but an empty extract is kept.

A followed link whose crawl yields no real title is a pseudo-result (its only title would be URL-derived) and must not enter the final ranking,

Test getting user's voting history

Test getting vote counts for multiple URLs

Test getting vote counts without providing URLs

Test getting vote counts for a single URL

Test HaGeZiBlacklistProvider rejects invalid list types.

Test HaGeZiBlacklistProvider with successful HTTP response.

A doc whose score equals the heap minimum must NOT be promoted — the check is strictly greater-than, so ties don't displace existing entries.

When the heap is full, a later doc with a higher score replaces the minimum and is still promoted — the heap min-replacement path must work.

Test the indexing workflow with fake Redis

Test integration with the main blacklist module.

Test the complete integration workflow

Test handling a large number of URLs (more than 160) that would fail with GET

A promoted page is crawled, its outbound links followed, and the followed docs appear in the final ranking — exercises page_fetched / link_followe

A malformed recipe is logged and skipped, not raised — so one bad file can't crash the whole app at import time. Valid recipes still load.

Test that a user can submit multiple WASM files

After training, NDCG on training data should be above 0.5 (better than random).

The dedup guard must never emit two consecutive identical `results` frames.

No feature should be NaN for any test case.

NUM_FEATURES constant should be 50.

Test that the unified OpenAPI docs page loads without errors

Test that the unified OpenAPI spec can be generated without errors

When the pipeline overruns the deadline, the stream still terminates with a `done` event whose reason is `timed_out`.

Body api_key field is deprecated but must still work for backward compatibility.

When both header and body key are present, the header key is used.

A search-scoped key must not grant access to the crawler results endpoint.