seedTrack mints a new artist per call when there is no mbid, so the genre
fixtures held one artist per track and the candidate query's artist_ids
counted 30 where the test meant 6.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A new system playlist kind, genre_mix: up to 10 "{Genre} mix" playlists per
user, rebuilt daily with the other system mixes.
Which genres: library genres with at least 25 playable tracks across 3
artists, ranked in tiers — taste-profile genres first (matched
case-insensitively, so ID3 "Post-Punk" and folksonomy "post-punk" are one),
then genres the user has played or liked, then the rest of the library, which
only tops a short shelf up to 4. A genre whose artist set is a near
duplicate of one already picked (Jaccard >= 0.5) is skipped, so "Lo-Fi" and
"Lo-Fi Hip Hop" don't both appear but "Alternative Rock" survives "Rock". A
genre that can't fill 15 tracks gives way to the next.
Each mix is 50 tracks: 3 in 5 the user already knows (played through or
liked), the rest unheard in that genre, interleaved, with the usual
per-album/per-artist caps. Tracks skipped twice and more often than played
are left out. The order changes daily and holds within a day. Picks are
stamped taste / taste_unheard, and genre_mix plays count toward rotation and
the recommendation metrics.
Migration 0102 adds genre_mix to the playlists variant constraints.
Web and Android each get a "Your genres" row in name order, hidden until a
mix exists. Kinds with several playlists per user now play by playlist id
wherever the client used to test seed_artist_id (web card, detail page
self-heal), keyed on the server's refreshable flag; on Android genre_mix
joins songs_like_artist as not refreshable. The web detail page's
"mix was rebuilt" reload now finds a Songs-like or genre mix by name instead
of calling the by-kind shuffle, which those kinds don't have.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
bvandeusenandClaude Opus 5.5