feat(server): "You might like" album/artist Home rows (#790)

Surface in-library albums/artists the listener doesn't actively spin but
is predicted to enjoy, derived from the same similarity + like-weighted
candidate engine that powers For-You — rolled up from track scores to
album/artist granularity. Built in the daily 3am BuildSystemPlaylists
pass, atomic-replaced alongside the system playlists, and read back by
/api/home (+ /api/home/index).

Cold-start gate: skips generation entirely below 20 distinct unskipped
tracks AND 5 distinct artists, so a thin profile ships empty rows rather
than near-random tiles.

- migration 0034: you_might_like_albums / you_might_like_artists (id+rank,
  CASCADE, per-user rank index).
- playlists/you_might_like.go: cold-start gate + similarity roll-up
  (sum-of-top-3 aggregation, per-artist album cap, daily-rotating via the
  same userIDHash jitter as For-You) + atomic-replace persist in the tx.
- recommendation/home.go: two new HomePayload sections with read-time
  cross-section dedup vs Most Played / Rediscover / Last Played, trimmed
  to 10 each.
- api: you_might_like_albums / you_might_like_artists on /api/home and
  /api/home/index, reusing albumRefFrom / artistRefFromCovered.
- tests: pure roll-up/aggregation/cap unit tests + DB-backed gate,
  sufficiency, and atomic-replace tests (all green vs real Postgres).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

internal/playlists/system.go

@@ -497,6 +497,13 @@ func BuildSystemPlaylists(ctx context.Context, pool *pgxpool.Pool, logger *slog.
 		built = append(built, out...)
 	}
 
+	// "You might like" Home rows reuse the same similarity engine but
+	// persist to dedicated album/artist tables (not playlist_tracks).
+	// Computed here (reads only) and atomic-replaced inside the tx below,
+	// alongside the system playlists. A failed/gated computation is
+	// handled by yml.built (leave prior rows vs. clear) — never fatal.
+	yml := buildYouMightLike(ctx, q, logger, userID, dateStr, now)
+
 	// Atomic replace inside a transaction.
 	tx, err := pool.Begin(ctx)
 	if err != nil {
@@ -530,6 +537,13 @@ func BuildSystemPlaylists(ctx context.Context, pool *pgxpool.Pool, logger *slog.
 		createdIDs = append(createdIDs, id)
 	}
 
+	if yml.built {
+		if err := persistYouMightLike(ctx, qtx, userID, yml); err != nil {
+			buildErr = fmt.Errorf("persist you-might-like: %w", err)
+			return buildErr
+		}
+	}
+
 	if err := tx.Commit(ctx); err != nil {
 		buildErr = fmt.Errorf("commit tx: %w", err)
 		return buildErr

internal/playlists/you_might_like.go

@@ -0,0 +1,227 @@
+// you_might_like.go builds the per-user "You might like" Home rows
+// (#790): in-library albums/artists the listener doesn't actively spin
+// but is predicted to enjoy. It reuses the For-You candidate engine
+// (similarity + like-weighted scoring) and rolls the per-track scores up
+// to album/artist granularity, gated on a minimum listening history so a
+// near-empty profile ships nothing rather than noise.
+//
+// Built in the same daily BuildSystemPlaylists pass and atomic-replaced
+// alongside the system playlists; read back by /api/home.
+package playlists
+
+import (
+	"context"
+	"fmt"
+	"log/slog"
+	"math/rand"
+	"sort"
+	"time"
+
+	"github.com/jackc/pgx/v5/pgtype"
+
+	"git.fabledsword.com/bvandeusen/minstrel/internal/db/dbq"
+	"git.fabledsword.com/bvandeusen/minstrel/internal/recommendation"
+)
+
+const (
+	// Cold-start gate. Below this much *real* listening the similarity
+	// roll-up degrades toward random fill, so You-might-like ships
+	// nothing rather than arbitrary tiles. Both bars must clear: distinct
+	// tracks rules out "played 3 songs on repeat," distinct artists rules
+	// out "hammered one album." Conservative defaults; raise if early
+	// rows feel random on thin libraries.
+	youMightLikeMinDistinctTracks  = 20
+	youMightLikeMinDistinctArtists = 5
+
+	// Persisted depth — a few more than the rendered HomeYouMightLikeLimit
+	// so the read-time cross-section dedup (vs Most Played / Rediscover)
+	// has headroom before trimming.
+	youMightLikeAlbumsN  = 12
+	youMightLikeArtistsN = 12
+
+	// One strongly-matched artist shouldn't fill the albums row.
+	youMightLikeMaxAlbumsPerArtist = 2
+
+	// Aggregation: sum of an entity's top-K track scores. Favors albums/
+	// artists with several good matches over a single outlier track.
+	youMightLikeAggTopK = 3
+)
+
+// youMightLikeResult carries one day's roll-up. built=false means the
+// computation failed and the caller must leave any existing rows intact;
+// built=true with nil slices means "explicitly empty" (cold-start gate or
+// no candidates) and the caller should clear the user's rows.
+type youMightLikeResult struct {
+	albumIDs  []pgtype.UUID
+	artistIDs []pgtype.UUID
+	built     bool
+}
+
+// buildYouMightLike computes the user's daily album/artist rolls. Reads
+// only — the caller persists inside the build tx. Never returns an error:
+// a failure logs and yields built=false so the prior day's rows survive.
+func buildYouMightLike(
+	ctx context.Context, q *dbq.Queries, logger *slog.Logger,
+	userID pgtype.UUID, dateStr string, now time.Time,
+) youMightLikeResult {
+	signal, err := q.CountListeningSignalForUser(ctx, userID)
+	if err != nil {
+		logger.Warn("you-might-like: listening-signal query failed; skipping",
+			"user_id", uuidStringPL(userID), "err", err)
+		return youMightLikeResult{built: false}
+	}
+	if signal.DistinctTracks < youMightLikeMinDistinctTracks ||
+		signal.DistinctArtists < youMightLikeMinDistinctArtists {
+		// Cold start: not enough listening to recommend from. built=true
+		// clears any stale rows (normally none) and ships an empty row.
+		return youMightLikeResult{built: true}
+	}
+
+	seeds, err := q.PickTopPlayedTracksForUser(ctx, userID)
+	if err != nil {
+		logger.Warn("you-might-like: seed query failed; skipping",
+			"user_id", uuidStringPL(userID), "err", err)
+		return youMightLikeResult{built: false}
+	}
+	seed := pickForYouSeedForDay(seeds, userID, dateStr)
+	if !seed.Valid {
+		return youMightLikeResult{built: true}
+	}
+
+	zeroVec := recommendation.SessionVector{Seed: true}
+	cands, err := recommendation.LoadCandidatesFromSimilarity(
+		ctx, q, userID, seed, 1, zeroVec,
+		[]pgtype.UUID{seed}, systemForYouSourceLimits(),
+	)
+	if err != nil {
+		logger.Warn("you-might-like: candidate load failed; skipping",
+			"user_id", uuidStringPL(userID), "err", err)
+		return youMightLikeResult{built: false}
+	}
+
+	albumIDs, artistIDs := rollUpCandidates(cands, userID, dateStr, now)
+	return youMightLikeResult{albumIDs: albumIDs, artistIDs: artistIDs, built: true}
+}
+
+// scoredEntity is an album or artist with its aggregated score.
+type scoredEntity struct {
+	id    pgtype.UUID
+	score float64
+}
+
+// rollUpCandidates scores every candidate track (systemMixWeights, jitter
+// seeded by userIDHash so near-ties rotate day-to-day) and aggregates the
+// scores up to album and artist via sum-of-top-K. Returns the top-N album
+// and artist IDs, with a per-artist cap on the album list.
+func rollUpCandidates(
+	cands []recommendation.Candidate, userID pgtype.UUID, dateStr string, now time.Time,
+) (albumIDs, artistIDs []pgtype.UUID) {
+	rng := rand.New(rand.NewSource(int64(userIDHash(userID, dateStr))))
+	albumScores := map[pgtype.UUID][]float64{}
+	artistScores := map[pgtype.UUID][]float64{}
+	albumArtist := map[pgtype.UUID]pgtype.UUID{}
+	for _, c := range cands {
+		s := recommendation.Score(c.Inputs, systemMixWeights, now, rng.Float64)
+		if c.Track.AlbumID.Valid {
+			albumScores[c.Track.AlbumID] = append(albumScores[c.Track.AlbumID], s)
+			albumArtist[c.Track.AlbumID] = c.Track.ArtistID
+		}
+		if c.Track.ArtistID.Valid {
+			artistScores[c.Track.ArtistID] = append(artistScores[c.Track.ArtistID], s)
+		}
+	}
+
+	rankedAlbums := capAlbumsPerArtist(
+		rankEntities(albumScores, dateStr), albumArtist, youMightLikeMaxAlbumsPerArtist)
+	albumIDs = topNEntityIDs(rankedAlbums, youMightLikeAlbumsN)
+	artistIDs = topNEntityIDs(rankEntities(artistScores, dateStr), youMightLikeArtistsN)
+	return albumIDs, artistIDs
+}
+
+// rankEntities aggregates each entity's track scores (sum-of-top-K) and
+// returns them sorted by score DESC, ties broken deterministically by
+// tieBreakHash(id, dateStr) so ordering is stable within a day.
+func rankEntities(scoresByEntity map[pgtype.UUID][]float64, dateStr string) []scoredEntity {
+	out := make([]scoredEntity, 0, len(scoresByEntity))
+	for id, scores := range scoresByEntity {
+		out = append(out, scoredEntity{id: id, score: sumTopK(scores, youMightLikeAggTopK)})
+	}
+	sort.SliceStable(out, func(i, j int) bool {
+		if out[i].score != out[j].score {
+			return out[i].score > out[j].score
+		}
+		return tieBreakHash(out[i].id, dateStr) < tieBreakHash(out[j].id, dateStr)
+	})
+	return out
+}
+
+// sumTopK sorts scores DESC and sums the highest k.
+func sumTopK(scores []float64, k int) float64 {
+	sort.Sort(sort.Reverse(sort.Float64Slice(scores)))
+	sum := 0.0
+	for i := 0; i < len(scores) && i < k; i++ {
+		sum += scores[i]
+	}
+	return sum
+}
+
+// capAlbumsPerArtist drops albums beyond maxPerArtist for any one artist,
+// preserving input (score) order.
+func capAlbumsPerArtist(
+	albums []scoredEntity, albumArtist map[pgtype.UUID]pgtype.UUID, maxPerArtist int,
+) []scoredEntity {
+	perArtist := map[pgtype.UUID]int{}
+	out := make([]scoredEntity, 0, len(albums))
+	for _, a := range albums {
+		art := albumArtist[a.id]
+		if art.Valid {
+			if perArtist[art] >= maxPerArtist {
+				continue
+			}
+			perArtist[art]++
+		}
+		out = append(out, a)
+	}
+	return out
+}
+
+// topNEntityIDs truncates to n and projects the IDs in rank order.
+func topNEntityIDs(entities []scoredEntity, n int) []pgtype.UUID {
+	if len(entities) > n {
+		entities = entities[:n]
+	}
+	out := make([]pgtype.UUID, 0, len(entities))
+	for _, e := range entities {
+		out = append(out, e.id)
+	}
+	return out
+}
+
+// persistYouMightLike atomic-replaces the user's You-might-like rows
+// inside the build tx. Called only when the roll-up was freshly built;
+// a failed computation leaves the prior rows untouched.
+func persistYouMightLike(
+	ctx context.Context, qtx *dbq.Queries, userID pgtype.UUID, r youMightLikeResult,
+) error {
+	if err := qtx.DeleteYouMightLikeAlbumsForUser(ctx, userID); err != nil {
+		return fmt.Errorf("delete you-might-like albums: %w", err)
+	}
+	for i, id := range r.albumIDs {
+		if err := qtx.InsertYouMightLikeAlbum(ctx, dbq.InsertYouMightLikeAlbumParams{
+			UserID: userID, AlbumID: id, Rank: int32(i),
+		}); err != nil {
+			return fmt.Errorf("insert you-might-like album: %w", err)
+		}
+	}
+	if err := qtx.DeleteYouMightLikeArtistsForUser(ctx, userID); err != nil {
+		return fmt.Errorf("delete you-might-like artists: %w", err)
+	}
+	for i, id := range r.artistIDs {
+		if err := qtx.InsertYouMightLikeArtist(ctx, dbq.InsertYouMightLikeArtistParams{
+			UserID: userID, ArtistID: id, Rank: int32(i),
+		}); err != nil {
+			return fmt.Errorf("insert you-might-like artist: %w", err)
+		}
+	}
+	return nil
+}

internal/playlists/you_might_like_db_test.go

@@ -0,0 +1,99 @@
+package playlists_test
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"github.com/jackc/pgx/v5/pgtype"
+	"github.com/jackc/pgx/v5/pgxpool"
+
+	"git.fabledsword.com/bvandeusen/minstrel/internal/playlists"
+)
+
+// countYouMightLike returns the persisted row counts for a user.
+func countYouMightLike(t *testing.T, pool *pgxpool.Pool, userID pgtype.UUID) (albums, artists int) {
+	t.Helper()
+	ctx := context.Background()
+	if err := pool.QueryRow(ctx,
+		`SELECT count(*) FROM you_might_like_albums WHERE user_id = $1`, userID,
+	).Scan(&albums); err != nil {
+		t.Fatalf("count albums: %v", err)
+	}
+	if err := pool.QueryRow(ctx,
+		`SELECT count(*) FROM you_might_like_artists WHERE user_id = $1`, userID,
+	).Scan(&artists); err != nil {
+		t.Fatalf("count artists: %v", err)
+	}
+	return albums, artists
+}
+
+// TestYouMightLike_ColdStartGate: a thin profile (3 artists × 3 tracks =
+// 9 distinct tracks, below the 20-track / 5-artist gate) must produce no
+// You-might-like rows — the build ships nothing rather than near-random
+// tiles.
+func TestYouMightLike_ColdStartGate(t *testing.T) {
+	pool := newPool(t)
+	logger := discardLogger()
+	u, _ := seedActiveLibrary(t, pool, "ymlcold", 3, 3)
+
+	if err := playlists.BuildSystemPlaylists(
+		context.Background(), pool, logger, u.ID, time.Now().UTC(), t.TempDir(),
+	); err != nil {
+		t.Fatalf("build: %v", err)
+	}
+
+	albums, artists := countYouMightLike(t, pool, u.ID)
+	if albums != 0 || artists != 0 {
+		t.Errorf("cold-start user should have no you-might-like rows; got %d albums, %d artists",
+			albums, artists)
+	}
+}
+
+// TestYouMightLike_SufficientActivityPopulates: a rich profile (8 artists
+// × 4 tracks = 32 distinct tracks, clearing both gate bars) populates the
+// You-might-like tables in the daily build.
+func TestYouMightLike_SufficientActivityPopulates(t *testing.T) {
+	pool := newPool(t)
+	logger := discardLogger()
+	u, _ := seedActiveLibrary(t, pool, "ymlrich", 8, 4)
+
+	if err := playlists.BuildSystemPlaylists(
+		context.Background(), pool, logger, u.ID, time.Now().UTC(), t.TempDir(),
+	); err != nil {
+		t.Fatalf("build: %v", err)
+	}
+
+	albums, artists := countYouMightLike(t, pool, u.ID)
+	if albums == 0 {
+		t.Error("rich user should have you-might-like album rows; got 0")
+	}
+	if artists == 0 {
+		t.Error("rich user should have you-might-like artist rows; got 0")
+	}
+}
+
+// TestYouMightLike_AtomicReplace: a second build replaces rather than
+// appends, so counts stay bounded by the persisted depth.
+func TestYouMightLike_AtomicReplace(t *testing.T) {
+	pool := newPool(t)
+	logger := discardLogger()
+	u, _ := seedActiveLibrary(t, pool, "ymlreplace", 8, 4)
+	ctx := context.Background()
+
+	for i := 0; i < 2; i++ {
+		if err := playlists.BuildSystemPlaylists(
+			ctx, pool, logger, u.ID, time.Now().UTC(), t.TempDir(),
+		); err != nil {
+			t.Fatalf("build %d: %v", i, err)
+		}
+	}
+
+	albums, artists := countYouMightLike(t, pool, u.ID)
+	if albums > 12 {
+		t.Errorf("albums = %d after two builds, want <= 12 (atomic replace)", albums)
+	}
+	if artists > 12 {
+		t.Errorf("artists = %d after two builds, want <= 12 (atomic replace)", artists)
+	}
+}

internal/playlists/you_might_like_test.go

@@ -0,0 +1,125 @@
+package playlists
+
+import (
+	"testing"
+	"time"
+
+	"github.com/jackc/pgx/v5/pgtype"
+
+	"git.fabledsword.com/bvandeusen/minstrel/internal/recommendation"
+)
+
+// fixedNow keeps recencyDecay stable across the pure roll-up tests.
+var fixedNow = time.Date(2026, 6, 11, 12, 0, 0, 0, time.UTC)
+
+// byteOf returns the per-test entity discriminator (makeCand sets it in
+// Bytes[15]) so assertions can name albums/artists by their small int.
+func byteOf(u pgtype.UUID) byte { return u.Bytes[15] }
+
+// countWithByte reports how many of ids carry the given Bytes[15] marker.
+func countWithByte(ids []pgtype.UUID, want byte) int {
+	n := 0
+	for _, id := range ids {
+		if byteOf(id) == want {
+			n++
+		}
+	}
+	return n
+}
+
+func TestSumTopK(t *testing.T) {
+	if got := sumTopK([]float64{1, 2, 3, 4}, 2); got != 7 {
+		t.Errorf("sumTopK top-2 of 1..4 = %v, want 7", got)
+	}
+	if got := sumTopK([]float64{5}, 3); got != 5 {
+		t.Errorf("sumTopK fewer-than-k = %v, want 5", got)
+	}
+	if got := sumTopK(nil, 3); got != 0 {
+		t.Errorf("sumTopK empty = %v, want 0", got)
+	}
+}
+
+func TestRollUpCandidates_MultiMatchAlbumRanksFirst(t *testing.T) {
+	// Album 10 (artist 100): three solid matches. Album 20 (artist 200):
+	// one slightly-stronger single match. Sum-of-top-3 should rank the
+	// multi-match album ahead of the single-match one. Similarity gaps are
+	// wide enough that the ±0.1 jitter can't reorder the result.
+	cands := []recommendation.Candidate{
+		makeCand(1, 10, 100, 0.8),
+		makeCand(2, 10, 100, 0.8),
+		makeCand(3, 10, 100, 0.8),
+		makeCand(4, 20, 200, 0.95),
+	}
+	albums, _ := rollUpCandidates(cands, testUserID, "2026-06-11", fixedNow)
+	if len(albums) < 2 {
+		t.Fatalf("want >=2 albums, got %d", len(albums))
+	}
+	if byteOf(albums[0]) != 10 {
+		t.Errorf("top album = %d, want 10 (multi-match beats single)", byteOf(albums[0]))
+	}
+}
+
+func TestRollUpCandidates_PerArtistAlbumCap(t *testing.T) {
+	// Artist 100 spans albums 10/11/12; the album row caps any one artist
+	// at youMightLikeMaxAlbumsPerArtist (2). Artist 200's album 20 should
+	// still appear.
+	cands := []recommendation.Candidate{
+		makeCand(1, 10, 100, 0.9),
+		makeCand(2, 11, 100, 0.85),
+		makeCand(3, 12, 100, 0.8),
+		makeCand(4, 20, 200, 0.7),
+	}
+	albums, _ := rollUpCandidates(cands, testUserID, "2026-06-11", fixedNow)
+	from100 := countWithByte(albums, 10) + countWithByte(albums, 11) + countWithByte(albums, 12)
+	if from100 > youMightLikeMaxAlbumsPerArtist {
+		t.Errorf("artist 100 contributed %d albums, want <= %d",
+			from100, youMightLikeMaxAlbumsPerArtist)
+	}
+	if countWithByte(albums, 20) != 1 {
+		t.Errorf("album 20 (artist 200) should survive the cap; got %d",
+			countWithByte(albums, 20))
+	}
+}
+
+func TestRollUpCandidates_ArtistRollupDistinct(t *testing.T) {
+	// Three artists; the artist roll-up should surface all three distinct
+	// artist IDs (no per-artist cap on the artist row).
+	cands := []recommendation.Candidate{
+		makeCand(1, 10, 100, 0.9),
+		makeCand(2, 11, 101, 0.6),
+		makeCand(3, 12, 102, 0.3),
+	}
+	_, artists := rollUpCandidates(cands, testUserID, "2026-06-11", fixedNow)
+	seen := map[byte]bool{}
+	for _, a := range artists {
+		seen[byteOf(a)] = true
+	}
+	for _, want := range []byte{100, 101, 102} {
+		if !seen[want] {
+			t.Errorf("artist %d missing from roll-up", want)
+		}
+	}
+}
+
+func TestCapAlbumsPerArtist_DropsBeyondCap(t *testing.T) {
+	mk := func(albumN, artistN int) (pgtype.UUID, pgtype.UUID) {
+		var al, ar pgtype.UUID
+		al.Valid, ar.Valid = true, true
+		al.Bytes[15], ar.Bytes[15] = byte(albumN), byte(artistN)
+		return al, ar
+	}
+	a10, ar1 := mk(10, 1)
+	a11, _ := mk(11, 1)
+	a12, _ := mk(12, 1)
+	a20, ar2 := mk(20, 2)
+	albumArtist := map[pgtype.UUID]pgtype.UUID{a10: ar1, a11: ar1, a12: ar1, a20: ar2}
+	in := []scoredEntity{{a10, 3}, {a11, 2}, {a12, 1}, {a20, 0.5}}
+	got := capAlbumsPerArtist(in, albumArtist, 2)
+	if len(got) != 3 {
+		t.Fatalf("len = %d, want 3 (artist 1 capped at 2 + artist 2's one)", len(got))
+	}
+	// Highest-scored two of artist 1 (albums 10, 11) kept; 12 dropped.
+	if byteOf(got[2].id) != 20 {
+		t.Errorf("third survivor = %d, want 20", byteOf(got[2].id))
+	}
+}