feat(recommendation): For You exploration attribution — taste vs fresh picks

Milestone #127 step 2 (#1249). For You deliberately blends two
populations — a head of top-scored taste picks and a tail sampled from
deeper ranking (the freshness injection) — but the metrics judged it as
one blob, so its skip rate couldn't distinguish "the taste engine is
missing" from "the freshness tax is too high". That number decides the
exploration share before we tune it.

- Migration 0038: nullable pick_kind ('taste'|'fresh') on both
  playlist_tracks (stamped at snapshot build) and play_events (frozen at
  play-ingestion — the snapshot rebuilds daily, so attribution cannot be
  reconstructed at read time).
- Builder: pickHeadAndTail marks head=taste / tail=fresh; the small-pool
  fallback is all taste (top-N-by-score IS the taste mechanism). Other
  variants persist NULL.
- Ingestion: for_you plays (live + offline replay) look the track up in
  the user's current snapshot; not found → unattributed, never guessed.
- Metrics: For You's row gains a breakdown (taste / fresh / earlier
  unattributed plays), parent row stays the sum; web card renders the
  sub-rows indented with the same baseline deltas + low-data dimming.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TsF3cNoKrqCYsU78cXC8U6

internal/db/dbq/events.sql.go

@@ -31,6 +31,33 @@ func (q *Queries) GetCurrentSessionVectorForUser(ctx context.Context, userID pgt
 	return session_vector_at_play, err
 }
 
+const getForYouPickKindForTrack = `-- name: GetForYouPickKindForTrack :one
+SELECT pt.pick_kind
+FROM playlist_tracks pt
+JOIN playlists p ON p.id = pt.playlist_id
+WHERE p.user_id = $1
+  AND p.system_variant = 'for_you'
+  AND pt.track_id = $2
+LIMIT 1
+`
+
+type GetForYouPickKindForTrackParams struct {
+	UserID  pgtype.UUID
+	TrackID pgtype.UUID
+}
+
+// Looks a track up in the user's CURRENT For You snapshot and returns
+// its pick_kind ('taste'/'fresh'). Used at play-ingestion time to
+// freeze exploration attribution onto the play_event — the snapshot is
+// rebuilt daily, so attribution can't be reconstructed at read time
+// (#1249). No row = track not in today's snapshot (caller stores NULL).
+func (q *Queries) GetForYouPickKindForTrack(ctx context.Context, arg GetForYouPickKindForTrackParams) (*string, error) {
+	row := q.db.QueryRow(ctx, getForYouPickKindForTrack, arg.UserID, arg.TrackID)
+	var pick_kind *string
+	err := row.Scan(&pick_kind)
+	return pick_kind, err
+}
+
 const getMostRecentPlaySessionForUser = `-- name: GetMostRecentPlaySessionForUser :one
 SELECT id, user_id, started_at, ended_at, last_event_at, track_count, client_id FROM play_sessions
 WHERE user_id = $1
@@ -54,7 +81,7 @@ func (q *Queries) GetMostRecentPlaySessionForUser(ctx context.Context, userID pg
 }
 
 const getOpenPlayEventForUser = `-- name: GetOpenPlayEventForUser :one
-SELECT id, user_id, track_id, session_id, started_at, ended_at, duration_played_ms, completion_ratio, was_skipped, client_id, session_vector_at_play, scrobbled_at, source FROM play_events
+SELECT id, user_id, track_id, session_id, started_at, ended_at, duration_played_ms, completion_ratio, was_skipped, client_id, session_vector_at_play, scrobbled_at, source, pick_kind FROM play_events
 WHERE user_id = $1 AND ended_at IS NULL
 ORDER BY started_at DESC
 LIMIT 1
@@ -79,12 +106,13 @@ func (q *Queries) GetOpenPlayEventForUser(ctx context.Context, userID pgtype.UUI
 		&i.SessionVectorAtPlay,
 		&i.ScrobbledAt,
 		&i.Source,
+		&i.PickKind,
 	)
 	return i, err
 }
 
 const getPlayEventByID = `-- name: GetPlayEventByID :one
-SELECT id, user_id, track_id, session_id, started_at, ended_at, duration_played_ms, completion_ratio, was_skipped, client_id, session_vector_at_play, scrobbled_at, source FROM play_events WHERE id = $1
+SELECT id, user_id, track_id, session_id, started_at, ended_at, duration_played_ms, completion_ratio, was_skipped, client_id, session_vector_at_play, scrobbled_at, source, pick_kind FROM play_events WHERE id = $1
 `
 
 func (q *Queries) GetPlayEventByID(ctx context.Context, id pgtype.UUID) (PlayEvent, error) {
@@ -104,15 +132,16 @@ func (q *Queries) GetPlayEventByID(ctx context.Context, id pgtype.UUID) (PlayEve
 		&i.SessionVectorAtPlay,
 		&i.ScrobbledAt,
 		&i.Source,
+		&i.PickKind,
 	)
 	return i, err
 }
 
 const insertPlayEvent = `-- name: InsertPlayEvent :one
 INSERT INTO play_events (
-    user_id, track_id, session_id, started_at, client_id, source
-) VALUES ($1, $2, $3, $4, $5, $6)
-RETURNING id, user_id, track_id, session_id, started_at, ended_at, duration_played_ms, completion_ratio, was_skipped, client_id, session_vector_at_play, scrobbled_at, source
+    user_id, track_id, session_id, started_at, client_id, source, pick_kind
+) VALUES ($1, $2, $3, $4, $5, $6, $7::text)
+RETURNING id, user_id, track_id, session_id, started_at, ended_at, duration_played_ms, completion_ratio, was_skipped, client_id, session_vector_at_play, scrobbled_at, source, pick_kind
 `
 
 type InsertPlayEventParams struct {
@@ -122,8 +151,11 @@ type InsertPlayEventParams struct {
 	StartedAt pgtype.Timestamptz
 	ClientID  *string
 	Source    *string
+	PickKind  *string
 }
 
+// pick_kind is non-NULL only for source='for_you' plays whose track was
+// found in the user's live For You snapshot at ingestion time (#1249).
 func (q *Queries) InsertPlayEvent(ctx context.Context, arg InsertPlayEventParams) (PlayEvent, error) {
 	row := q.db.QueryRow(ctx, insertPlayEvent,
 		arg.UserID,
@@ -132,6 +164,7 @@ func (q *Queries) InsertPlayEvent(ctx context.Context, arg InsertPlayEventParams
 		arg.StartedAt,
 		arg.ClientID,
 		arg.Source,
+		arg.PickKind,
 	)
 	var i PlayEvent
 	err := row.Scan(
@@ -148,6 +181,7 @@ func (q *Queries) InsertPlayEvent(ctx context.Context, arg InsertPlayEventParams
 		&i.SessionVectorAtPlay,
 		&i.ScrobbledAt,
 		&i.Source,
+		&i.PickKind,
 	)
 	return i, err
 }
@@ -291,7 +325,7 @@ SET ended_at = $2,
     completion_ratio = $4,
     was_skipped = $5
 WHERE id = $1
-RETURNING id, user_id, track_id, session_id, started_at, ended_at, duration_played_ms, completion_ratio, was_skipped, client_id, session_vector_at_play, scrobbled_at, source
+RETURNING id, user_id, track_id, session_id, started_at, ended_at, duration_played_ms, completion_ratio, was_skipped, client_id, session_vector_at_play, scrobbled_at, source, pick_kind
 `
 
 type UpdatePlayEventEndedParams struct {
@@ -328,6 +362,7 @@ func (q *Queries) UpdatePlayEventEnded(ctx context.Context, arg UpdatePlayEventE
 		&i.SessionVectorAtPlay,
 		&i.ScrobbledAt,
 		&i.Source,
+		&i.PickKind,
 	)
 	return i, err
 }

internal/db/dbq/models.go

@@ -383,6 +383,7 @@ type PlayEvent struct {
 	SessionVectorAtPlay []byte
 	ScrobbledAt         pgtype.Timestamptz
 	Source              *string
+	PickKind            *string
 }
 
 type PlaySession struct {
@@ -433,6 +434,7 @@ type PlaylistTrack struct {
 	AlbumTitle  string
 	DurationSec int32
 	AddedAt     pgtype.Timestamptz
+	PickKind    *string
 }
 
 type RegistrationSetting struct {

internal/db/dbq/playlists.sql.go

@@ -12,7 +12,7 @@ import (
 )
 
 const appendPlaylistTrack = `-- name: AppendPlaylistTrack :one
-INSERT INTO playlist_tracks (playlist_id, position, track_id, title, artist_name, album_title, duration_sec)
+INSERT INTO playlist_tracks (playlist_id, position, track_id, title, artist_name, album_title, duration_sec, pick_kind)
 SELECT
     $1::uuid,
     COALESCE((SELECT MAX(position) + 1 FROM playlist_tracks WHERE playlist_id = $1::uuid), 0),
@@ -20,24 +20,27 @@ SELECT
     t.title,
     artists.name,
     albums.title,
-    (t.duration_ms / 1000)::integer
+    (t.duration_ms / 1000)::integer,
+    $2::text
 FROM tracks t
 JOIN albums  ON albums.id  = t.album_id
 JOIN artists ON artists.id = t.artist_id
-WHERE t.id = $2::uuid
-RETURNING playlist_id, position, track_id, title, artist_name, album_title, duration_sec, added_at
+WHERE t.id = $3::uuid
+RETURNING playlist_id, position, track_id, title, artist_name, album_title, duration_sec, added_at, pick_kind
 `
 
 type AppendPlaylistTrackParams struct {
 	PlaylistID pgtype.UUID
+	PickKind   *string
 	TrackID    pgtype.UUID
 }
 
 // Inserts at the next available position. Snapshot fields are copied
 // from the tracks/albums/artists join at insert time. tracks.duration_ms
-// is converted to seconds for the snapshot.
+// is converted to seconds for the snapshot. pick_kind is NULL for manual
+// playlists; the For You builder stamps 'taste'/'fresh' (#1249).
 func (q *Queries) AppendPlaylistTrack(ctx context.Context, arg AppendPlaylistTrackParams) (PlaylistTrack, error) {
-	row := q.db.QueryRow(ctx, appendPlaylistTrack, arg.PlaylistID, arg.TrackID)
+	row := q.db.QueryRow(ctx, appendPlaylistTrack, arg.PlaylistID, arg.PickKind, arg.TrackID)
 	var i PlaylistTrack
 	err := row.Scan(
 		&i.PlaylistID,
@@ -48,6 +51,7 @@ func (q *Queries) AppendPlaylistTrack(ctx context.Context, arg AppendPlaylistTra
 		&i.AlbumTitle,
 		&i.DurationSec,
 		&i.AddedAt,
+		&i.PickKind,
 	)
 	return i, err
 }
@@ -257,7 +261,7 @@ func (q *Queries) ListAllPlaylistTracksForCollage(ctx context.Context, arg ListA
 }
 
 const listPlaylistTracks = `-- name: ListPlaylistTracks :many
-SELECT pt.playlist_id, pt.position, pt.track_id, pt.title, pt.artist_name, pt.album_title, pt.duration_sec, pt.added_at,
+SELECT pt.playlist_id, pt.position, pt.track_id, pt.title, pt.artist_name, pt.album_title, pt.duration_sec, pt.added_at, pt.pick_kind,
        t.id           AS live_track_id,
        albums.id      AS album_id,
        artists.id     AS artist_id
@@ -278,6 +282,7 @@ type ListPlaylistTracksRow struct {
 	AlbumTitle  string
 	DurationSec int32
 	AddedAt     pgtype.Timestamptz
+	PickKind    *string
 	LiveTrackID pgtype.UUID
 	AlbumID     pgtype.UUID
 	ArtistID    pgtype.UUID
@@ -305,6 +310,7 @@ func (q *Queries) ListPlaylistTracks(ctx context.Context, playlistID pgtype.UUID
 			&i.AlbumTitle,
 			&i.DurationSec,
 			&i.AddedAt,
+			&i.PickKind,
 			&i.LiveTrackID,
 			&i.AlbumID,
 			&i.ArtistID,

internal/db/dbq/recommendation_metrics.sql.go

@@ -15,6 +15,7 @@ const recommendationSourceMetricsForUser = `-- name: RecommendationSourceMetrics
 
 SELECT
     pe.source,
+    pe.pick_kind,
     count(*)::bigint AS plays,
     count(*) FILTER (WHERE pe.was_skipped)::bigint AS skips,
     count(pe.completion_ratio)::bigint AS completion_n,
@@ -22,7 +23,7 @@ SELECT
 FROM play_events pe
 WHERE pe.user_id = $1
   AND pe.started_at > now() - ($2::float8 * INTERVAL '1 day')
-GROUP BY pe.source
+GROUP BY pe.source, pe.pick_kind
 ORDER BY plays DESC
 `
 
@@ -33,6 +34,7 @@ type RecommendationSourceMetricsForUserParams struct {
 
 type RecommendationSourceMetricsForUserRow struct {
 	Source        *string
+	PickKind      *string
 	Plays         int64
 	Skips         int64
 	CompletionN   int64
@@ -50,6 +52,8 @@ type RecommendationSourceMetricsForUserRow struct {
 // avg_completion correctly.
 // $1 user_id, $2 window_days. plays/skips are counts; avg_completion is the
 // mean completion ratio over the completion_n plays that recorded one.
+// pick_kind splits For You plays into taste/fresh/unattributed (#1249);
+// it is NULL for every other source, so those still group to one row.
 func (q *Queries) RecommendationSourceMetricsForUser(ctx context.Context, arg RecommendationSourceMetricsForUserParams) ([]RecommendationSourceMetricsForUserRow, error) {
 	rows, err := q.db.Query(ctx, recommendationSourceMetricsForUser, arg.UserID, arg.Column2)
 	if err != nil {
@@ -61,6 +65,7 @@ func (q *Queries) RecommendationSourceMetricsForUser(ctx context.Context, arg Re
 		var i RecommendationSourceMetricsForUserRow
 		if err := rows.Scan(
 			&i.Source,
+			&i.PickKind,
 			&i.Plays,
 			&i.Skips,
 			&i.CompletionN,

internal/db/migrations/0038_for_you_pick_kind.down.sql

@@ -0,0 +1,2 @@
+ALTER TABLE play_events DROP COLUMN pick_kind;
+ALTER TABLE playlist_tracks DROP COLUMN pick_kind;

internal/db/migrations/0038_for_you_pick_kind.up.sql

@@ -0,0 +1,34 @@
+-- For You exploration attribution (milestone #127 step 2, Scribe #1249).
+--
+-- For You is deliberately composed of two populations: a head of
+-- top-scored similarity matches ("taste" picks) and a tail sampled from
+-- deeper in the candidate ranking ("fresh" picks — the freshness
+-- injection). Until now the metrics could only judge For You as one
+-- blob, so a high skip rate couldn't be read as "the taste engine is
+-- missing" vs "the freshness tax is too high".
+--
+-- Two denormalized columns, both nullable:
+--
+-- 1. playlist_tracks.pick_kind — stamped by the system-playlist builder
+--    at snapshot build time. NULL for manual playlists and for system
+--    variants that don't split (Discover, Songs like X, discovery
+--    mixes). System snapshots are atomically replaced daily, so the
+--    column never goes stale within a snapshot.
+--
+-- 2. play_events.pick_kind — stamped at play-ingestion time by looking
+--    the track up in the user's CURRENT For You snapshot. Attribution
+--    must be frozen at play time: For You rebuilds daily, so a 30-day
+--    metrics window spans ~30 snapshots and a join-at-read against the
+--    live snapshot would misattribute nearly everything. Plays that
+--    predate this feature (or whose track already rotated out, e.g. a
+--    late offline replay) stay NULL → reported as unattributed.
+
+ALTER TABLE playlist_tracks
+    ADD COLUMN pick_kind text
+    CONSTRAINT playlist_tracks_pick_kind_check
+        CHECK (pick_kind IN ('taste', 'fresh'));
+
+ALTER TABLE play_events
+    ADD COLUMN pick_kind text
+    CONSTRAINT play_events_pick_kind_check
+        CHECK (pick_kind IN ('taste', 'fresh'));

internal/db/queries/events.sql

@@ -24,11 +24,27 @@ ORDER BY started_at DESC
 LIMIT 1;
 
 -- name: InsertPlayEvent :one
+-- pick_kind is non-NULL only for source='for_you' plays whose track was
+-- found in the user's live For You snapshot at ingestion time (#1249).
 INSERT INTO play_events (
-    user_id, track_id, session_id, started_at, client_id, source
-) VALUES ($1, $2, $3, $4, $5, $6)
+    user_id, track_id, session_id, started_at, client_id, source, pick_kind
+) VALUES ($1, $2, $3, $4, $5, $6, sqlc.narg(pick_kind)::text)
 RETURNING *;
 
+-- name: GetForYouPickKindForTrack :one
+-- Looks a track up in the user's CURRENT For You snapshot and returns
+-- its pick_kind ('taste'/'fresh'). Used at play-ingestion time to
+-- freeze exploration attribution onto the play_event — the snapshot is
+-- rebuilt daily, so attribution can't be reconstructed at read time
+-- (#1249). No row = track not in today's snapshot (caller stores NULL).
+SELECT pt.pick_kind
+FROM playlist_tracks pt
+JOIN playlists p ON p.id = pt.playlist_id
+WHERE p.user_id = $1
+  AND p.system_variant = 'for_you'
+  AND pt.track_id = $2
+LIMIT 1;
+
 -- name: UpdatePlayEventEnded :one
 -- Closes a play_event by id with the given ended_at, duration, and skip flag.
 -- completion_ratio is computed from duration_played_ms and the track duration

internal/db/queries/playlists.sql

@@ -68,8 +68,9 @@ ORDER BY pt.position;
 -- name: AppendPlaylistTrack :one
 -- Inserts at the next available position. Snapshot fields are copied
 -- from the tracks/albums/artists join at insert time. tracks.duration_ms
--- is converted to seconds for the snapshot.
-INSERT INTO playlist_tracks (playlist_id, position, track_id, title, artist_name, album_title, duration_sec)
+-- is converted to seconds for the snapshot. pick_kind is NULL for manual
+-- playlists; the For You builder stamps 'taste'/'fresh' (#1249).
+INSERT INTO playlist_tracks (playlist_id, position, track_id, title, artist_name, album_title, duration_sec, pick_kind)
 SELECT
     sqlc.arg(playlist_id)::uuid,
     COALESCE((SELECT MAX(position) + 1 FROM playlist_tracks WHERE playlist_id = sqlc.arg(playlist_id)::uuid), 0),
@@ -77,7 +78,8 @@ SELECT
     t.title,
     artists.name,
     albums.title,
-    (t.duration_ms / 1000)::integer
+    (t.duration_ms / 1000)::integer,
+    sqlc.narg(pick_kind)::text
 FROM tracks t
 JOIN albums  ON albums.id  = t.album_id
 JOIN artists ON artists.id = t.artist_id

internal/db/queries/recommendation_metrics.sql

@@ -11,8 +11,11 @@
 -- name: RecommendationSourceMetricsForUser :many
 -- $1 user_id, $2 window_days. plays/skips are counts; avg_completion is the
 -- mean completion ratio over the completion_n plays that recorded one.
+-- pick_kind splits For You plays into taste/fresh/unattributed (#1249);
+-- it is NULL for every other source, so those still group to one row.
 SELECT
     pe.source,
+    pe.pick_kind,
     count(*)::bigint AS plays,
     count(*) FILTER (WHERE pe.was_skipped)::bigint AS skips,
     count(pe.completion_ratio)::bigint AS completion_n,
@@ -20,5 +23,5 @@ SELECT
 FROM play_events pe
 WHERE pe.user_id = $1
   AND pe.started_at > now() - ($2::float8 * INTERVAL '1 day')
-GROUP BY pe.source
+GROUP BY pe.source, pe.pick_kind
 ORDER BY plays DESC;