perf(for-you): cap my_saved_artists to 200 most-recent

[dylanjeffers]

Summary

The similar_artists CTE in /v1/users/{id}/feed/for-you self-joins saves against the my_saved_artists set. For long-tenure power users (e.g. an account with thousands of saved artists) the join blows up and the planner times out — observed in prod: the request hangs >60s and returns nothing.

Cap my_saved_artists to the 200 most-recently saved artists. Recency is the right axis to cut on:

• Old saves are weak signal of current taste anyway
• 200 artists still gives the collaborative-filter step plenty of surface to find similar-artist candidates
• Bounds the saves self-join cost to predictable size regardless of user tenure

Change

-- Before
my_saved_artists AS (
    SELECT DISTINCT t.owner_id AS artist_id
    FROM my_saved_tracks mst
    JOIN tracks t ON t.track_id = mst.track_id
),

-- After
my_saved_artists AS (
    SELECT t.owner_id AS artist_id, MAX(s.created_at) AS last_saved_at
    FROM saves s
    JOIN tracks t ON t.track_id = s.save_item_id
    WHERE s.user_id = @userId
      AND s.save_type = 'track'
      AND s.is_current = true
      AND s.is_delete = false
    GROUP BY t.owner_id
    ORDER BY last_saved_at DESC
    LIMIT 200
),

CTE output shape (column artist_id) is preserved, so the downstream IN (SELECT artist_id FROM my_saved_artists) and NOT IN (...) consumers in similar_artists are unchanged.

Test plan

• ✅ All 9 existing TestV1FeedForYou_* tests pass locally against the test DB (fixtures have <200 saved artists, so the cap doesn't kick in and behavior is identical to before for the test cases).
• ✅ go build ./api/... / go vet ./api/... clean.
• After deploy: re-curl /v1/users/eYZmn/feed/for-you?user_id=eYZmn&limit=5 (notjulian, deep save history) — should return 200 in <2s instead of hanging.

Follow-ups not in this PR

• Worth adding EXPLAIN ANALYZE instrumentation behind a flag to catch this class of regression earlier next time.
• The my_artist_affinity CTE also unions saves+reposts+plays for the user unbounded — likely the next-slowest piece for power users. Worth a similar bound in a follow-up if this fix doesn't fully resolve the timeout.

🤖 Generated with Claude Code