feat(recommender): nightly ALS with offline metrics and CF candidates in the wave (M6)

- Implicit scores from play events and reactions per the signals table
  (90-day half-life, weaker skips outside the wave, dislikes override).
- numpy ALS for implicit feedback; time-split evaluation on the last
  7 days (NDCG@10, Recall@50) against a popularity baseline; a new model
  is activated only if NDCG@10 does not regress by more than 2%.
- Vectors are written to track_embeddings.cf and user_profiles_rec.taste_cf,
  versions to model_versions; NightlyTrainer runs at TRAIN_HOUR and on
  first start, `pipu-rec train` runs it manually.
- Wave: cf and item_cf candidate sources and a 0.10 cf term in the v1
  score; impressions are tagged heuristic-v1+cf.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

services/recommender/pipu_rec/wave/store.py

@@ -39,7 +39,8 @@ def load_index(conn: psycopg.Connection[Any]) -> CatalogIndex:
                 WHERE ta.track_id = t.id AND ta.role = 'main' ORDER BY ta.position LIMIT 1),
                 al.album_artist_id::text),
             e.audio::text, f.bpm, f.energy, f.danceability, f.valence, f.acousticness,
-            COALESCE((SELECT sum(play_count) FROM user_track_stats s WHERE s.track_id = t.id), 0)
+            COALESCE((SELECT sum(play_count) FROM user_track_stats s WHERE s.track_id = t.id), 0),
+            e.cf::text
         FROM tracks t
         JOIN albums al ON al.id = t.album_id
         JOIN media_files mf ON mf.id = t.primary_file_id AND mf.missing_since IS NULL
@@ -60,6 +61,7 @@ def load_index(conn: psycopg.Connection[Any]) -> CatalogIndex:
         raw_emb=[parse_vector(r[2]) for r in rows],
         features=feats,
         plays=[int(r[8]) for r in rows],
+        raw_cf=[parse_vector(r[9]) for r in rows],
     )
 
 
@@ -123,11 +125,15 @@ def load_profile(conn: psycopg.Connection[Any], user_id: str) -> Profile:
         scores[tid] = raw * math.exp(-math.log(2) * (now - ts) / 86400 / HALF_LIFE_DAYS)
     disliked: set[str] = set()
     liked: set[str] = set()
+    recent_likes: list[str] = []
     for tid, reaction in conn.execute(
-        "SELECT track_id::text, reaction::text FROM track_reactions WHERE user_id = %s", (user_id,)
+        "SELECT track_id::text, reaction::text FROM track_reactions WHERE user_id = %s "
+        "ORDER BY created_at DESC",
+        (user_id,),
     ):
         if reaction == "like":
             liked.add(tid)
+            recent_likes.append(tid)
             scores[tid] = scores.get(tid, 0.0) + 3.0
         else:
             disliked.add(tid)
@@ -149,8 +155,11 @@ def load_profile(conn: psycopg.Connection[Any], user_id: str) -> Profile:
             (user_id,),
         )
     }
-    row = conn.execute("SELECT exploration FROM user_profiles_rec WHERE user_id = %s", (user_id,)).fetchone()
+    row = conn.execute(
+        "SELECT exploration, taste_cf::text FROM user_profiles_rec WHERE user_id = %s", (user_id,)
+    ).fetchone()
     exploration = float(row[0]) if row else 0.15
+    taste_cf = parse_vector(row[1]) if row else None
     if len(played) < 30:
         exploration = max(exploration, 0.4)  # холодный старт: больше исследования на первых треках
     day = 86400
@@ -163,6 +172,8 @@ def load_profile(conn: psycopg.Connection[Any], user_id: str) -> Profile:
         liked_recent={t for t in liked if now - last.get(t, 0) < day},
         rediscover={t for t, s in scores.items() if s > 1.5 and now - last.get(t, now) > 60 * day},
         exploration=exploration,
+        taste_cf=taste_cf.tolist() if taste_cf is not None else None,
+        recent_likes=recent_likes,
     )