feat(recommender): LightGBM lambdarank ranker with impression features (M6)

The wave logs each served candidate's feature vector to rec_impression_features.
Nightly training joins them with impression outcomes, validates on the last week
and activates the model only when it beats the v1 order and the active model.
The service loads the active ranker and uses its order in place of the v1 score.

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

services/recommender/pipu_rec/wave/service.py

@@ -2,18 +2,22 @@
 
 from __future__ import annotations
 
+import logging
 from typing import Literal
 
 import numpy as np
 import psycopg
 from pydantic import BaseModel, Field
 
+from pipu_rec.train.ranker import RankerHolder
 from pipu_rec.wave.contexts import Contexts, Rule
-from pipu_rec.wave.engine import Pick, build_batch, model_version, taste_centroids, update_session
+from pipu_rec.wave.engine import Context, Pick, build_batch, model_version, taste_centroids, update_session
 from pipu_rec.wave.store import IndexHolder, SessionStore, load_profile, seed_tracks
 
 MAX_SERVED_MEMORY = 300
 
+log = logging.getLogger(__name__)
+
 Outcome = Literal["complete", "skip", "skip_early", "like", "dislike"]
 
 
@@ -55,6 +59,12 @@ class WaveService:
         self.index = IndexHolder(dsn)
         self.sessions = SessionStore(redis_url)
         self.contexts = contexts
+        self.ranker = RankerHolder(dsn)
+
+    def reload(self) -> None:
+        """После ночного обучения: перечитать каталог (векторы ALS) и активный ранжировщик."""
+        self.index.refresh()
+        self.ranker.reload()
 
     def rules(self, activity: str | None, mood: str | None) -> list[Rule]:
         """Занятие и настроение комбинируются — пересечение профилей; неизвестный ключ игнорируется."""
@@ -74,6 +84,7 @@ class WaveService:
         update_session(
             index, session, [(f.track_id, f.outcome) for f in req.feedback], taste_centroids(index, profile)
         )
+        ranker = self.ranker.get()
         picks: list[Pick] = build_batch(
             index,
             profile,
@@ -83,12 +94,31 @@ class WaveService:
             exclude=set(req.exclude_ids),
             seed_ids=seeds,
             rng=np.random.default_rng(),
+            ranker=ranker,
+            ctx=Context(has_activity=bool(req.activity), has_mood=bool(req.mood)),
         )
+        self.log_features(req.session_id, picks)
         session.served = (session.served + [p.id for p in picks])[-MAX_SERVED_MEMORY:]
         artists = [index.artist[index.pos[p.id]] for p in picks]
         session.recent_artists = (session.recent_artists + artists)[-10:]
         self.sessions.put(req.session_id, session)
         return BatchResponse(
             tracks=[BatchTrack(id=p.id, source=p.source, score=p.score) for p in picks],
-            model_ver=model_version(index, profile),
+            model_ver=model_version(index, profile, ranker),
         )
+
+    def log_features(self, session_id: str, picks: list[Pick]) -> None:
+        """Признаки показанных треков — обучающие данные ранжировщика v2. Сбой записи не ломает волну."""
+        rows = [(session_id, p.id, p.features) for p in picks if p.features is not None]
+        if not rows:
+            return
+        try:
+            with psycopg.connect(self.dsn, autocommit=True) as conn, conn.cursor() as cur:
+                cur.executemany(
+                    "INSERT INTO rec_impression_features (wave_session_id, track_id, features) "
+                    "VALUES (%s, %s, %s) "
+                    "ON CONFLICT DO NOTHING",
+                    rows,
+                )
+        except psycopg.Error as e:
+            log.warning("impression features not saved: %s", e)