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/train/signals.py

@@ -0,0 +1,64 @@
+"""Неявная оценка пары (пользователь, трек) из событий — таблица docs/recommendations.md#3."""
+
+from __future__ import annotations
+
+import math
+from collections import defaultdict
+from dataclasses import dataclass
+
+HALF_LIFE_DAYS = 90.0
+DAY_S = 86400.0
+
+LIKE = 3.0
+DISLIKE = -5.0
+
+
+@dataclass(frozen=True)
+class Interaction:
+    user: str
+    track: str
+    ts: float  # unix-время события
+    weight: float  # вклад до затухания
+
+
+def event_weight(kind: str, played_ms: int | None, duration_ms: int | None, in_wave: bool) -> float | None:
+    """Вклад одного прослушивания. Скип вне волны слабее (×0.5): человек мог просто листать альбом."""
+    played = played_ms or 0
+    if kind == "play_end":
+        return 1.0
+    if kind != "skip":
+        return None
+    if duration_ms and played >= 0.5 * duration_ms:
+        return 0.7  # засчитанное прослушивание без доигрывания
+    if played < 10_000:
+        w = -1.0
+    elif played < 30_000:
+        w = -0.7
+    else:
+        w = -0.2
+    return w if in_wave else w * 0.5
+
+
+def decay(ts: float, now: float) -> float:
+    """Экспоненциальное затухание с полупериодом 90 дней."""
+    return math.exp(-math.log(2) * max(now - ts, 0.0) / DAY_S / HALF_LIFE_DAYS)
+
+
+def implicit_scores(
+    interactions: list[Interaction], now: float, before: float | None = None
+) -> dict[str, dict[str, float]]:
+    """user → track → implicit_score; before — учитывать только события раньше этого момента.
+    Дизлайк перекрывает всё: оценка трека становится DISLIKE (исключение из рекомендаций)."""
+    out: dict[str, dict[str, float]] = defaultdict(dict)
+    disliked: set[tuple[str, str]] = set()
+    for it in interactions:
+        if before is not None and it.ts >= before:
+            continue
+        if it.weight <= DISLIKE:
+            disliked.add((it.user, it.track))
+            continue
+        row = out[it.user]
+        row[it.track] = row.get(it.track, 0.0) + it.weight * decay(it.ts, now)
+    for user, track in disliked:
+        out[user][track] = DISLIKE
+    return dict(out)