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/tests/test_train.py

@@ -0,0 +1,110 @@
+"""ALS, неявные оценки, офлайн-метрики и правило активации модели."""
+
+import numpy as np
+
+from pipu_rec.train.als import train_als
+from pipu_rec.train.metrics import evaluate, ndcg_at_k, recall_at_k
+from pipu_rec.train.pipeline import holdout, should_activate
+from pipu_rec.train.signals import DISLIKE, Interaction, event_weight, implicit_scores
+from pipu_rec.wave.engine import CatalogIndex, Profile, Session, build_batch, model_version
+
+DAY = 86400.0
+
+
+def test_event_weights_follow_spec() -> None:
+    assert event_weight("play_end", 200_000, 200_000, True) == 1.0
+    assert event_weight("skip", 120_000, 200_000, True) == 0.7  # ≥ 50% — засчитано
+    assert event_weight("skip", 5_000, 200_000, True) == -1.0
+    assert event_weight("skip", 20_000, 200_000, True) == -0.7
+    assert event_weight("skip", 5_000, 200_000, False) == -0.5  # вне волны — слабее
+    assert event_weight("play_start", 0, 200_000, True) is None
+
+
+def test_implicit_scores_decay_and_dislike() -> None:
+    now = 1000 * DAY
+    inter = [
+        Interaction("u", "a", now, 1.0),
+        Interaction("u", "b", now - 90 * DAY, 1.0),  # полупериод — половина веса
+        Interaction("u", "c", now, 3.0),
+        Interaction("u", "c", now, DISLIKE),  # дизлайк перекрывает
+    ]
+    s = implicit_scores(inter, now)["u"]
+    assert s["a"] == 1.0
+    assert abs(s["b"] - 0.5) < 1e-9
+    assert s["c"] == DISLIKE
+
+
+def two_taste_world() -> dict[str, dict[str, float]]:
+    """Две группы слушателей с непересекающимися вкусами: rock_* и jazz_*."""
+    rng = np.random.default_rng(1)
+    scores: dict[str, dict[str, float]] = {}
+    for g in ("rock", "jazz"):
+        for u in range(15):
+            liked = rng.choice(30, size=12, replace=False)
+            scores[f"{g}_user{u}"] = {f"{g}_{i}": 1.0 + float(rng.random()) for i in liked}
+    return scores
+
+
+def test_als_learns_taste_groups() -> None:
+    model = train_als(two_taste_world(), factors=8, iterations=10)
+    assert model is not None
+    s = model.scores("rock_user0")
+    rock = np.mean([s[i] for i, t in enumerate(model.items) if t.startswith("rock")])
+    jazz = np.mean([s[i] for i, t in enumerate(model.items) if t.startswith("jazz")])
+    assert rock > jazz + 0.3
+
+
+def test_metrics() -> None:
+    assert ndcg_at_k(["a", "b"], {"a"}) == 1.0
+    assert 0 < ndcg_at_k(["x", "a"], {"a"}) < 1
+    assert recall_at_k(["a", "x"], {"a", "b"}) == 0.5
+    res = evaluate(["u", "v"], {"u": {"a"}}, lambda u: ["a"])
+    assert res == {"users": 1, "ndcg@10": 1.0, "recall@50": 1.0}
+
+
+def test_holdout_keeps_only_new_positives() -> None:
+    now = 100 * DAY
+    inter = [
+        Interaction("u", "old", now - 30 * DAY, 1.0),
+        Interaction("u", "old", now - DAY, 1.0),  # уже был в обучении — не «новый»
+        Interaction("u", "new", now - DAY, 1.0),
+        Interaction("u", "skipped", now - DAY, -1.0),
+    ]
+    train, relevant = holdout(inter, now)
+    assert set(train["u"]) == {"old"}
+    assert relevant == {"u": {"new"}}
+
+
+def test_activation_gate() -> None:
+    cur = {"offline": {"als": {"ndcg@10": 0.30}}}
+    assert should_activate({"als": {"ndcg@10": 0.295}}, cur)  # −1.7% — допустимо
+    assert not should_activate({"als": {"ndcg@10": 0.28}}, cur)
+    assert should_activate({"users": 0}, cur)  # не на чем проверить
+    assert should_activate({"als": {"ndcg@10": 0.1}}, None)
+
+
+def test_engine_uses_cf_source() -> None:
+    n = 40
+    ids = [f"t{i}" for i in range(n)]
+    cf = [
+        np.array([1.0, 0.0], dtype=np.float32) if i < 20 else np.array([0.0, 1.0], dtype=np.float32)
+        for i in range(n)
+    ]
+    index = CatalogIndex.build(
+        ids=ids, artist=[f"a{i}" for i in range(n)], raw_emb=[None] * n, features={}, plays=[0] * n, raw_cf=cf
+    )
+    profile = Profile(
+        scores={},
+        disliked=set(),
+        disliked_artists=set(),
+        recent=set(),
+        played=set(),
+        liked_recent=set(),
+        rediscover=set(),
+        exploration=0.0,
+        taste_cf=[1.0, 0.0],
+    )
+    picks = build_batch(index, profile, Session(), [], n=5, rng=np.random.default_rng(0))
+    assert len(picks) == 5
+    assert sum(int(p.id[1:]) < 20 for p in picks) >= 4  # почти всё — из «своей» половины по ALS
+    assert model_version(index, profile).endswith("+cf")