feat(recommender): Thompson-sampling bandit over candidate sources (M6)

Nightly Beta(α, β) per user and source from wave impressions (reward:
complete or like, ×0.99 daily decay) stored in user_profiles_rec.bandit;
the wave adds 0.2·(θ − 0.5) per candidate source, sources without data
stay neutral. The user's exploration share now follows how well `fresh`
does against the other sources.

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

services/recommender/pipu_rec/train/bandit.py

@@ -0,0 +1,68 @@
+"""Бандит по источникам кандидатов (docs/recommendations.md#6.6): Beta(α, β) на пару (пользователь, источник).
+
+Награда — трек из источника доигран или лайкнут; скип и дизлайк — неудача. Наблюдения затухают ×0.99 в день,
+чтобы бандит подстраивался под смену вкуса. Доля «нового» (exploration) следует за успехом источника `fresh`.
+"""
+
+from __future__ import annotations
+
+import json
+from collections import defaultdict
+from typing import Any
+
+import psycopg
+
+DAILY_DECAY = 0.99
+REWARD = {"complete": 1.0, "like": 1.0, "skip": 0.0, "skip_early": 0.0, "dislike": 0.0}
+EXPLORATION_MIN, EXPLORATION_MAX = 0.05, 0.4
+
+Params = dict[str, list[float]]  # источник → [α, β]
+
+
+def compute(rows: list[tuple[str, str, str, float]], now: float) -> dict[str, Params]:
+    """rows: (user, source, outcome, served_at) → user → source → [α, β] с приором Beta(1, 1)."""
+    out: dict[str, Params] = defaultdict(dict)
+    for user, source, outcome, ts in rows:
+        r = REWARD.get(outcome)
+        if r is None:
+            continue
+        w = DAILY_DECAY ** max((now - ts) / 86400.0, 0.0)
+        ab = out[user].setdefault(source, [1.0, 1.0])
+        ab[0] += w * r
+        ab[1] += w * (1.0 - r)
+    return {u: {s: [round(a, 4), round(b, 4)] for s, (a, b) in p.items()} for u, p in out.items()}
+
+
+def exploration(params: Params, default: float = 0.15) -> float:
+    """Доля «нового»: успешнее `fresh` относительно остальных источников — больше исследования."""
+    fresh = params.get("fresh")
+    others = [a / (a + b) for s, (a, b) in params.items() if s != "fresh"]
+    if fresh is None or not others:
+        return default
+    rate = fresh[0] / (fresh[0] + fresh[1])
+    base = sum(others) / len(others)
+    value = default * (rate / base if base > 0 else 1.0)
+    return round(min(max(value, EXPLORATION_MIN), EXPLORATION_MAX), 3)
+
+
+def update_bandits(conn: psycopg.Connection[Any], now: float) -> int:
+    """Пересчитывает параметры по показам волны за 90 дней. Возвращает число пользователей."""
+    rows = conn.execute(
+        """
+        SELECT ws.user_id::text, wi.candidate_source, wi.outcome, extract(epoch FROM wi.served_at)::float8
+        FROM wave_impressions wi JOIN wave_sessions ws ON ws.id = wi.wave_session_id
+        WHERE wi.outcome IS NOT NULL AND wi.served_at > now() - interval '90 days'
+        """
+    ).fetchall()
+    params = compute([(u, s, o, float(ts)) for u, s, o, ts in rows], now)
+    with conn.transaction(), conn.cursor() as cur:
+        cur.executemany(
+            """
+            INSERT INTO user_profiles_rec (user_id, bandit, exploration, updated_at)
+            VALUES (%s, %s, %s, now())
+            ON CONFLICT (user_id) DO UPDATE SET bandit = EXCLUDED.bandit, exploration = EXCLUDED.exploration,
+                updated_at = now()
+            """,
+            [(u, json.dumps(p), exploration(p)) for u, p in params.items()],
+        )
+    return len(params)