"""Odtworzenie uczenia feedu Edge zdarzenie po zdarzeniu (bez zagladania w przyszlosc).

Dla kazdego wyswietlenia (complete/skip/impression/autoswipe) zapisuje cechy znane W TEJ
CHWILI: skladniki obecnej punktacji z wag urzadzenia sprzed zdarzenia oraz liczniki klipu
z wczesniejszych zdarzen wszystkich urzadzen. Wynik CSV na stdout.

Uzyte do przestrojenia punktacji 2026-10-02 (patrz app/clips_feed.py, score_candidate).
Wybor tagow na starcie (onboarding) nie jest zdarzeniem, wiec tu go brakuje.

  docker compose exec -T worker sh -c "cd /srv && python3 scripts/edge_replay_features.py | gzip -c" > edge.csv.gz
Analiza lokalnie: python scripts/edge_tuning_analysis.py edge.csv (pandas + scikit-learn).
"""
import csv
import math
import sys
from collections import defaultdict
from datetime import timedelta

from sqlalchemy import text

from app import clips_feed as cf
from app.db import session_scope

with session_scope() as s:
    ev = s.execute(text("""
        select device_id, clip_id, kind, dwell_ms, reason, slot, created_at
        from clip_feed_events where clip_id is not null order by created_at""")).all()
    ids = list({e.clip_id for e in ev})
    clips = {}
    for i in range(0, len(ids), 5000):
        for r in s.execute(text("""select id, creator, likes, published_at, niches, duration_sec, width, height,
                                          has_audio, creator_verified
                                   from clips where id = any(:ids)"""), {"ids": ids[i:i + 5000]}).all():
            clips[r.id] = r
    tags = defaultdict(list)
    for i in range(0, len(ids), 5000):
        for cid, t in s.execute(text("select clip_id, tag from clip_tags where clip_id = any(:ids)"), {"ids": ids[i:i + 5000]}).all():
            tags[cid].append(t)
    n_all = s.execute(text("select count(*) from clips")).scalar()
    tag_df = dict(s.execute(text("select tag, count(*) from clip_tags group by tag")).all())
idf = {t: math.log((n_all + 1) / (n + 1)) for t, n in tag_df.items()}
print(f"zdarzen {len(ev)}, klipow {len(clips)}, tagow {len(idf)}", file=sys.stderr)

aff = defaultdict(cf.Affinity)
last_day = {}
dev_n = defaultdict(int)
# liczniki klipu z wczesniejszych zdarzen (wszystkie urzadzenia)
c_seen = defaultdict(int)
c_comp = defaultdict(int)
c_skip = defaultdict(int)
c_like = defaultdict(int)
liked = set((e.device_id, e.clip_id) for e in ev if e.kind == "like")

w = csv.writer(sys.stdout)
w.writerow(["dev", "ts", "slot", "kind", "dwell", "liked", "tag_sum", "tag_pos", "tag_n", "niche_sum", "creator_w",
            "tag_idf_sum", "tag_idf_pos", "pop", "fresh", "dur", "vertical", "audio", "verified", "c_seen", "c_comp", "c_skip", "c_like", "dev_n"])
OUTCOME = {"complete", "skip", "impression", "autoswipe"}
for e in ev:
    c = clips.get(e.clip_id)
    if c is None:
        continue
    a = aff[e.device_id]
    day = e.created_at.date()
    prev = last_day.get(e.device_id)
    if prev is not None and day > prev:
        for _ in range((day - prev).days):
            a.tags = cf.decay_weights(a.tags)
            a.niches = cf.decay_weights(a.niches)
            a.creators = cf.decay_weights(a.creators)
    last_day[e.device_id] = day
    t = tags.get(e.clip_id, [])
    niches = list(c.niches or [])
    if e.kind in OUTCOME:
        ts_vals = [a.tags.get(x, 0.0) for x in t]
        fresh = int(bool(c.published_at and c.published_at >= e.created_at - timedelta(days=cf.FRESH_DAYS)))
        vert = int(bool(c.width and c.height and c.height > c.width))
        w.writerow([e.device_id[:12], e.created_at.isoformat(), e.slot or "", e.kind, e.dwell_ms or 0,
                    int((e.device_id, e.clip_id) in liked),
                    round(sum(ts_vals), 3), sum(1 for v in ts_vals if v > 0), len(t),
                    round(sum(a.niches.get(n, 0.0) for n in niches), 3),
                    round(a.creators.get(c.creator or "", 0.0), 3),
                    round(sum(a.tags.get(x, 0.0) * idf.get(x, 0.0) for x in t), 3),
                    round(sum(idf.get(x, 0.0) for x in t if a.tags.get(x, 0.0) > 0), 3),
                    round(math.log1p(max(c.likes or 0, 0)), 3), fresh, round(c.duration_sec or 0, 1), vert,
                    int(bool(c.has_audio)), int(bool(c.creator_verified)),
                    c_seen[e.clip_id], c_comp[e.clip_id], c_skip[e.clip_id], c_like[e.clip_id], dev_n[e.device_id]])
        c_seen[e.clip_id] += 1
        if e.kind == "complete":
            c_comp[e.clip_id] += 1
        elif e.kind == "skip":
            c_skip[e.clip_id] += 1
        dev_n[e.device_id] += 1
    elif e.kind == "like":
        c_like[e.clip_id] += 1
    cf.apply_event(a, kind=e.kind, dwell_ms=e.dwell_ms, reason=e.reason, tags=t, niches=niches, creator=c.creator)