scraper.py

"""
藥妝店商品爬蟲 — 康是美 / 寶雅 (91APP) / 屈臣氏
已驗證方法:
  康是美: https://shop.cosmed.com.tw/search?q={keyword}
  寶雅:   https://www.poyabuy.com.tw/search?q={keyword}
  兩家都用 91APP 前端，selector: a.product-card__vertical
  屈臣氏: 先試直接呼叫 api.watsons.com.tw OAuth token（不用瀏覽器）；
          失敗才退回 Playwright Firefox 繞過 Cloudflare 取 token
"""

import re
import unicodedata
import urllib.parse
import concurrent.futures
import requests as _req

from playwright.sync_api import sync_playwright, TimeoutError as PWTimeout

TIMEOUT = 25_000
WATSONS_TIMEOUT = 35_000
UA = (
    "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) "
    "AppleWebKit/537.36 (KHTML, like Gecko) Chrome/124.0.0.0 Safari/537.36"
)
FIREFOX_UA = (
    "Mozilla/5.0 (Macintosh; Intel Mac OS X 10.15; rv:124.0) "
    "Gecko/20100101 Firefox/124.0"
)

def _to_ascii(s: str) -> str:
    """將帶重音的 Unicode 字母正規化為 ASCII（ü→u、é→e），讓 NARÜKO 視為 NARUKO。"""
    return ''.join(c for c in unicodedata.normalize('NFD', s) if ord(c) < 128)


# 促銷標籤模式 — 這些不是商品名稱，解析時跳過
_PROMO_RE = re.compile(
    r"^(POYA限定|買\d+送\d+|買一送一|買二送一|下單請選購\d+件|限定|特惠|"
    r"已售完|即將缺貨|\d+%\s*OFF|組合優惠|限時|新品|爆款|\d+折)$",
    re.IGNORECASE,
)

# 相關性檢查用的英文停用詞
_EN_SKIP = frozenset({'ml', 'spf', 'pa', 'uva', 'uvb', 'and', 'the', 'for', 'plus', 'ultra'})
# 組合商品偵測：「X條/入/件」、「套組」、「A+B」、「雙入」、「年度組」等
_COMBO_RE = re.compile(
    r'[2-9][條入件支管根]|\d+[條入件支管根]裝?|套組|超值組|試用組|組合[裝包]?|\w\+\w'
    r'|[兩雙][入條件]|[三四五六七八九][入條件]'       # 中文數字組合（雙入、兩條、三件…）
    r'|年度[組套]?|禮盒[組套]?|加購[組套]?'           # 年度組、禮盒組、加購組
    r'|[×x\*]\s*[2-9]\d*'                             # ×4、*2、x3 等乘量寫法（175g*4、175g×2）
    r'|\d+\s*[×x\*]\s*\d+',                           # 175g×2 的另一種位置（數字在前）
    re.IGNORECASE
)
# 規格提取：「140g」、「230ml」、「50片」等
_SIZE_RE = re.compile(r'(\d+(?:\.\d+)?)\s*(g|ml|mL|公克|毫升|oz|片)', re.IGNORECASE)


def _extract_sizes(text: str) -> set[str]:
    """提取文字中的規格（正規化，如 '140g'、'230ml'）。"""
    sizes = set()
    for m in _SIZE_RE.finditer(text):
        num = m.group(1)
        unit = m.group(2).lower().replace('公克', 'g').replace('毫升', 'ml').replace('mL', 'ml')
        sizes.add(f"{num}{unit}")
    return sizes


def _relevance_score(product_name: str, keyword: str) -> int:
    """
    多層次相關性分數：
    -1 = 品牌不符（排除）
     0 = 品牌符合但無產品特徵詞，或產品型態末字不符
    1+ = 產品特徵詞匹配分
    +2 = 規格（g/ml）也符合（加在特徵分之上）
    """
    name = product_name
    name_l = name.lower()
    # 將 ü/é 等重音字母正規化為 ASCII，讓 NARÜKO 作為完整 token 識別
    kw_norm = _to_ascii(keyword)
    name_l_ascii = _to_ascii(name).lower()
    zh_terms = re.findall(r'[一-鿿]{2,}', keyword)
    en_tokens = [t for t in re.findall(r'[A-Za-z][A-Za-z-]*[A-Za-z]|[A-Za-z]{2,}', kw_norm)
                 if t.lower() not in _EN_SKIP]
    if not zh_terms and not en_tokens:
        return 1
    first_en = kw_norm.find(en_tokens[0]) if en_tokens else len(kw_norm)
    first_zh = keyword.find(zh_terms[0]) if zh_terms else len(keyword)
    en_brand_primary = bool(en_tokens) and first_en < first_zh
    zh_brand_2 = zh_terms[0][:2] if zh_terms else ""
    # 品牌檢查（英文品牌用 ASCII 正規化後比對；屈臣氏等有時只列中文品名，英文找不到才退而用中文）
    if en_brand_primary:
        en_brand = en_tokens[0].lower()
        if len(en_brand) >= 3:
            if en_brand not in name_l_ascii:
                # 英文品牌找不到 → 退而比對中文品牌名（如屈臣氏只有「海倫仙度絲」沒有「Head」）
                if not zh_brand_2 or zh_brand_2 not in name:
                    return -1
        else:
            if zh_brand_2 and zh_brand_2 not in name:
                return -1
    else:
        if zh_brand_2 and zh_brand_2 not in name:
            return -1
    # zh_terms[0] 是品牌詞時整個跳過
    # 新增：英文品牌在前且有 ≥2 個中文詞 → 第一個必定是中文品牌名（如「高露潔」「海倫仙度絲」）
    skip_brand = (
        not en_brand_primary
        or (en_tokens and len(en_tokens[0]) < 3)
        or (zh_terms and len(zh_terms[0]) <= 2)
        or (en_brand_primary and zh_terms and len(zh_terms) >= 2)
    )
    # 產品型態錨點：取特徵詞中最長的那個的末 2 字，必須出現在商品名稱中。
    # 用最長詞（通常是產品類型「全面修復霜」）而非最後一詞（可能是修飾詞「升級版」），
    # 防止「修復霜」被「面膜 升級版」蒙混通過。
    if len(zh_terms) >= 2:
        feature_start = 1 if skip_brand else 0
        ft = zh_terms[feature_start:]
        if ft:
            anchor = max(ft, key=len)[-2:]
            if anchor not in name:
                return 0
    # 滑動視窗 2-gram：只計產品特徵詞，品牌詞全部跳過
    # 特例：只有一個中文詞時（如「雅漾亮顏潤色防曬乳」），去掉前 2 字（品牌）後當特徵詞
    match_keys: list[str] = []
    for idx, term in enumerate(zh_terms):
        if idx == 0 and skip_brand:
            if len(zh_terms) > 1:
                continue  # 多個詞時正常跳過品牌詞
            # 只有單一長中文詞：去掉前 2 字（品牌部分）
            term = term[2:] if len(term) > 3 else term
        for i in range(len(term) - 1):
            chunk = term[i:i + 2]
            if chunk not in match_keys:
                match_keys.append(chunk)

    # 前段 gram 強制比對：特徵詞 ≥6 個 2-gram 時，前三個 gram 中
    # 至少一個必須出現在商品名（辨識起始型態，防止 size_bonus 讓錯誤商品蒙混過門檻）。
    # 用「任一」而非「第一個」，容許店家略去詞首的小字（如省略「清爽」→「極效」仍可通過）。
    if len(match_keys) >= 6 and not any(g in name for g in match_keys[:3]):
        return 0

    base = sum(1 for k in match_keys if k in name)

    # 規格加分：base ≥ 1（產品類型已符合）且規格也一致時才加分
    kw_sizes = _extract_sizes(keyword)
    p_sizes = _extract_sizes(name)
    size_bonus = 2 if base >= 1 and kw_sizes and p_sizes and (kw_sizes & p_sizes) else 0

    return base + size_bonus


def _min_required_score(keyword: str) -> int:
    """
    keyword 有 ≥2 個不重複特徵 2-gram 時要求 score ≥ 2，
    避免「牙膏」「乳液」等通用詞讓不相關商品通過篩選。
    """
    kw_norm = _to_ascii(keyword)
    zh_terms = re.findall(r'[一-鿿]{2,}', keyword)
    en_tokens = [t for t in re.findall(r'[A-Za-z][A-Za-z-]*[A-Za-z]|[A-Za-z]{2,}', kw_norm)
                 if t.lower() not in _EN_SKIP]
    first_en = kw_norm.find(en_tokens[0]) if en_tokens else len(kw_norm)
    first_zh = keyword.find(zh_terms[0]) if zh_terms else len(keyword)
    en_brand_primary = bool(en_tokens) and first_en < first_zh
    skip_brand = (
        not en_brand_primary
        or (en_tokens and len(en_tokens[0]) < 3)
        or (zh_terms and len(zh_terms[0]) <= 2)
        or (en_brand_primary and zh_terms and len(zh_terms) >= 2)
    )
    grams: set[str] = set()
    for idx, term in enumerate(zh_terms):
        if idx == 0 and skip_brand:
            if len(zh_terms) > 1:
                continue
            term = term[2:] if len(term) > 3 else term
        for i in range(len(term) - 1):
            grams.add(term[i:i + 2])
    # 特徵詞越長，要求匹配的門檻越高：
    # ≥6 grams → 3（防止「去油光霧感防曬乳」混入「清爽極效全護防曬乳」結果）
    # ≥2 grams → 2
    # 其餘     → 1
    if len(grams) >= 6:
        return 3
    elif len(grams) >= 2:
        return 2
    else:
        return 1


def _size_ok(product_name: str, keyword: str) -> bool:
    """
    規格是否在合理範圍（差距 < 2 倍）。
    無規格資訊時回傳 True（不懲罰）。
    用於排序：回傳 False 的商品排到後面，但不會被過濾掉。
    """
    kw_sizes = _extract_sizes(keyword)
    p_sizes = _extract_sizes(product_name)
    if not kw_sizes or not p_sizes:
        return True
    try:
        kw_val = min(float(sz.rstrip('gml')) for sz in kw_sizes)
        p_val = min(float(sz.rstrip('gml')) for sz in p_sizes)
        return min(kw_val, p_val) <= 0 or max(kw_val, p_val) / min(kw_val, p_val) < 2.0
    except (ValueError, ZeroDivisionError):
        return True


def _size_diff(product_name: str, keyword: str) -> float:
    """回傳商品尺寸與搜尋尺寸的差距（數值越小越優先）；無尺寸資訊時回傳 inf。"""
    kw_sizes = _extract_sizes(keyword)
    p_sizes = _extract_sizes(product_name)
    if not kw_sizes or not p_sizes:
        return float('inf')
    try:
        kw_val = min(float(s.rstrip('gml')) for s in kw_sizes)
        p_val = min(float(s.rstrip('gml')) for s in p_sizes)
        return abs(kw_val - p_val)
    except Exception:
        return float('inf')


def _check_relevance(product_name: str, keyword: str) -> bool:
    """品牌符合且特徵詞匹配達到最低門檻（多特徵詞時要求 ≥2）。"""
    return _relevance_score(product_name, keyword) >= _min_required_score(keyword)


def _parse_91app_card(card) -> tuple[str, str, str]:
    """
    從 a.product-card__vertical 的 innerText 解析商品名稱、現售價、原價。
    91APP 卡片格式：原價（較高）在前，折後價（較低）在後。
    回傳 (name, current_price, original_price)，無折扣時 original_price 為空字串。
    """
    text = card.inner_text()
    lines = [l.strip() for l in text.split("\n") if l.strip()]

    name = ""
    for line in lines:
        if not _PROMO_RE.match(line) and len(line) > 2:
            name = line
            break

    # 找出所有價格，取最小值為現售價、最大值為原價
    prices = []
    for line in lines:
        m = re.search(r"NT\$\s*([\d,]+)", line)
        if m:
            prices.append(int(m.group(1).replace(",", "")))

    if not prices:
        return name, "", ""

    current = min(prices)
    original = max(prices)
    current_str = f"NT${current}"
    original_str = f"NT${original}" if original != current else ""

    return name, current_str, original_str


def _scrape_91app(store_name: str, base_url: str, keyword: str) -> dict | None:
    """通用 91APP 搜尋爬蟲。"""
    # 搜尋 URL 去掉規格數字（40ml、340g）和組合包數量（2入、3條…）
    # 讓店家回傳各規格/各入數的結果；排序時 _size_ok 挑規格、_COMBO_RE 挑單入
    search_kw = re.sub(
        r'\s*\d+(?:\.\d+)?\s*(?:g|ml|mL|公克|毫升|oz|片)\b', '',
        keyword, flags=re.IGNORECASE
    )
    search_kw = re.sub(
        r'\s*[2-9]\s*[條入件支管根]', '',
        search_kw
    ).strip() or keyword
    # 若英文品牌在前且有多個中文詞（如「CeraVe 適樂膚 長效潤澤修復霜」），
    # 第一個中文詞是英文品牌的中文翻譯，AND 兩個品牌名讓 91APP 找不到商品 → 移除
    _sq_zh = re.findall(r'[一-鿿]{2,}', search_kw)
    _sq_en = [t for t in re.findall(r'[A-Za-z][A-Za-z-]*[A-Za-z]|[A-Za-z]{2,}',
                                     _to_ascii(search_kw)) if t.lower() not in _EN_SKIP]
    if (_sq_en and len(_sq_zh) >= 2
            and search_kw.find(_sq_en[0]) < search_kw.find(_sq_zh[0])):
        search_kw = search_kw.replace(_sq_zh[0], '', 1).strip()
    # normalize double spaces after replace
    search_kw = ' '.join(search_kw.split())
    encoded = urllib.parse.quote(search_kw)
    search_url = f"{base_url}/search?q={encoded}"
    print(f"[91app] {store_name} search_url={search_url!r}", flush=True)

    try:
        with sync_playwright() as pw:
            browser = pw.chromium.launch(headless=True, args=["--no-sandbox"])
            page = browser.new_page(user_agent=UA)
            try:
                page.goto(search_url, wait_until="networkidle", timeout=TIMEOUT)
            except PWTimeout:
                pass

            cards = page.query_selector_all("a.product-card__vertical")
            print(f"[91app] {store_name} cards={len(cards)} min_req={_min_required_score(keyword)}", flush=True)
            # 收集所有相關商品候選，(score, is_combo, name, price, orig, href)
            candidates: list[tuple] = []
            min_req = _min_required_score(keyword)
            for card in cards[:8]:
                href = card.get_attribute("href") or ""
                name, price, original_price = _parse_91app_card(card)
                if not name or not price:
                    continue
                s = _relevance_score(name, keyword)
                print(f"[91app] {store_name} name={name!r} score={s}", flush=True)
                if s < min_req:
                    continue
                combo = bool(_COMBO_RE.search(name))
                candidates.append((s, combo, name, price, original_price, href))
            browser.close()
        # 排序：非組合優先 → 規格相符優先（差距 < 2 倍）→ 分數高 → 尺寸最接近
        # 規格不符的候選排到後面但不過濾，確保找不到相符規格時仍能回傳結果
        if not candidates:
            return None
        candidates.sort(key=lambda x: (
            x[1],                                        # 非組合優先
            0 if _size_ok(x[2], keyword) else 1,        # 規格相符優先
            -x[0],                                       # 分數高優先
            _size_diff(x[2], keyword),                   # 尺寸差距小優先
        ))
        s, combo, name, price, original_price, href = candidates[0]
        product_url = f"{base_url}{href}" if href.startswith("/") else href
        return {
            "store": store_name,
            "name": name,
            "price": price,
            "original_price": original_price,
            "url": product_url,
            "real": True,
        }
    except Exception:
        return None


def scrape_cosmed(keyword: str) -> dict | None:
    return _scrape_91app("康是美", "https://shop.cosmed.com.tw", keyword)


def scrape_poya(keyword: str) -> dict | None:
    return _scrape_91app("寶雅", "https://www.poyabuy.com.tw", keyword)


def _parse_watsons_json(data: dict, keyword: str) -> dict | None:
    """從 Watsons API JSON 解析商品資訊：用 _relevance_score 計分，優先非組合商品。"""
    products = data.get("products", [])
    if not products:
        return None

    # 過濾：相關性分數 ≥ min_required（多特徵詞時要求 ≥2，防止「牙膏」等通用詞誤選）
    min_req = _min_required_score(keyword)
    scored = []
    for p in products:
        s = _relevance_score(p.get("name", ""), keyword)
        if s >= min_req:
            combo = bool(_COMBO_RE.search(p.get("name", "")))
            scored.append((s, combo, p))

    if not scored:
        return None

    # 排序：非組合優先 → 分數高 → 尺寸最接近搜尋規格
    scored.sort(key=lambda x: (x[1], -x[0], _size_diff(x[2].get("name", ""), keyword)))
    _, _, matched = scored[0]

    name = matched.get("name", keyword)
    price_info = matched.get("price", {})
    raw_price = price_info.get("value") if isinstance(price_info, dict) else price_info
    price = f"NT${int(raw_price)}" if raw_price else "請洽門市"

    # Watsons API 原價欄位
    # strikeThroughPrice 可能是純數字（float）或 dict，需分別處理
    # elabOldPrice / basePrice / wasPrice 是 dict with .value
    def _extract_price_value(v):
        if isinstance(v, (int, float)) and v:
            return float(v)
        if isinstance(v, dict):
            return v.get("value")
        return None

    list_price_info = price_info.get("listPrice", {}) if isinstance(price_info, dict) else {}
    raw_base = (
        _extract_price_value(matched.get("strikeThroughPrice"))
        or _extract_price_value(matched.get("elabOldPrice"))
        or _extract_price_value(matched.get("basePrice"))
        or _extract_price_value(matched.get("wasPrice"))
        or _extract_price_value(list_price_info)
    )
    original_price = f"NT${int(raw_base)}" if raw_base and raw_base != raw_price else ""

    code = matched.get("code", "")
    encoded_kw = urllib.parse.quote(keyword)
    url = (
        f"https://www.watsons.com.tw/p/{code}"
        if code
        else f"https://www.watsons.com.tw/search?text={encoded_kw}"
    )

    return {
        "store": "屈臣氏",
        "name": name,
        "price": price,
        "original_price": original_price,
        "url": url,
        "real": True,
    }


def _watsons_api_search(cf, api_url: str, headers: dict, keyword: str, token: str | None) -> dict | None:
    """用 curl_cffi 呼叫一次 Watsons 搜尋 API（不帶或帶 token）。"""
    try:
        h = {**headers}
        if token:
            h["Authorization"] = f"bearer {token}"
        r = cf.get(api_url, headers=h, impersonate="chrome120", timeout=12)
        if r.ok:
            return _parse_watsons_json(r.json(), keyword)
    except Exception:
        pass
    return None


def _watsons_via_curl(keyword: str) -> dict | None:
    """
    用 curl_cffi 模擬 Chrome TLS 指紋，繞過 Cloudflare bot 檢測。
    漸進式搜尋：完整中文詞 → 退回品牌名；每個搜尋詞都試不帶/帶 token。
    """
    try:
        from curl_cffi import requests as cf
    except ImportError:
        return None

    zh_terms = re.findall(r'[一-鿿]{2,}', keyword)

    # 搜尋詞清單：由精確到寬鬆，中文精確詞優先，英文品牌兜底
    search_attempts: list[str] = []

    # 1. 中文詞精確搜尋（放最前，最重要）
    if zh_terms:
        brand = zh_terms[0]
        if len(zh_terms) > 1:
            product = zh_terms[1]
            search_attempts.append(brand + product)          # 全部合併（最精確）
            for n in (4, 2):                                 # 品牌 + 產品詞前 n 字
                if len(product) >= n:
                    combo = brand + product[:n]
                    if combo not in search_attempts:
                        search_attempts.append(combo)
        if brand not in search_attempts:
            search_attempts.append(brand)                    # 品牌名（保底）
    else:
        search_attempts.append(keyword)

    # 2. 若英文品牌開頭（如 NARÜKO、Colgate），英文品牌搜尋作為後備
    #    放最後：中文搜尋找不到時才用英文（避免寬泛英文搜尋蓋掉精確中文結果）
    if keyword and keyword[0].isascii() and keyword[0].isalpha():
        kw_ascii = _to_ascii(keyword)
        _en_m = re.match(r'^([A-Za-z][A-Za-z0-9\-]*(?:\s+[A-Za-z0-9\-]+)*)', kw_ascii)
        if _en_m:
            _en_brand = _en_m.group(1).strip()
            if zh_terms:
                _combo = _en_brand + zh_terms[0][:4]  # e.g. "NARUKO茶樹"
                if _combo not in search_attempts:
                    search_attempts.append(_combo)
            if _en_brand not in search_attempts:
                search_attempts.append(_en_brand)      # e.g. "NARUKO"（最後兜底）

    base_headers = {
        "Accept": "application/json",
        "Accept-Language": "zh-TW,zh;q=0.9,en-US;q=0.8",
        "Origin": "https://www.watsons.com.tw",
    }

    # 先嘗試不帶 token 的所有搜尋詞（SAP Hybris 有時允許匿名）
    for st in search_attempts:
        kw = urllib.parse.quote(st)
        api_url = (
            f"https://api.watsons.com.tw/api/v2/wtctw/products/search"
            f"?fields=FULL&query={kw}&pageSize=10&currentPage=0&lang=zh_TW&curr=TWD"
        )
        headers = {**base_headers, "Referer": f"https://www.watsons.com.tw/search?text={kw}"}
        result = _watsons_api_search(cf, api_url, headers, keyword, token=None)
        if result:
            return result

    # 取 OAuth token（只做一次）
    token = None
    first_kw = urllib.parse.quote(search_attempts[0])
    oauth_headers = {
        **base_headers,
        "Content-Type": "application/x-www-form-urlencoded",
        "Referer": f"https://www.watsons.com.tw/search?text={first_kw}",
    }
    for cid, csec in [
        ("mobile", "secret"),
        ("client", "secret"),
        ("trusted_client", "secret"),
        ("mobile", ""),
        ("wtctw_mobile", "secret"),
    ]:
        try:
            tr = cf.post(
                "https://api.watsons.com.tw/oauth/token",
                data={"grant_type": "client_credentials", "client_id": cid, "client_secret": csec},
                headers=oauth_headers,
                impersonate="chrome120",
                timeout=8,
            )
            if tr.ok:
                token = tr.json().get("access_token")
                if token:
                    break
        except Exception:
            continue

    if not token:
        return None

    # 帶 token 再試所有搜尋詞
    for st in search_attempts:
        kw = urllib.parse.quote(st)
        api_url = (
            f"https://api.watsons.com.tw/api/v2/wtctw/products/search"
            f"?fields=FULL&query={kw}&pageSize=10&currentPage=0&lang=zh_TW&curr=TWD"
        )
        headers = {**base_headers, "Referer": f"https://www.watsons.com.tw/search?text={kw}"}
        result = _watsons_api_search(cf, api_url, headers, keyword, token=token)
        if result:
            return result

    return None


def _watsons_via_firefox(keyword: str) -> dict | None:
    """Playwright Firefox 備援：攔截 OAuth token 後呼叫 API。"""
    zh_terms = re.findall(r'[一-鿿]{2,}', keyword)
    # 中文詞優先；英文品牌開頭但找不到中文詞時才用英文品牌
    if zh_terms:
        search_term = "".join(zh_terms[:2])   # 例如 "雅漾清爽極效全護防曬乳"
    elif keyword and keyword[0].isascii() and keyword[0].isalpha():
        kw_ascii = _to_ascii(keyword)
        _en_m = re.match(r'^([A-Za-z][A-Za-z0-9\-]*(?:\s+[A-Za-z0-9\-]+)*)', kw_ascii)
        search_term = _en_m.group(1).strip() if _en_m else keyword
    else:
        search_term = keyword
    kw = urllib.parse.quote(search_term)
    page_url = f"https://www.watsons.com.tw/search?text={kw}"

    try:
        with sync_playwright() as pw:
            browser = pw.firefox.launch(headless=True)
            context = browser.new_context(
                user_agent=FIREFOX_UA,
                locale="zh-TW",
                extra_http_headers={
                    "Accept-Language": "zh-TW,zh;q=0.8,en-US;q=0.5,en;q=0.3",
                    "Accept": "text/html,application/xhtml+xml,application/xml;q=0.9,image/avif,image/webp,*/*;q=0.8",
                    "Upgrade-Insecure-Requests": "1",
                },
            )
            page = context.new_page()
            page.add_init_script(
                "Object.defineProperty(navigator, 'webdriver', {get: () => undefined});"
            )

            access_token = None
            search_data = None

            def on_resp(resp):
                nonlocal access_token, search_data
                try:
                    if "oauth/token" in resp.url and not access_token:
                        body = resp.json()
                        access_token = body.get("access_token")
                    elif "products/search" in resp.url and not search_data:
                        search_data = resp.json()
                except Exception:
                    pass

            page.on("response", on_resp)
            try:
                page.goto(page_url, wait_until="networkidle", timeout=WATSONS_TIMEOUT)
            except Exception:
                pass

            context.close()
            browser.close()

        # 優先用瀏覽器攔截到的搜尋結果（最準確）
        if search_data:
            return _parse_watsons_json(search_data, keyword)

        # 其次用 token 再呼叫 API
        if access_token:
            kw_api = urllib.parse.quote(search_term)
            api_url = (
                f"https://api.watsons.com.tw/api/v2/wtctw/products/search"
                f"?fields=FULL&query={kw_api}&pageSize=10&currentPage=0&lang=zh_TW&curr=TWD"
            )
            try:
                resp = _req.get(
                    api_url,
                    headers={
                        "Authorization": f"bearer {access_token}",
                        "Accept": "application/json",
                        "Accept-Language": "zh-TW,zh;q=0.9",
                        "User-Agent": FIREFOX_UA,
                        "Origin": "https://www.watsons.com.tw",
                        "Referer": f"https://www.watsons.com.tw/search?text={kw}",
                    },
                    timeout=15,
                )
                if resp.ok:
                    return _parse_watsons_json(resp.json(), keyword)
            except Exception:
                pass
    except Exception:
        pass

    return None


def scrape_watsons(keyword: str) -> dict | None:
    """
    屈臣氏搜尋三段式：
    1. curl_cffi 模擬 Chrome TLS 指紋（輕量、快速）
    2. Playwright Firefox 備援（慢但能攔截 token）
    """
    result = _watsons_via_curl(keyword)
    if result:
        return result
    return _watsons_via_firefox(keyword)


def discover_cosmed(keyword: str) -> dict | None:
    """
    二段式搜尋的第一階段：在康是美以寬鬆條件搜尋，回傳真實架上商品的完整結果 dict。
    評分 ≥ 1 即可（品牌符合 + 至少一個特徵 gram），取非組合商品最高分者。
    回傳格式與 scrape_cosmed 相同：{"store", "name", "price", "url", "real"}
    """
    base_url = "https://shop.cosmed.com.tw"
    try:
        encoded = urllib.parse.quote(keyword)
        search_url = f"{base_url}/search?q={encoded}"
        with sync_playwright() as pw:
            browser = pw.chromium.launch(headless=True, args=["--no-sandbox"])
            page = browser.new_page(user_agent=UA)
            try:
                page.goto(search_url, wait_until="networkidle", timeout=TIMEOUT)
            except Exception:
                pass
            base_url = "https://shop.cosmed.com.tw"
            cards = page.query_selector_all("a.product-card__vertical")
            best_score = -1
            best_result = None
            for card in cards[:8]:
                href = card.get_attribute("href") or ""
                name, price, original_price = _parse_91app_card(card)
                if not name or not price:
                    continue
                if _COMBO_RE.search(name):
                    continue
                s = _relevance_score(name, keyword)
                if s >= 1 and s > best_score:
                    best_score = s
                    product_url = f"{base_url}{href}" if href.startswith("/") else href
                    best_result = {
                        "store": "康是美",
                        "name": name,
                        "price": price,
                        "original_price": original_price,
                        "url": product_url,
                        "real": True,
                    }
            browser.close()
        return best_result
    except Exception:
        return None


def discover_watsons(keyword: str) -> dict | None:
    """
    快速探索屈臣氏商品名稱（僅走 curl 快速路徑，不啟動 Playwright）。
    回傳格式與 scrape_watsons 相同：{"store", "name", "price", "url", "real"}
    """
    return _watsons_via_curl(keyword)


def scrape_watsons_and_poya(keyword: str) -> list[dict]:
    """屈臣氏 + 寶雅平行搜尋，供二段式搜尋的第二階段使用。"""
    tasks = {"屈臣氏": scrape_watsons, "寶雅": scrape_poya}
    results = []
    with concurrent.futures.ThreadPoolExecutor(max_workers=3) as executor:
        future_map = {executor.submit(fn, keyword): store for store, fn in tasks.items()}
        try:
            for future in concurrent.futures.as_completed(future_map, timeout=65):
                try:
                    data = future.result()
                    if data:
                        results.append(data)
                except Exception:
                    pass
        except concurrent.futures.TimeoutError:
            for future in future_map:
                if future.done():
                    try:
                        data = future.result()
                        if data:
                            results.append(data)
                    except Exception:
                        pass
    order = ["屈臣氏", "寶雅"]
    results.sort(key=lambda x: order.index(x["store"]) if x["store"] in order else 99)
    return results


def scrape_all_stores(keyword: str) -> list[dict]:
    """
    並行爬取三家，回傳成功結果（最多 3 筆），依屈臣氏→康是美→寶雅排序。
    每筆格式: {"store", "name", "price", "url", "real"}
    """
    print(f"[scraper] keyword={keyword!r}  min_req={_min_required_score(keyword)}", flush=True)
    tasks = {
        "屈臣氏": scrape_watsons,
        "康是美": scrape_cosmed,
        "寶雅": scrape_poya,
    }
    results = []
    with concurrent.futures.ThreadPoolExecutor(max_workers=3) as executor:
        future_map = {executor.submit(fn, keyword): store for store, fn in tasks.items()}
        try:
            for future in concurrent.futures.as_completed(future_map, timeout=65):
                try:
                    data = future.result()
                    if data:
                        results.append(data)
                except Exception:
                    pass
        except concurrent.futures.TimeoutError:
            for future, store in future_map.items():
                if future.done():
                    try:
                        data = future.result()
                        if data:
                            results.append(data)
                    except Exception:
                        pass

    order = ["屈臣氏", "康是美", "寶雅"]
    results.sort(key=lambda x: order.index(x["store"]) if x["store"] in order else 99)
    return results