Neo-ZQYY/scripts/analyze_store_manager_quality.py

"""App2 财务洞察 · 店长视角评分方法 v2。

围绕"店长读了这 12 条能做什么 / 学到什么"做评估，三大维度：

【1】准确性（客观正确 · 越高越好）
  - 数字取自 payload，无编造
  - 引用权威字段（非原始指标兜底）
  - 遵守业务规则（手动调整不拆明细 / 禁行业数字 / 禁单期推测 / 对比口径）
  - 数据缺失/0 支出主动标注

【2】洞察深度（业务价值 · 越高越好）
  - 非显性信号挖掘（"消耗>充值=存量消费"/"差值=平均让利"等解读）
  - 多指标协同分析（seq 11 结构失衡 + 原因互相印证）
  - 根因而非表象（如"数据录入缺失"而不只"支出为 0"）
  - seq 12 跟踪 4 要素齐全（指标 / 阈值 / 节奏 / 触发动作）
  - 避免空洞表达（"关注 XX" / "加强 XX" 被扣分）

【3】稳定性（可靠性 · 越高越好；稳定 ≠ 僵化）
  - 健康度评级方向一致（同数据下 10 次评级应相似，灯色众数占比 ≥ 80%）
  - 关键原因收敛（seq 11 原因 1/2 所引 key signal 的 IoU ≥ 60%）
  - seq 12 跟踪指标选择一致（10 次中 TOP 指标命中率高）
  - 字数/时长波动小（内容饱满但不冗余 · CV 低）

用法：
  PYTHONIOENCODING=utf-8 .venv/Scripts/python.exe scripts/analyze_store_manager_quality.py \\
    --compare export/ai-ab-test/round_a export/ai-ab-test/round_b \\
              export/ai-ab-test/round_v5 export/ai-ab-test/round_v5_1
"""
from __future__ import annotations

import argparse
import json
import re
import sys
from collections import Counter
from pathlib import Path
from statistics import mean, stdev


# ===== 核心业务关键词库（店长语汇） =====

# 挖掘深度关键信号
DEEP_SIGNALS = [
    # 储值卡负债动态
    ('消耗.*充值|充值.*消耗', '储值卡充消对比'),
    ('存量消费|复购乏力|复购.*收缩|复购.*减弱', '复购解读'),
    ('负债.*累积|兑付压力|负债.*减轻', '负债解读'),
    # 客单价双口径差值
    ('按成交收入.*按发生额|按发生额.*按成交收入', '客单价双口径对比'),
    ('每单.*让利|让利.*量级|让利.*金额', '让利量化'),
    # 业态归因
    ('业态|定位|散客|团购为主|车站|商场', '业态归因'),
    # 数据完整性
    ('数据.*缺失|数据.*完整|录入.*缺失|支出.*全.*0|支出.*为.*0|虚高风险|净利.*可信', '数据质量质疑'),
    # 协同恶化/结构失衡
    ('结构失衡|协同|多指标.*同|同向恶化', '结构性洞察'),
    # 时间规律深度
    ('周六.*周.*倍|周末.*工作日|旺淡日.*倍', '周规律倍率'),
    ('同周|同星期|同周基线', '基线识别'),
]

# 空洞表达（应被扣分）
HOLLOW_PATTERNS = [
    r'^关注\s*[^，。]+$',
    r'建议关注',
    r'加强\s*(\w+)?运营',
    r'提升\s*(\w+)?管理',
    r'需要重视',
    r'应当注意',
]

# 趋势词 · 必须同句内有 % 或绝对值数字（否则单期推测违规）
TREND_WORDS = ['下滑', '下降', '上升', '提升', '收缩', '萎缩', '承压', '走弱', '走强',
               '加剧', '恶化', '持续', '显著', '大幅', '明显', '锐减', '攀升']


def _iter_rounds(dir_path: Path) -> list[dict]:
    files = sorted(dir_path.glob('round_*.json'))
    out = []
    for f in files:
        data = json.loads(f.read_text(encoding='utf-8'))
        data['_file'] = f.name
        out.append(data)
    return out


def _get_seq(insights: list, seq: int) -> dict | None:
    for ins in insights:
        if isinstance(ins, dict) and ins.get('seq') == seq:
            return ins
    return None


def analyze_accuracy(insights: list) -> dict:
    """准确性评分（每项 0/1）。"""
    a_texts = ' '.join((_get_seq(insights, s) or {}).get('content') or '' for s in (1, 2))
    b_texts = ' '.join((_get_seq(insights, s) or {}).get('content') or '' for s in (3, 4))
    c_texts = ' '.join((_get_seq(insights, s) or {}).get('content') or '' for s in (5, 6))
    d_texts = ' '.join((_get_seq(insights, s) or {}).get('content') or '' for s in (7, 8))
    seq9 = (_get_seq(insights, 9) or {}).get('content') or ''
    seq10 = (_get_seq(insights, 10) or {}).get('content') or ''
    all_text = ' '.join((ins.get('content') or '') for ins in insights if isinstance(ins, dict))

    # 1. 对比口径显式引用（V5.1 H1 硬性要求）
    calib_explicit = bool(re.search(r'对比口径|同期对齐|同天数对齐|当期\s*\d+\s*天.*上期|\d+\s*天同期', a_texts))

    # 2. 权威字段引用（C 储值卡余额变化 / A 单位经济环比）
    authority_c = any(k in c_texts for k in ('期初', '期末', '余额变化', '其他调整'))
    authority_a = bool(re.search(r'[+-]?\d+\.?\d*%', a_texts))  # 有实际环比数字

    # 3. 规则合规
    no_industry_number = not bool(re.search(r'行业(警戒线|均值|标准|参考|基准|水平|经验值|通常).*\d+\.?\d*%?', all_text))
    no_manual_detail = not bool(re.search(r'(抹零|免单)\s*\d+[\.\d]*\s*元', b_texts))

    # 4. 单期推测违规计数（趋势词同句内是否有数字锚点）
    speculation_count = 0
    for sent in re.split(r'[。；\n]', all_text):
        if not sent.strip():
            continue
        has_trend_word = any(w in sent for w in TREND_WORDS)
        has_number = bool(re.search(r'[+-]?\d+\.?\d*%|\d{3,}元|\d+\s*元|\-\d+', sent))
        if has_trend_word and not has_number:
            # 允许"禁止推测"等元指令跳过
            if '推测' in sent or '禁' in sent or '不能' in sent:
                continue
            speculation_count += 1

    # 5. 数据完整性标注（D 板块 0 支出主动指出）
    data_integrity_flagged = any(k in d_texts for k in
                                 ('支出.*0', '数据缺失', '录入', '不完整', '虚高', '无法评估', '可信度'))

    # 6. 每条含数字
    ins_with_number = sum(1 for ins in insights if isinstance(ins, dict) and
                          re.search(r'\d+', ins.get('content') or ''))
    number_rate = ins_with_number / len(insights) if insights else 0

    return {
        'calib_explicit': int(calib_explicit),
        'authority_c': int(authority_c),
        'authority_a_env_bi': int(authority_a),
        'no_industry_number': int(no_industry_number),
        'no_manual_detail': int(no_manual_detail),
        'speculation_count': speculation_count,   # 越低越好
        'data_integrity_flagged': int(data_integrity_flagged),
        'number_rate': round(number_rate, 2),
    }


def analyze_depth(insights: list) -> dict:
    """洞察深度评分。"""
    all_text = ' '.join((ins.get('content') or '') for ins in insights if isinstance(ins, dict))

    # 1. 深度信号命中数（10 类里命中几类）
    deep_hits: list[str] = []
    for pattern, name in DEEP_SIGNALS:
        if re.search(pattern, all_text):
            deep_hits.append(name)
    deep_hit_count = len(deep_hits)

    # 2. 空洞表达计数（每条按句扫描）
    hollow_count = 0
    for ins in insights:
        if not isinstance(ins, dict):
            continue
        body = ins.get('content') or ''
        for pattern in HOLLOW_PATTERNS:
            if re.search(pattern, body):
                hollow_count += 1
                break

    # 3. seq 11 top 2 原因结构 + 每个原因是否有意义解读（不只数字堆砌）
    seq11 = (_get_seq(insights, 11) or {}).get('content') or ''
    has_r1 = bool(re.search(r'原因\s*1|1\s*[\)）\.、:：]|①|原因一', seq11))
    has_r2 = bool(re.search(r'原因\s*2|2\s*[\)）\.、:：]|②|原因二', seq11))
    f_top2 = int(has_r1 and has_r2)
    # 意义解读：原因文字里是否有解读词（"收缩/虚高/失衡/风险/无法评估"等因果意义）
    has_semantic = bool(re.search(r'收缩|虚高|失衡|风险|无法评估|可信|压力|乏力|不足|崩塌|疲态|暴露',
                                   seq11))
    f_top2_semantic = int(has_r1 and has_r2 and has_semantic)

    # 4. seq 12 跟踪 4 要素齐全（指标/阈值/节奏/动作）
    seq12 = (_get_seq(insights, 12) or {}).get('content') or ''
    has_indicator = bool(re.search(r'[储值卡余额|会员|客单价|成交收入|现金流入|现金流出|支出|占比|助教|订单]', seq12))
    has_rhythm = bool(re.search(r'(每[周月日天]|每\s*\d+|双周|旬|定期|周期性|连续)', seq12))
    has_threshold = bool(re.search(r'<\s*-?\d+|>\s*-?\d+|≥|≤|达到|跌破|超过|低于|目标', seq12))
    has_action = bool(re.search(r'(启动|触发|召回|发起|立即|核查|补录|校准|活动|赠金|审批|预算)', seq12))
    seq12_4elem = int(has_indicator and has_rhythm and has_threshold and has_action)

    # 5. 协同分析识别（seq 11 是否讲到多指标联动）
    collaborative = bool(re.search(r'(\+|与|及|和|协同|同时|叠加|共同|双收缩|双下降).*(\+|与|及|和)', seq11)) or \
                    bool(re.search(r'\d+.*\d+.*\d+', seq11))  # 至少 3 个数字说明多信号

    return {
        'deep_hit_count': deep_hit_count,            # 0-10
        'deep_hits': deep_hits,
        'hollow_count': hollow_count,                # 越低越好
        'f_top2_structure': f_top2,
        'f_top2_with_semantic': f_top2_semantic,
        'seq12_4elem_complete': seq12_4elem,
        'collaborative_analysis': int(collaborative),
    }


def analyze_stability_round(insights: list) -> dict:
    """单轮提取稳定性分析所需的"指纹"。"""
    # 1. 健康度灯色
    seq11 = (_get_seq(insights, 11) or {}).get('content') or ''
    light = 'unknown'
    if re.search(r'🔴|红灯', seq11):
        light = 'red'
    elif re.search(r'🟡|黄灯', seq11):
        light = 'yellow'
    elif re.search(r'🟢|绿灯', seq11):
        light = 'green'

    # 2. 关键 key signals（从 seq 11 提取命中的业务关键词集合）
    KEY_SIGNAL_VOCAB = [
        '会员占比', '会员订单占比',
        '客单价',
        '储值卡余额', '储值卡',
        '支出.*0', '支出缺失', '数据缺失',
        '复购',
        '优惠',
        '助教成本', '人力成本',
        '成交收入', '现金流入',
    ]
    signals = set()
    for kw in KEY_SIGNAL_VOCAB:
        if re.search(kw, seq11):
            signals.add(kw)

    # 3. seq 12 选取的跟踪指标
    seq12_title = (_get_seq(insights, 12) or {}).get('title') or ''
    seq12_body = (_get_seq(insights, 12) or {}).get('content') or ''
    tracking_indicator = 'unknown'
    for kw, label in [
        ('储值卡余额', '储值卡余额'),
        ('会员占比|会员订单占比', '会员占比'),
        ('客单价', '客单价'),
        ('现金流出|支出', '现金流出/支出'),
        ('成交收入', '成交收入'),
        ('助教|人力成本', '助教/人力成本'),
    ]:
        if re.search(kw, seq12_title + ' ' + seq12_body):
            tracking_indicator = label
            break

    # 4. 字数
    avg_len = 0
    if insights:
        avg_len = sum(len(ins.get('content') or '') for ins in insights if isinstance(ins, dict)) / len(insights)

    return {
        'light': light,
        'seq11_signals': signals,
        'tracking_indicator': tracking_indicator,
        'avg_content_len': round(avg_len, 1),
    }


def summarize(rounds_data: list[dict], label: str, perf: list[dict] | None = None) -> dict:
    """聚合 n 次的三层评分。

    rounds_data: round_*.json 原始数据列表（含 parsed）
    perf: 外部传入性能数据（duration_s, tokens），从 meta 里取
    """
    accuracies = []
    depths = []
    stab_fingerprints = []

    for rd in rounds_data:
        insights = (rd.get('parsed') or {}).get('insights') or []
        if not isinstance(insights, list) or not insights:
            continue
        accuracies.append(analyze_accuracy(insights))
        depths.append(analyze_depth(insights))
        stab_fingerprints.append(analyze_stability_round(insights))

    n = len(accuracies)
    if n == 0:
        return {'label': label, 'n': 0}

    # === 准确性聚合 ===
    acc_scores = {
        'calib_explicit_rate': mean(a['calib_explicit'] for a in accuracies),
        'authority_c_rate': mean(a['authority_c'] for a in accuracies),
        'authority_a_env_bi_rate': mean(a['authority_a_env_bi'] for a in accuracies),
        'no_industry_number_rate': mean(a['no_industry_number'] for a in accuracies),
        'no_manual_detail_rate': mean(a['no_manual_detail'] for a in accuracies),
        'speculation_avg': mean(a['speculation_count'] for a in accuracies),
        'data_integrity_flagged_rate': mean(a['data_integrity_flagged'] for a in accuracies),
        'number_rate_avg': mean(a['number_rate'] for a in accuracies),
    }

    # === 洞察深度聚合 ===
    depth_scores = {
        'deep_hit_avg': mean(d['deep_hit_count'] for d in depths),
        'deep_hit_union': len(set.union(*[set(d['deep_hits']) for d in depths])) if depths else 0,
        'hollow_avg': mean(d['hollow_count'] for d in depths),
        'f_top2_rate': mean(d['f_top2_structure'] for d in depths),
        'f_top2_with_semantic_rate': mean(d['f_top2_with_semantic'] for d in depths),
        'seq12_4elem_rate': mean(d['seq12_4elem_complete'] for d in depths),
        'collab_analysis_rate': mean(d['collaborative_analysis'] for d in depths),
    }

    # === 稳定性聚合（核心：众数占比 / 交并比 / CV）===
    # 灯色众数占比
    light_counter = Counter(s['light'] for s in stab_fingerprints)
    light_mode_rate = light_counter.most_common(1)[0][1] / n

    # seq 11 signals 交并比（跨 n 轮的平均 IoU）
    iou_scores = []
    for i in range(n):
        for j in range(i + 1, n):
            s1, s2 = stab_fingerprints[i]['seq11_signals'], stab_fingerprints[j]['seq11_signals']
            union = s1 | s2
            if not union:
                continue
            iou_scores.append(len(s1 & s2) / len(union))
    seq11_iou = mean(iou_scores) if iou_scores else 0

    # tracking indicator 一致性（众数占比）
    ti_counter = Counter(s['tracking_indicator'] for s in stab_fingerprints)
    tracking_mode_rate = ti_counter.most_common(1)[0][1] / n

    # 字数 CV（低为稳定）
    content_lens = [s['avg_content_len'] for s in stab_fingerprints]
    content_cv = (stdev(content_lens) / mean(content_lens)) if len(content_lens) > 1 and mean(content_lens) > 0 else 0

    # 性能 CV
    perf_data = perf or []
    durations = [p['duration_s'] for p in perf_data if p.get('duration_s') is not None]
    tokens = [p['tokens'] for p in perf_data if p.get('tokens') is not None]
    duration_cv = (stdev(durations) / mean(durations)) if len(durations) > 1 and mean(durations) > 0 else 0
    tokens_cv = (stdev(tokens) / mean(tokens)) if len(tokens) > 1 and mean(tokens) > 0 else 0

    stab_scores = {
        'light_mode': light_counter.most_common(1)[0][0],
        'light_mode_rate': light_mode_rate,
        'light_distribution': dict(light_counter),
        'seq11_signal_iou': seq11_iou,
        'tracking_mode': ti_counter.most_common(1)[0][0],
        'tracking_mode_rate': tracking_mode_rate,
        'content_len_cv': content_cv,
        'duration_cv': duration_cv,
        'tokens_cv': tokens_cv,
        'duration_mean': mean(durations) if durations else 0,
        'tokens_mean': mean(tokens) if tokens else 0,
    }

    # === 综合评分（0-100 百分制，三维加权）===
    # 准确性 40% + 洞察深度 35% + 稳定性 25%
    acc_composite = (
        acc_scores['calib_explicit_rate'] * 0.25 +
        acc_scores['authority_c_rate'] * 0.15 +
        acc_scores['authority_a_env_bi_rate'] * 0.10 +
        acc_scores['no_industry_number_rate'] * 0.15 +
        acc_scores['no_manual_detail_rate'] * 0.10 +
        max(0, 1 - acc_scores['speculation_avg'] / 5) * 0.15 +  # 5 次推测扣到 0
        acc_scores['data_integrity_flagged_rate'] * 0.10
    )
    depth_composite = (
        min(depth_scores['deep_hit_avg'] / 5, 1) * 0.30 +           # 深度信号 5 类以上得满
        max(0, 1 - depth_scores['hollow_avg'] / 3) * 0.15 +          # 空洞 3 次扣到 0
        depth_scores['f_top2_with_semantic_rate'] * 0.25 +
        depth_scores['seq12_4elem_rate'] * 0.20 +
        depth_scores['collab_analysis_rate'] * 0.10
    )
    stab_composite = (
        stab_scores['light_mode_rate'] * 0.30 +                      # 同灯色占比
        stab_scores['seq11_signal_iou'] * 0.25 +                     # 原因信号交并比
        stab_scores['tracking_mode_rate'] * 0.25 +                   # 跟踪指标一致
        max(0, 1 - stab_scores['content_len_cv'] * 2) * 0.10 +       # CV 0.5 扣到 0
        max(0, 1 - stab_scores['duration_cv'] * 2) * 0.10
    )
    overall = acc_composite * 0.4 + depth_composite * 0.35 + stab_composite * 0.25

    return {
        'label': label,
        'n': n,
        'accuracy': acc_scores,
        'depth': depth_scores,
        'stability': stab_scores,
        'composite': {
            'accuracy': round(acc_composite * 100, 1),
            'depth': round(depth_composite * 100, 1),
            'stability': round(stab_composite * 100, 1),
            'overall': round(overall * 100, 1),
        },
    }


def _load_perf(dir_path: Path) -> list[dict]:
    """从 round_XX.json 的 meta 提取 duration/tokens。"""
    out = []
    for f in sorted(dir_path.glob('round_*.json')):
        data = json.loads(f.read_text(encoding='utf-8'))
        meta = data.get('meta') or {}
        out.append({'duration_s': meta.get('duration_s'), 'tokens': meta.get('tokens')})
    return out


def print_summary(s: dict) -> None:
    if s['n'] == 0:
        print(f"\n=== {s['label'].upper()} 数据为空 ===")
        return
    c = s['composite']
    print(f"\n=== Round {s['label'].upper()} · 店长视角评分（n={s['n']}）===")
    print(f"  【综合评分】 {c['overall']:.1f} / 100")
    print(f"     准确性 {c['accuracy']:.1f}  |  洞察深度 {c['depth']:.1f}  |  稳定性 {c['stability']:.1f}")

    print(f"\n  -- 准确性明细 --")
    a = s['accuracy']
    print(f"    对比口径显式引用:      {a['calib_explicit_rate']:.0%}")
    print(f"    C 权威字段(余额变化):  {a['authority_c_rate']:.0%}")
    print(f"    A 权威字段(环比数字):  {a['authority_a_env_bi_rate']:.0%}")
    print(f"    禁行业数字合规:        {a['no_industry_number_rate']:.0%}")
    print(f"    禁手动调整拆明细合规:  {a['no_manual_detail_rate']:.0%}")
    print(f"    单期推测违规/次 (低好): {a['speculation_avg']:.1f}")
    print(f"    数据完整性标注:        {a['data_integrity_flagged_rate']:.0%}")

    print(f"\n  -- 洞察深度明细 --")
    d = s['depth']
    print(f"    深度信号命中/次 (满 10): {d['deep_hit_avg']:.1f}")
    print(f"    跨轮覆盖信号数:         {d['deep_hit_union']} / 10")
    print(f"    空洞表达/次 (低好):     {d['hollow_avg']:.1f}")
    print(f"    seq 11 top 2 结构:      {d['f_top2_rate']:.0%}")
    print(f"    seq 11 有意义解读:      {d['f_top2_with_semantic_rate']:.0%}")
    print(f"    seq 12 四要素齐全:      {d['seq12_4elem_rate']:.0%}")
    print(f"    多指标协同分析:         {d['collab_analysis_rate']:.0%}")

    print(f"\n  -- 稳定性明细 --")
    st = s['stability']
    print(f"    评级众数 ({st['light_mode']}) 占比:   {st['light_mode_rate']:.0%}  [{st['light_distribution']}]")
    print(f"    seq 11 原因信号 IoU:    {st['seq11_signal_iou']:.0%}")
    print(f"    跟踪指标众数 ({st['tracking_mode']}) 占比: {st['tracking_mode_rate']:.0%}")
    print(f"    字数 CV (低好):         {st['content_len_cv']:.2f}")
    print(f"    时长 CV (低好):         {st['duration_cv']:.2f}   均值 {st['duration_mean']:.1f}s")
    print(f"    tokens CV (低好):       {st['tokens_cv']:.2f}   均值 {st['tokens_mean']:.0f}")


def print_compare(*summaries: dict) -> None:
    labels = [s['label'].upper() for s in summaries]
    header = ' vs '.join(labels)
    col_w = 12
    print(f"\n======= {header} 店长视角综合评分 =======")
    print(f"{'指标':<34}" + ''.join(f"{l:>{col_w}}" for l in labels))
    print('-' * (34 + col_w * len(labels)))

    def _row(name: str, values: list, fmt: str = 'float', higher_better: bool = True) -> None:
        cells = []
        for v in values:
            if fmt == 'percent':
                cells.append(f'{v:.0%}')
            elif fmt == 'int':
                cells.append(str(v))
            elif fmt == 'float1':
                cells.append(f'{v:.1f}')
            else:
                cells.append(f'{v:.2f}')
        print(f"{name:<34}" + ''.join(f"{c:>{col_w}}" for c in cells))

    print('【综合】')
    _row('  总分 / 100',                [s['composite']['overall'] for s in summaries], 'float1')
    _row('  准确性 (40%)',              [s['composite']['accuracy'] for s in summaries], 'float1')
    _row('  洞察深度 (35%)',            [s['composite']['depth'] for s in summaries], 'float1')
    _row('  稳定性 (25%)',              [s['composite']['stability'] for s in summaries], 'float1')
    print('【准确性】')
    _row('  对比口径显式 ★',            [s['accuracy']['calib_explicit_rate'] for s in summaries], 'percent')
    _row('  C 权威字段',                [s['accuracy']['authority_c_rate'] for s in summaries], 'percent')
    _row('  数据完整性标注',             [s['accuracy']['data_integrity_flagged_rate'] for s in summaries], 'percent')
    _row('  单期推测违规/次 ↓',         [s['accuracy']['speculation_avg'] for s in summaries], 'float1')
    _row('  禁行业数字合规',             [s['accuracy']['no_industry_number_rate'] for s in summaries], 'percent')
    print('【洞察深度】')
    _row('  深度信号命中/次',            [s['depth']['deep_hit_avg'] for s in summaries], 'float1')
    _row('  跨轮信号覆盖 /10',          [s['depth']['deep_hit_union'] for s in summaries], 'int')
    _row('  空洞表达/次 ↓',             [s['depth']['hollow_avg'] for s in summaries], 'float1')
    _row('  seq 11 top 2 结构',         [s['depth']['f_top2_rate'] for s in summaries], 'percent')
    _row('  seq 11 有意义解读 ★',       [s['depth']['f_top2_with_semantic_rate'] for s in summaries], 'percent')
    _row('  seq 12 四要素齐全',          [s['depth']['seq12_4elem_rate'] for s in summaries], 'percent')
    _row('  多指标协同分析',             [s['depth']['collab_analysis_rate'] for s in summaries], 'percent')
    print('【稳定性（同数据下越一致越好）】')
    _row('  评级众数占比 ★',            [s['stability']['light_mode_rate'] for s in summaries], 'percent')
    _row('  seq 11 原因信号 IoU',       [s['stability']['seq11_signal_iou'] for s in summaries], 'percent')
    _row('  跟踪指标众数占比 ★',        [s['stability']['tracking_mode_rate'] for s in summaries], 'percent')
    _row('  字数 CV ↓',                [s['stability']['content_len_cv'] for s in summaries], 'float')
    _row('  时长 CV ↓',                [s['stability']['duration_cv'] for s in summaries], 'float')
    _row('  均时长 s',                  [s['stability']['duration_mean'] for s in summaries], 'float1')
    _row('  均 tokens',                 [s['stability']['tokens_mean'] for s in summaries], 'float1')
    print()
    for s in summaries:
        print(f"  {s['label'].upper()} 灯色分布: {s['stability']['light_distribution']}")


def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument('--dir', help='单目录分析')
    parser.add_argument('--compare', nargs='+', metavar='DIR', help='多目录对比（2-5 个）')
    args = parser.parse_args()

    if args.compare:
        dirs = [Path(d) for d in args.compare]
        summaries = []
        for d in dirs:
            rounds = _iter_rounds(d)
            if not rounds:
                print(f'[skip] {d} 无数据')
                continue
            perf = _load_perf(d)
            label = d.name.replace('round_', '')
            summaries.append(summarize(rounds, label, perf))
        for s in summaries:
            print_summary(s)
        print_compare(*summaries)
        tag = '_'.join(s['label'] for s in summaries)
        out_path = Path(f'export/ai-ab-test/_manager_quality_{tag}.json')
        out_path.write_text(json.dumps({s['label']: s for s in summaries}, ensure_ascii=False, indent=2, default=str),
                             encoding='utf-8')
        print(f'\n[done] 店长视角评分 JSON 已存: {out_path}')
    elif args.dir:
        rounds = _iter_rounds(Path(args.dir))
        perf = _load_perf(Path(args.dir))
        label = Path(args.dir).name.replace('round_', '')
        s = summarize(rounds, label, perf)
        print_summary(s)
    else:
        parser.error('需指定 --dir 或 --compare')


if __name__ == '__main__':
    main()