# ⚠️ 由 scripts/build-skills.mjs 从 shared/examples/brand.yaml 同步生成，不要直接改这里。
# 店铺品牌视觉规范 —— 所有生图技能读这一份，保证几百个 SKU 看起来像同一家店。
#   node scripts/brand.mjs --brand brand.yaml --for flat-lay
#   node scripts/gen.mjs --task flat-lay --brand brand.yaml --prompt '...'

brand:
  name: 示例品牌
  # 一句话概括调性，会原样进 prompt
  tone: quiet minimalist, warm and lived-in, never glossy or commercial

model:
  # 锁模特：给一张脸的参考图，所有技能都会把它作为最后一张参考图传入
  reference: assets/model/face-a.jpg
  description: East Asian woman, late twenties, natural makeup, shoulder-length black hair
  body: slim, height around 168cm

photography:
  background: seamless off-white studio backdrop, RGB 248 248 246
  lighting: soft large softbox from camera left, gentle fill, no hard shadows
  camera: 85mm equivalent, eye level, shallow depth of field
  grade: neutral white balance around 5200K, low contrast, slightly lifted blacks
  crop: full body with headroom, product centered

layout:
  # 给带排版的技能（主图 / 详情页）用
  margin: at least 8% empty margin on all sides
  typeface: clean sans-serif, no decorative fonts
  text_color: near-black on light background

forbid:
  - no visible brand logos other than the product's own
  - no text or watermark
  - no exaggerated poses or dramatic wind effects
  - no oversaturated colors

# 可选：把这些直接写进合规目标，生成时就按平台要求出图
compliance:
  platform: amazon

# `claude-sonnet-5` 参数清单

本技能默认用的模型的完整参数。日常只需要「参数约定」里那几个，
这份清单在需要用到非常规参数时再看。

**CRITICAL INSTRUCTION FOR AGENT**:
Run the `dlazy claude-sonnet-5` command to get results.

```bash
dlazy claude-sonnet-5 -h

Options:
  --prompt <prompt>     Prompt
  --images [images...]  Images [image: url or local path] (max 10)
  --videos [videos...]  Videos [video: url or local path] (max 8)
  --dry-run             Print payload without executing the tool
  --no-wait             Return generateId immediately for async tasks
  --timeout <seconds>   Max seconds to wait for async completion (default:
                        "1800")
  --input <jsonOrFile>  Inline JSON or @path/to/file.json — merged under flag
                        values (flags win)
  --save <path>         Download the result asset to this local path (mkdir +
                        retry handled for you). A destination path — NOT a
                        response format; for stdout shape use --format
  --batch <n>           Fan-out N parallel runs (cloud tools only) (default:
                        "1")
  -h, --help            display help for command
```

> Any flag also accepts pipe references — `-` (auto-pick from upstream stdin), `@N` (n-th output), `@N.path` (jsonpath into output), `@*` (all primary values), `@stdin` / `@stdin:path` (whole envelope). See `dlazy --help` for details.

---

换其他后端时参数由 `scripts/gen.mjs` 统一翻译，见 [`provider-cli.md`](provider-cli.md)。

<!-- 由 scripts/build-skills.mjs 从 shared/references/platform-specs.md 同步生成，不要直接改这里。 -->
# 平台图片规格（可机检子集）

`scripts/check_listing.py` 内置的规则来源与口径。**平台规则会变，以各平台最新官方文档为准**；
需要覆盖时写一份 JSON 用 `--rules` 传入，结构与下表字段一一对应。

---

## 字段含义

| 字段 | 含义 |
| --- | --- |
| `pure_white_bg` | 是否要求纯白 RGB(255,255,255) 背景 |
| `bg_tolerance` | 判定「纯白」允许的单通道偏差 |
| `bg_coverage` | 边缘一圈需要有多大比例落在容差内 |
| `min_long_side` / `recommend_long_side` / `max_long_side` | 最长边像素 |
| `min_occupancy` | 商品包围盒面积 ÷ 画面面积 的下限 |
| `allow_alpha` | 是否允许透明通道 |
| `allow_border` | 是否允许描边 / 外框 |
| `formats` / `max_bytes` / `aspect` | 允许格式、体积上限、允许比例 |

---

## 内置规则

| 平台 | 纯白底 | 最长边（下限 / 建议） | 主体占比 | 比例 | 体积上限 |
| --- | --- | --- | --- | --- | --- |
| `amazon` | 是 | 1000 / 1600 | ≥ 85% | 不限 | 10 MB |
| `tiktok-shop` | 否 | 800 / 1600 | ≥ 60% | 1:1 或 3:4 | 5 MB |
| `temu` | 是 | 800 / 1350 | ≥ 70% | 1:1 | 3 MB |
| `shopee` | 否 | 500 / 1024 | ≥ 55% | 1:1 | 2 MB |
| `shopify` | 否 | 1024 / 2048 | 不限 | 不限 | 20 MB |
| `taobao` | 是 | 800 / 1200 | ≥ 70% | 1:1 | 3 MB |

---

## 几个容易踩的点

- **透明 PNG**：Amazon 会把透明像素转成黑色。永远压平成白底 JPEG 再传。
- **「白底」不等于「看起来是白的」**：棚拍的浅灰墙（约 RGB 208）肉眼像白，机检直接判不合格。
- **主体占比**：留白过多是最常见的驳回原因之一，比分辨率不够更常见。
- **文字 / 水印 / 拼图**：像素层测不了，交给 `detect-task` 的视觉模型或人工。
- **自动修复的边界**：`--fix` 能压白底、按占比重构画布、补分辨率、压体积；
  它**不会**修图，也不会去水印——那是 `item-repair` 和 `remove-watermark` 的活。

<!-- 由 scripts/build-skills.mjs 从 shared/references/provider-cli.md 同步生成，不要直接改这里。 -->
# 后端调用参考

技能正文只写「要生成什么」。认证、计费、错误码、输出结构这些每个技能都一样的东西放在这里，
**用到时再读**，不占技能的常驻上下文。

---

## 一、认证

### 默认后端 dLazy

```bash
dlazy login            # 设备码流程，远程 shell 也能用，自动写入本地配置
dlazy auth set <KEY>   # 已有 key 时直接写入
```

key 存在用户配置目录（macOS/Linux `~/.dlazy/config.json`，Windows `%USERPROFILE%\.dlazy\config.json`），
权限限本机用户。也可以每次调用用环境变量 `DLAZY_API_KEY` 传入。

手动获取：登录 [dlazy.com](https://dlazy.com) → [API Key 页面](https://dlazy.com/dashboard/organization/api-key)。
key 按组织隔离，可随时轮换或吊销。

### 其他后端

本技能库不锁定单一厂商。配好任意一家的 key 即可跑：

| 后端 | 环境变量 | 说明 |
| --- | --- | --- |
| `dlazy` | `dlazy login` 或 `DLAZY_API_KEY` | 默认，最省事 |
| `openai` | `OPENAI_API_KEY` | 走 `/v1/images/edits` 与 `/v1/images/generations` |
| `gemini` | `GEMINI_API_KEY` | Nano Banana 系列 |
| `fal` | `FAL_KEY` | |
| `replicate` | `REPLICATE_API_TOKEN` | |
| `ark` | `ARK_API_KEY` + `ARK_MODEL` | 火山方舟，模型 ID 需按开通情况填 |

选路优先级：`--provider` 参数 > `PROVIDER` 环境变量 > 第一个配了 key 的 > `dlazy`。

```bash
node scripts/gen.mjs --doctor     # 看当前哪个后端可用
```

各后端的模型 ID 可用 `GEN_MODEL_OPENAI` / `GEN_MODEL_GEMINI` / `GEN_MODEL_FAL` /
`GEN_MODEL_REPLICATE` / `GEN_MODEL_ARK` 覆盖。**厂商目录会变，以各家最新文档为准。**

---

## 二、两种调用方式

### 方式 A：统一入口（推荐）

```bash
node scripts/gen.mjs --task <技能名> --prompt '...' --images a.jpg b.jpg --save out.jpg
```

它负责：后端选路、默认尺寸档位、失败重试（429/5xx 指数退避）、落盘建目录、成本估算。

```bash
node scripts/gen.mjs --task flat-lay --prompt '...' --dry-run   # 不调用不计费，只看要发什么
node scripts/gen.mjs --help
```

### 方式 B：直接用 dLazy CLI

不想引入 Node 依赖时，技能正文里的 `dlazy ...` 命令可以原样执行，效果等价。

```bash
npx @dlazy/cli@1.2.3 <command>     # 不装全局二进制
```

- CLI 源码：[github.com/dlazy-ai/cli](https://github.com/dlazy-ai/cli) · npm 包 `@dlazy/cli`

---

## 三、数据流向

调用 dLazy 时：提示词与参数发往 `api.dlazy.com`；传入的本地图片会上传到 `files.dlazy.com`
供模型读取；产出 URL 同样托管在 `files.dlazy.com`。这是云端生成 API 的通用形态。
换成其他后端时，数据流向对应厂商，不经过 dLazy。

---

## 四、输出结构

`gen.mjs`（加 `--json`）：

```json
{
  "ok": true,
  "task": "flat-lay",
  "provider": "dlazy",
  "model": "gpt-image-2",
  "files": ["docs/flat-lay/output-sku001.jpg"],
  "texts": [],
  "estimatedCredits": 60,
  "elapsedMs": 58213
}
```

dLazy CLI 原生：

```json
{
  "ok": true,
  "result": {
    "tool": "gpt-image-2",
    "data": { "urls": ["https://files.dlazy.com/data/ai/....jpg"] },
    "savedPath": "docs/flat-lay/example-output.jpg"
  }
}
```

加 `--no-wait` 的异步任务不返回 `data`，返回 `task: { generateId, status }`，
用 `dlazy status <generateId> --wait` 轮询。

文本类模型（如质检）产出在 `result.data.texts[0]`：

```bash
dlazy claude-sonnet-5 --prompt '...' --images x.jpg \
  | python3 -c 'import sys,json;print(json.load(sys.stdin)["result"]["data"]["texts"][0])'
```

---

## 五、错误处理

| Code | 类型 | 示例 |
| --- | --- | --- |
| 401 | 未授权 / 无 key | `ok: false, code: "unauthorized"` |
| 501 | 缺必填参数 | `error: required option '--prompt <prompt>' not specified` |
| 502 | 本地文件读不到 | `Error: Image file not found: ...` |
| 503 | 余额不足 | `ok: false, code: "insufficient_balance"` |
| 503 | 服务端错误 | `HTTP status code error (500)` |
| 504 | 异步任务失败 | `=== Generation Failed ===` / `Prompt violates safety policy` |

**给 Agent 的硬性要求**

1. 命中 `insufficient_balance` → 明确告诉用户算力不足，并给出充值入口
   <https://dlazy.com/dashboard/organization/settings?tab=credits>
2. 命中 `unauthorized` / 缺 key → 告诉用户去 <https://dlazy.com/dashboard/organization/api-key>
   取 key，用 `dlazy auth set <key>` 存好再继续。
3. 用 `gen.mjs` 时，429 与 5xx 已自动重试；仍失败才向用户报错。
4. **不要**为了「跑通」而偷偷降级参数（尺寸、档位、批量），先问用户。

#!/usr/bin/env python3
# ⚠️ 由 scripts/build-skills.mjs 从 shared/scripts/check_listing.py 同步生成，不要直接改这里。
"""
check_listing.py —— 上架前的客观合规校验。

detect-task 用视觉模型判断「像不像真的」，这个脚本判断「传上去会不会被驳回」。
两者互补：一个主观、一个客观；一个要花算力、一个只读像素。

  python3 scripts/check_listing.py main.jpg --platform amazon
  python3 scripts/check_listing.py imgs/*.jpg --platform amazon --json
  python3 scripts/check_listing.py raw.png --platform amazon --fix out/

退出码：0 全部通过（含仅警告）；1 存在驳回风险；2 参数或读图错误。

⚠️ 平台规则会变。本文件内置的是可机检子集，以各平台最新官方文档为准；
   需要覆盖时用 --rules your-rules.json（结构见 references/platform-specs.md）。
"""
from __future__ import annotations

import argparse
import json
import os
import sys

try:
    from PIL import Image, ImageChops
except ImportError:
    sys.exit("需要 Pillow：pip install Pillow")

# ---------------------------------------------------------------- 规则集

RULES = {
    "amazon": {
        "label": "Amazon 主图",
        "pure_white_bg": True, "bg_tolerance": 2, "bg_coverage": 0.98,
        "min_long_side": 1000, "recommend_long_side": 1600, "max_long_side": 10000,
        "min_occupancy": 0.85, "allow_alpha": False, "allow_border": False,
        "formats": ["JPEG", "PNG", "TIFF", "GIF"], "max_bytes": 10 * 1024**2,
        "aspect": None,
        "notes": "主图禁文字/logo/水印/拼图；透明 PNG 会被转黑底，必须压平为白底 JPEG。",
    },
    "tiktok-shop": {
        "label": "TikTok Shop 主图",
        "pure_white_bg": False, "bg_tolerance": 8, "bg_coverage": 0.90,
        "min_long_side": 800, "recommend_long_side": 1600, "max_long_side": 8000,
        "min_occupancy": 0.60, "allow_alpha": False, "allow_border": False,
        "formats": ["JPEG", "PNG"], "max_bytes": 5 * 1024**2,
        "aspect": [(1, 1), (3, 4)],
        "notes": "禁边框与水印；1:1 或 3:4。",
    },
    "temu": {
        "label": "Temu 主图",
        "pure_white_bg": True, "bg_tolerance": 6, "bg_coverage": 0.95,
        "min_long_side": 800, "recommend_long_side": 1350, "max_long_side": 8000,
        "min_occupancy": 0.70, "allow_alpha": False, "allow_border": False,
        "formats": ["JPEG", "PNG"], "max_bytes": 3 * 1024**2,
        "aspect": [(1, 1)],
        "notes": "主图须方图、干净背景、无促销文字。",
    },
    "shopee": {
        "label": "Shopee 主图",
        "pure_white_bg": False, "bg_tolerance": 10, "bg_coverage": 0.85,
        "min_long_side": 500, "recommend_long_side": 1024, "max_long_side": 8000,
        "min_occupancy": 0.55, "allow_alpha": False, "allow_border": True,
        "formats": ["JPEG", "PNG"], "max_bytes": 2 * 1024**2,
        "aspect": [(1, 1)],
        "notes": "方图；单张 ≤2MB。",
    },
    "shopify": {
        "label": "Shopify 商品图",
        "pure_white_bg": False, "bg_tolerance": 12, "bg_coverage": 0.0,
        "min_long_side": 1024, "recommend_long_side": 2048, "max_long_side": 20000,
        "min_occupancy": 0.0, "allow_alpha": True, "allow_border": True,
        "formats": ["JPEG", "PNG", "WEBP"], "max_bytes": 20 * 1024**2,
        "aspect": None,
        "notes": "平台无硬性约束；2048×2048 方图便于缩放与放大镜。",
    },
    "taobao": {
        "label": "淘宝 / 1688 主图",
        "pure_white_bg": True, "bg_tolerance": 6, "bg_coverage": 0.92,
        "min_long_side": 800, "recommend_long_side": 1200, "max_long_side": 8000,
        "min_occupancy": 0.70, "allow_alpha": False, "allow_border": False,
        "formats": ["JPEG", "PNG"], "max_bytes": 3 * 1024**2,
        "aspect": [(1, 1)],
        "notes": "主图方图；白底图另有独立坑位要求。",
    },
}

PASS, WARN, FAIL = "pass", "warn", "fail"
MARK = {PASS: "✓", WARN: "!", FAIL: "✗"}

# ---------------------------------------------------------------- 度量

def border_ring(img, frac=0.02):
    """取四边一圈像素，用来判定背景色。"""
    w, h = img.size
    d = max(1, int(min(w, h) * frac))
    px = img.load()
    out = []
    for y in list(range(d)) + list(range(h - d, h)):
        for x in range(0, w, max(1, w // 200)):
            out.append(px[x, y])
    for x in list(range(d)) + list(range(w - d, w)):
        for y in range(0, h, max(1, h // 200)):
            out.append(px[x, y])
    return out


def bg_stats(img, tolerance):
    """返回 (背景基色, 纯白比例, 与基色一致的比例)。"""
    ring = border_ring(img)
    if not ring:
        return (255, 255, 255), 0.0, 0.0
    white = sum(1 for p in ring if all(abs(c - 255) <= tolerance for c in p[:3]))
    # 众数近似：把颜色量化到 8 级再统计
    buckets = {}
    for p in ring:
        k = tuple(c // 32 for c in p[:3])
        buckets[k] = buckets.get(k, 0) + 1
    mode_k = max(buckets, key=buckets.get)
    members = [p[:3] for p in ring if tuple(c // 32 for c in p[:3]) == mode_k]
    base = tuple(round(sum(c[i] for c in members) / len(members)) for i in range(3))
    same = sum(1 for p in ring
               if sum((a - b) ** 2 for a, b in zip(p[:3], base)) ** 0.5 <= 24)
    return base, white / len(ring), same / len(ring)


def subject_bbox(img, base, thresh=18):
    """相对背景基色做差，得到主体包围盒与占位面积。"""
    w, h = img.size
    bg = Image.new("RGB", img.size, tuple(int(c) for c in base))
    diff = ImageChops.difference(img.convert("RGB"), bg).convert("L")
    mask = diff.point(lambda v: 255 if v > thresh else 0)
    box = mask.getbbox()
    if not box:
        return None, 0.0, 0.0
    ink = sum(mask.histogram()[1:]) / float(w * h)
    x0, y0, x1, y1 = box
    return box, ((x1 - x0) * (y1 - y0)) / float(w * h), ink


def has_border(img, base):
    """最外一圈是否是一条与背景不同的均匀细边。"""
    w, h = img.size
    if min(w, h) < 20:
        return False
    px = img.convert("RGB").load()
    ring = [px[x, 0] for x in range(0, w, max(1, w // 100))] + \
           [px[x, h - 1] for x in range(0, w, max(1, w // 100))]
    if not ring:
        return False
    far = sum(1 for p in ring
              if sum((a - b) ** 2 for a, b in zip(p, base)) ** 0.5 > 40)
    return far / len(ring) > 0.9


def ratio_ok(w, h, allowed, tol=0.02):
    if not allowed:
        return True, f"{w}:{h}"
    r = w / h
    for a, b in allowed:
        if abs(r - a / b) <= tol:
            return True, f"{a}:{b}"
    return False, f"{r:.3f}"

# ---------------------------------------------------------------- 校验

def check(path, rules):
    try:
        img = Image.open(path)
        img.load()
    except Exception as e:  # noqa: BLE001
        return {"file": path, "error": str(e), "checks": [], "verdict": "error"}

    fmt = img.format
    size_bytes = os.path.getsize(path)
    w, h = img.size
    has_alpha = img.mode in ("RGBA", "LA", "PA") or "transparency" in img.info
    rgb = img.convert("RGB")
    base, white_frac, same_frac = bg_stats(rgb, rules["bg_tolerance"])
    box, occ, ink = subject_bbox(rgb, base)

    out = []

    def add(name, level, detail, measured=None, required=None):
        out.append({"check": name, "level": level, "detail": detail,
                    "measured": measured, "required": required})

    # 尺寸
    long_side = max(w, h)
    if long_side < rules["min_long_side"]:
        add("分辨率", FAIL, f"最长边 {long_side}px，低于下限", long_side, rules["min_long_side"])
    elif long_side < rules["recommend_long_side"]:
        add("分辨率", WARN, f"最长边 {long_side}px，达标但不足以触发放大镜",
            long_side, rules["recommend_long_side"])
    elif long_side > rules["max_long_side"]:
        add("分辨率", FAIL, f"最长边 {long_side}px 超上限", long_side, rules["max_long_side"])
    else:
        add("分辨率", PASS, f"{w}×{h}", long_side, rules["recommend_long_side"])

    # 比例
    ok, got = ratio_ok(w, h, rules["aspect"])
    add("画面比例", PASS if ok else FAIL,
        f"{w}:{h}" if ok else f"实际 {got}，不在允许比例内", got,
        rules["aspect"] and "/".join(f"{a}:{b}" for a, b in rules["aspect"]))

    # 背景
    if rules["pure_white_bg"]:
        need = rules["bg_coverage"]
        if white_frac >= need:
            add("纯白背景", PASS, f"边缘 {white_frac:.1%} 为 RGB(255,255,255)±{rules['bg_tolerance']}",
                round(white_frac, 4), need)
        else:
            add("纯白背景", FAIL,
                f"边缘仅 {white_frac:.1%} 为纯白，背景基色约 RGB{tuple(int(c) for c in base)}",
                round(white_frac, 4), need)
    else:
        add("背景一致性", PASS if same_frac >= rules["bg_coverage"] else WARN,
            f"边缘 {same_frac:.1%} 颜色一致（基色 RGB{tuple(int(c) for c in base)}）",
            round(same_frac, 4), rules["bg_coverage"])

    # 主体占比
    if rules["min_occupancy"] > 0:
        if occ >= rules["min_occupancy"]:
            add("主体占比", PASS, f"包围盒占画面 {occ:.1%}", round(occ, 4), rules["min_occupancy"])
        else:
            add("主体占比", FAIL,
                f"包围盒仅占 {occ:.1%}，商品太小（留白过多）", round(occ, 4), rules["min_occupancy"])

    # 透明通道
    if has_alpha and not rules["allow_alpha"]:
        add("透明通道", FAIL, "含 alpha；上传后透明像素可能被转为黑色，须压平为白底", True, False)
    else:
        add("透明通道", PASS, "无 alpha" if not has_alpha else "允许 alpha", has_alpha, rules["allow_alpha"])

    # 边框
    if not rules["allow_border"]:
        b = has_border(rgb, base)
        add("边框", FAIL if b else PASS,
            "检出一圈与背景不同的描边" if b else "无描边", b, False)

    # 格式与体积
    add("文件格式", PASS if fmt in rules["formats"] else FAIL,
        f"{fmt}", fmt, "/".join(rules["formats"]))
    add("文件体积", PASS if size_bytes <= rules["max_bytes"] else FAIL,
        f"{size_bytes / 1024**2:.2f} MB", size_bytes, rules["max_bytes"])

    # 色彩模式
    add("色彩模式", PASS if img.mode in ("RGB", "RGBA", "L", "P") else WARN,
        img.mode, img.mode, "RGB")

    # 需模型判定的项 —— 不假装能测
    add("文字 / 水印 / 拼图", "manual",
        "像素层判不了，交给 detect-task 或人工过一眼", None, None)

    fails = [c for c in out if c["level"] == FAIL]
    warns = [c for c in out if c["level"] == WARN]
    verdict = "reject-risk" if fails else ("warn" if warns else "pass")
    return {
        "file": path, "platform": rules["label"], "verdict": verdict,
        "width": w, "height": h, "format": fmt, "bytes": size_bytes,
        "bgBase": [int(c) for c in base], "whiteFraction": round(white_frac, 4),
        "occupancy": round(occ, 4), "inkCoverage": round(ink, 4), "bbox": box,
        "checks": out, "notes": rules["notes"],
    }

# ---------------------------------------------------------------- 自动修复

def fix(path, rules, outdir):
    """压白底 → 按目标占比裁切 → 补足分辨率 → 存合规 JPEG。"""
    img = Image.open(path)
    img.load()
    if img.mode in ("RGBA", "LA", "PA") or "transparency" in img.info:
        rgba = img.convert("RGBA")
        canvas = Image.new("RGB", rgba.size, (255, 255, 255))
        canvas.paste(rgba, mask=rgba.split()[-1])
        img = canvas
    else:
        img = img.convert("RGB")

    base, _, _ = bg_stats(img, rules["bg_tolerance"])
    box, occ, _ = subject_bbox(img, base)

    if rules["pure_white_bg"] and tuple(int(c) for c in base) != (255, 255, 255):
        # 把接近基色的像素刷成纯白，保留主体
        bg = Image.new("RGB", img.size, tuple(int(c) for c in base))
        diff = ImageChops.difference(img, bg).convert("L")
        mask = diff.point(lambda v: 255 if v > 18 else 0)
        white = Image.new("RGB", img.size, (255, 255, 255))
        img = Image.composite(img, white, mask)
        base = (255, 255, 255)

    target = rules["min_occupancy"] or 0.85
    if box and occ < target:
        x0, y0, x1, y1 = box
        bw, bh = x1 - x0, y1 - y0
        crop = img.crop(box)
        square = bool(rules["aspect"]) and (1, 1) in rules["aspect"]
        if square:
            # 方图：包围盒非正方时，占比的物理上限是 bw*bh/max^2
            side = max(int(round((bw * bh / target) ** 0.5)), bw, bh)
            cw = ch = side
        else:
            # 自由比例：画布跟随包围盒长宽比，占比可精确落在 target
            k = target ** 0.5
            cw, ch = max(int(round(bw / k)), bw), max(int(round(bh / k)), bh)
        canvas = Image.new("RGB", (cw, ch), tuple(int(c) for c in base))
        canvas.paste(crop, ((cw - bw) // 2, (ch - bh) // 2))
        img = canvas

    need = rules["recommend_long_side"]
    if max(img.size) < need:
        s = need / max(img.size)
        img = img.resize((round(img.width * s), round(img.height * s)), Image.LANCZOS)

    os.makedirs(outdir, exist_ok=True)
    stem = os.path.splitext(os.path.basename(path))[0]
    dst = os.path.join(outdir, f"{stem}-fixed.jpg")
    q = 92
    img.save(dst, "JPEG", quality=q, subsampling=0, optimize=True)
    while os.path.getsize(dst) > rules["max_bytes"] and q > 60:
        q -= 8
        img.save(dst, "JPEG", quality=q, subsampling=0, optimize=True)
    return dst

# ---------------------------------------------------------------- 输出

def render(reports):
    lines = []
    for r in reports:
        if r.get("error"):
            lines.append(f"✗ {r['file']} —— 读图失败：{r['error']}")
            continue
        head = {"pass": "通过", "warn": "通过（有提醒）", "reject-risk": "有驳回风险"}[r["verdict"]]
        lines.append(f"\n{r['file']}  ·  {r['platform']}  ·  {head}")
        lines.append(f"  {r['width']}×{r['height']} {r['format']} {r['bytes']/1024**2:.2f}MB"
                     f" · 背景 RGB{tuple(r['bgBase'])} · 主体占比 {r['occupancy']:.1%}")
        lines.append("")
        for c in r["checks"]:
            mark = MARK.get(c["level"], "?")
            lines.append(f"  {mark} {c['check']:<16} {c['detail']}")
        lines.append(f"\n  平台备注：{r['notes']}")
    return "\n".join(lines)


def main():
    ap = argparse.ArgumentParser(description="上架前客观合规校验")
    ap.add_argument("images", nargs="+")
    ap.add_argument("--platform", default="amazon",
                    help="/".join(RULES) + "（默认 amazon）")
    ap.add_argument("--rules", help="自定义规则 JSON，覆盖内置规则集")
    ap.add_argument("--json", action="store_true")
    ap.add_argument("--fix", metavar="OUTDIR", help="自动修复并输出到该目录")
    a = ap.parse_args()

    table = dict(RULES)
    if a.rules:
        with open(a.rules, encoding="utf8") as f:
            for k, v in json.load(f).items():
                table[k] = {**table.get(k, RULES["amazon"]), **v}
    if a.platform not in table:
        sys.exit(f"未知平台 {a.platform}，可选：{'/'.join(table)}")
    rules = table[a.platform]

    reports = [check(p, rules) for p in a.images]

    if a.fix:
        for r in reports:
            if r.get("error"):
                continue
            r["fixed"] = fix(r["file"], rules, a.fix)
        reports = [check(r["fixed"], rules) | {"source": r["file"]}
                   if r.get("fixed") else r for r in reports]

    if a.json:
        print(json.dumps(reports, ensure_ascii=False, indent=2))
    else:
        print(render(reports))

    sys.exit(1 if any(r.get("verdict") == "reject-risk" for r in reports) else 0)


if __name__ == "__main__":
    main()