@wji.ai/cli 0.9.1 → 0.9.4

package/skill/short-drama/SKILL.md

@@ -36,8 +36,11 @@ Either way, this skill only says **what to build**; the operations doc says how
 
 **How the drama is written** comes from `{baseDir}/reference/craft-story.md` (the 短剧 method:
 brief, three phases, 分集目录, 爽点 / 钩子 libraries, the episode formula, the character bible) and
-`{baseDir}/reference/dialogue-craft.md` (lines as action, 语体 rules, the polish pass). Read the
+`{baseDir}/reference/dialogue-craft.md` (lines as action, 语体 rules, the dialogue review). Read the
 first before stage 1 and the second before stage 4. The gates encode what those two can check.
+The full dialogue method is the user's **`w-film-dialogue-editor`** skill, installed beside this
+one: its language guides and corpus are read while the lines are written and again in the required
+dialogue review after the script.
 
 **Re-read the guideline.** The 创作指引 (`render guideline`) is the drama's top rule document.
 Print it before writing any character, any episode and any storyboard, and print the speakers'
@@ -277,6 +280,7 @@ Read `{baseDir}/reference/stage-script.md`.
    writes `episodes/E0N.json` with target seconds, hook/cliff, claimed beats, the episode design
    (`function` / `task` / `change` / `hookType`, `satisfactionBeat` to set) and a `seedNote`
    (synopsis, the 爽点 to realise, register, guideline priorities, every speaker's `bible.speech`,
+   the dialogue skill's guides and per-speaker corpus queries in `dialogue`,
    ids, previous cliff, budget). Fill `scenes[].flow` — action beats and line beats alternating,
    **common actions only**, hook image inside the first 3 beats and by 15 s, a `turn: true` beat
    at least every 30 s, the 爽点 on an action (`satisfactionBeat`), cliff as the last beat; lines
@@ -286,9 +290,13 @@ Read `{baseDir}/reference/stage-script.md`.
    Delete `seedNote` when done.
 2. `validate --stage script --ep N` — duration math, line length, speakers, hook window, the
    rhythm, the design, filmability, upstream ids; the register gate only advises. Over budget: cut
-   actions first; under: add conflict, not small talk. When dialogue is the weakness, run the
-   polish pass: `render dialogue --ep N` → fill the 逐句校对表 → apply → re-validate.
-3. Sync: `render script --ep N` (the 短剧 episode format with the 创作自检 table) → `story/` dataset
+   actions first; under: add conflict, not small talk.
+3. **Dialogue review** (required, gate `dialogue-review`): `seed dialogue --ep N` → read the
+   `w-film-dialogue-editor` framework and the guide of each 语体 it prints, query the corpus per
+   speaker → fill the diagnosis, level, one row per line (issue / revised / reason / playable
+   note) and the guide's 自检 → `apply dialogue --ep N` → re-validate. A line edited later is
+   reviewed again. Details in `stage-script.md`.
+4. Sync: `render script --ep N` (the 短剧 episode format with the 创作自检 table) → `story/` dataset
    `E0N 剧本` (type `screenplay`, `viewMode=content`); bind `wji.screenplayDatasetId` with `--ep N`.
 
 ## Stage 5 — Storyboard (per episode)
@@ -307,6 +315,21 @@ prompt, alongside the camera package and each speaker's delivery.
 
 **Two people talking** (gate `two-person-framing`, sheets mode): a two-shot once, then over-the-shoulder (`ots`) for the exchange, a clean single for the punchline. A single names only its person; every look states a screen direction (`eyeline`); the sides stay locked. See `reference/craft-shots.md`.
 
+**Every picture says what it is, and what NOT to copy from it** (gate `ref-binding`, sheets mode).
+The model reads one string: a legend line that only says "a character sheet" for two different
+people is how faces get swapped, and a shot that never names its `@imageK` is a shot where the
+model invents the face or the prop. So — one legend line per picture, naming a person by age /
+hair / clothes from the character bible, the room by its name, a prop by its name; every picture
+that is **not** a person also says what may not be copied from it (*take only its light ratios,
+contrast and colour temperature — never its composition or its people*); the room is named once
+for the whole segment (`Every shot stays inside @image1 …`); every person and every prop is bound
+by its label again **inside its own shot**; no connected picture may go unused; and a segment may
+not carry more pictures than the model takes (`refs` cap). Wording patterns and the six roles:
+`reference/reference-binding.md`. A prompt that was written or edited by hand (or copied from a
+production run) can be proofread before any image is spent:
+`node {baseDir}/reference/prompt-audit.mjs --project <project> --ep 1 --segment E0N-k --prompt <file>`
+— it reports per-shot unbound people and props, indistinguishable legend lines and unused pictures.
+
 **Every cut is lit and composed in the preset's look.** Read the preset's guide
 (`{baseDir}/reference/cinematography-<preset>.md`) before the first storyboard. Name the source
 in `light.source`, stage the person against it, and choose the camera side that gives the

package/skill/short-drama/examples/night-shift/episodes/E01.json

@@ -1879,5 +1879,177 @@
     "canvasId": "",
     "shotFolderId": "",
     "displayNodes": {}
+  },
+  "dialogueReview": {
+    "level": "轻修",
+    "guides": [
+      "lang-zh-cn.md"
+    ],
+    "diagnosis": {
+      "goal": "小夏想守住别人托付的袋子；老陈想弄清袋子从哪来",
+      "conflict": "老陈步步试探，小夏用「别人的东西」挡住",
+      "subtext": "老陈认得穿雨衣的人，却不说破；小夏开始怀疑他",
+      "voice": "小夏短句、收得急；老陈慢、一字一顿，遮住名字也分得出",
+      "rhythm": "问答都短，每两三句就被动作打断，节奏成立"
+    },
+    "rows": [
+      {
+        "scene": 1,
+        "beat": 2,
+        "speaker": "C03",
+        "original": "帮我收着。明天有人来拿。",
+        "issue": "无：两个短句，交代托付又不解释",
+        "revised": "帮我收着。明天有人来拿。",
+        "reason": "",
+        "note": "压着嗓子，你不是请求，是把麻烦塞给她"
+      },
+      {
+        "scene": 1,
+        "beat": 3,
+        "speaker": "C01",
+        "original": "你是谁？我们不代收——",
+        "issue": "无：被打断的半句，正好是人已走远",
+        "revised": "你是谁？我们不代收——",
+        "reason": "",
+        "note": "追问，话到一半发现对方根本不听"
+      },
+      {
+        "scene": 2,
+        "beat": 3,
+        "speaker": "C02",
+        "original": "又是你值夜班。",
+        "issue": "无：熟人寒暄，带着她天天值夜的信息",
+        "revised": "又是你值夜班。",
+        "reason": "",
+        "note": "先吃一口再说，你在装作若无其事"
+      },
+      {
+        "scene": 2,
+        "beat": 4,
+        "speaker": "C01",
+        "original": "陈叔，你怎么天天半夜来。",
+        "issue": "无：顺口一问，眼睛在袋子上，潜台词在动作里",
+        "revised": "陈叔，你怎么天天半夜来。",
+        "reason": "",
+        "note": "嘴上问他，心思全在袋子上，别看他"
+      },
+      {
+        "scene": 2,
+        "beat": 5,
+        "speaker": "C02",
+        "original": "老了，睡不着。",
+        "issue": "无：一句带过，把真正的原因藏住",
+        "revised": "老了，睡不着。",
+        "reason": "",
+        "note": "笑着挡回去，不让她再问"
+      },
+      {
+        "scene": 2,
+        "beat": 7,
+        "speaker": "C02",
+        "original": "这是什么？",
+        "issue": "无：最短的问题，压力在筷子停住",
+        "revised": "这是什么？",
+        "reason": "",
+        "note": "声音不变，你在试探，不是好奇"
+      },
+      {
+        "scene": 2,
+        "beat": 8,
+        "speaker": "C01",
+        "original": "刚才一个人放这儿的，说明天有人来拿。",
+        "issue": "无：耸肩一句，转述托付人的话，不加解释",
+        "revised": "刚才一个人放这儿的，说明天有人来拿。",
+        "reason": "",
+        "note": "耸肩，你在敷衍，想把话题带过"
+      },
+      {
+        "scene": 2,
+        "beat": 10,
+        "speaker": "C01",
+        "original": "别人的东西。",
+        "issue": "无：三个字划界线",
+        "revised": "别人的东西。",
+        "reason": "",
+        "note": "手按住抽屉，你在拒绝，不用抬高声音"
+      },
+      {
+        "scene": 2,
+        "beat": 11,
+        "speaker": "C02",
+        "original": "小夏，那个人穿雨衣？",
+        "issue": "无：叫名字一次，用在试探的关键处",
+        "revised": "小夏，那个人穿雨衣？",
+        "reason": "",
+        "note": "很轻，你在确认一件你早就知道的事"
+      },
+      {
+        "scene": 2,
+        "beat": 13,
+        "speaker": "C02",
+        "original": "三年前，这店里丢过一样东西。",
+        "issue": "无：只给半句往事，留钩子",
+        "revised": "三年前，这店里丢过一样东西。",
+        "reason": "",
+        "note": "背对她说，你在决定要说多少"
+      },
+      {
+        "scene": 2,
+        "beat": 15,
+        "speaker": "C02",
+        "original": "别把它交给任何人。",
+        "issue": "无：命令句，第一次正眼看她",
+        "revised": "别把它交给任何人。",
+        "reason": "",
+        "note": "一字一顿，你在把责任压到她身上"
+      },
+      {
+        "scene": 2,
+        "beat": 17,
+        "speaker": "VO",
+        "original": "他怎么知道那人穿雨衣。",
+        "issue": "无：心声一句，把怀疑交给观众",
+        "revised": "他怎么知道那人穿雨衣。",
+        "reason": "",
+        "note": "压得很低，像自己不敢承认这个念头"
+      },
+      {
+        "scene": 2,
+        "beat": 19,
+        "speaker": "C02",
+        "original": "锁上门。",
+        "issue": "无：三个字收尾，命令里有恐惧",
+        "revised": "锁上门。",
+        "reason": "",
+        "note": "不回头，你在逼她立刻照做"
+      }
+    ],
+    "selfCheck": [
+      {
+        "item": "读出声：有没有播音腔的句子？",
+        "ok": true,
+        "note": "全部短句口语，无书面长句"
+      },
+      {
+        "item": "语气词密度：一句最多一个，没有粤语残留（啦／喇／啫）",
+        "ok": true,
+        "note": "几乎不用语气词"
+      },
+      {
+        "item": "网络用语符合年龄：长辈角色不用",
+        "ok": true,
+        "note": "老陈六十多岁，无网络用语"
+      },
+      {
+        "item": "地域口音只给少数角色",
+        "ok": true,
+        "note": "全员标准普通话"
+      },
+      {
+        "item": "英语夹杂符合行业身份",
+        "ok": true,
+        "note": "无英语夹杂，便利店夜班不需要"
+      }
+    ]
   }
-}
+}

package/skill/short-drama/examples/night-shift/project.json

@@ -588,6 +588,8 @@
       "primary": true,
       "reusePlan": "",
       "summary": "凌晨两点的便利店：亮得刺眼、空得发冷，收银台是全店唯一有人气的地方，所有的戏都在那一米柜台前后发生。",
+      "condition": "lived-in",
+      "conditionReason": "这家便利店每晚都在营业：柜台和抽屉有日常使用的痕迹，不是残破，也不是刚搭好的布景。",
       "anchors": [
         {
           "name": "关东煮灯箱",
@@ -647,6 +649,8 @@
       "primary": true,
       "reusePlan": "",
       "summary": "店门外一米宽的雨棚：里面是灯，外面是雨，站在这里的人一半亮一半黑——所有来历不明的人都从这里进出。",
+      "condition": "lived-in",
+      "conditionReason": "雨棚在雨里站了很多年，红白条已经褪色，但店还在开，不是新搭的布景。",
       "anchors": [
         {
           "name": "褪色的红白雨棚",
@@ -709,6 +713,8 @@
       ],
       "scale": "handheld",
       "summary": "一块旧的 3.5 寸机械硬盘，装在防静电袋里，标签被撕掉一半。",
+      "condition": "worn",
+      "conditionReason": "这是三年前的监控硬盘：标签撕掉一半，接口生锈，不是主办方刚发的新道具。",
       "anchors": [
         {
           "name": "撕掉一半的标签",

package/skill/short-drama/reference/dialogue-craft.md

@@ -1,9 +1,12 @@
 # Dialogue craft — lines as action, in the right 语体
 
-Distilled from the user's `w-film-dialogue-editor` skill. Read it before writing a script's
-lines and before any polish pass (`render dialogue --ep N`). What can be checked is the
-**advisory** `dialogue-register` gate — it reports, it never blocks, because a 穿越 character
-speaking 古装 in a modern drama is right, not wrong.
+A summary of the user's `w-film-dialogue-editor` skill, which is installed beside this one and is
+the source of truth: its revision framework, one language guide per 语体 and a corpus of real
+usage queried with `scripts/ref.mjs`. This page is the short version to keep in mind; the guide of
+the drama's 语体 is read in full before writing lines and again in the dialogue review (§3). Two
+gates: the **advisory** `dialogue-register` notes lines that break the register (a 穿越 character
+speaking 古装 in a modern drama is right, not wrong), and the **required** `dialogue-review` holds
+the storyboard until every line has been reviewed.
 
 ## 1. A line is an action under pressure
 
@@ -94,14 +97,16 @@ a 6–18-month half-life: two or three signature words per character, no more. P
 vertical short drama allows damn / hell / ass / limited bitch, avoids the f-word; the first three
 seconds are clean.
 
-## 3. The polish pass — `render dialogue --ep N`
+## 3. The dialogue review — `seed dialogue` → fill → `apply dialogue`
 
-Produces the reference's 逐句校对表: 原句 filled, 问题 / 修改句 / 修改理由 / 导演／演员提示 to fill,
-with each speaker's `bible.speech` summary in the last column. Levels: 轻修 (keep the sentence,
-fix tone, rhythm, filler), 中修 (restructure, cut explanation, sharpen the attack and defence),
-重修 (re-design the beats and strategies, keep the story intent). Do it when the user asks, when
-dialogue is the visible weakness, or when the register gate keeps noting the same scene. The
-filled table is appended to the screenplay dataset under `## 对白校对`.
+Required after every script (`stage-script.md` § Dialogue review). `seed dialogue --ep N` writes
+`episode.dialogueReview` — one row per line with 原句 filled — and prints the guide and the corpus
+query for each speaker. Fill it the way the dialogue skill reviews a scene: the diagnosis first,
+then per line 问题 / 修改句 / 修改理由 / 导演／演员提示, then the guide's 自检清单. Levels: 轻修 (keep
+the sentence, fix tone, rhythm, filler), 中修 (restructure, cut explanation, sharpen the attack and
+defence), 重修 (re-design the beats and strategies, keep the story intent). `apply dialogue` writes
+the revised lines into the script; `render dialogue` prints the filled table, which is appended to
+the screenplay dataset under `## 对白校对`.
 
 | 原句 | 问题 | 修改句 | 修改理由 |
 | --- | --- | --- | --- |

package/skill/short-drama/reference/prompt-audit.mjs

@@ -0,0 +1,308 @@
+#!/usr/bin/env node
+/**
+ * prompt-audit.mjs — reference-binding proofreader for AI short-drama video prompts.
+ *
+ * It does NOT rewrite the prompt. It reads the prompt as TEXT, compares it with what the
+ * project says must be on screen (cuts → characters / props / scene), and reports every place
+ * where the model could mistake which @imageN is a person, which is the room, which is a prop,
+ * or where a shot simply never mentions the reference it depends on.
+ *
+ * usage:
+ *   node prompt-audit.mjs --project <dir> --ep 1 --segment E01-02 --prompt <file> [--json]
+ *
+ * exit code: 0 clean (warnings allowed), 1 errors found.
+ */
+
+import { readFileSync } from 'node:fs';
+import { resolve, join } from 'node:path';
+
+// ── args ────────────────────────────────────────────────────────────────────
+const argv = process.argv.slice(2);
+const arg = (name, fallback = null) => {
+  const i = argv.indexOf(`--${name}`);
+  return i >= 0 && argv[i + 1] && !argv[i + 1].startsWith('--') ? argv[i + 1] : fallback;
+};
+const PROJECT = arg('project');
+const EP = Number(arg('ep', 1));
+const SEGMENT = arg('segment');
+const PROMPT_FILE = arg('prompt');
+const AS_JSON = argv.includes('--json');
+if (!PROJECT || !SEGMENT || !PROMPT_FILE) {
+  console.error('usage: node prompt-audit.mjs --project <dir> --ep 1 --segment E01-02 --prompt <file> [--json]');
+  process.exit(2);
+}
+
+const project = JSON.parse(readFileSync(resolve(PROJECT, 'project.json'), 'utf8'));
+const epFile = join(resolve(PROJECT, 'episodes'), `E${String(EP).padStart(2, '0')}.json`);
+const episode = JSON.parse(readFileSync(epFile, 'utf8'));
+const prompt = readFileSync(resolve(PROMPT_FILE), 'utf8')
+  .split('\n')
+  .filter((l) => !l.startsWith('#'))          // drop `prompt` command's header comment line
+  .join('\n')
+  .trim();
+
+const seg = (episode.storyboard?.segments || []).find((s) => s.id === SEGMENT);
+if (!seg) {
+  console.error(`no segment ${SEGMENT} in ${epFile}`);
+  process.exit(2);
+}
+const scene = (episode.scenes || [])[Number(seg.sceneIndex) - 1] || {};
+const params = project.params || {};
+const charById = (id) => (project.characters || []).find((c) => c.id === id);
+const sceneById = (id) => (project.scenes || []).find((s) => s.id === id);
+const propById = (id) => (project.props || []).find((p) => p.id === id);
+
+// ── what the project says the segment needs ────────────────────────────────
+/** Reference plan in connection order, the same way the engine derives it (env → people → props → voices). */
+function planOf(seg) {
+  const imgs = [];
+  if (Array.isArray(seg.refOverride) && seg.refOverride.length) {
+    for (const r of seg.refOverride) imgs.push({ entity: r.entity, asset: r.asset || 'sheet', declared: r.role || null });
+  } else {
+    if (scene.sceneId) {
+      imgs.push({ entity: scene.sceneId, asset: 'sheet', declared: 'scene' });
+      const se = sceneById(scene.sceneId);
+      if (scene.lighting && se?.wji?.assets?.lighting?.[scene.lighting]) imgs.push({ entity: scene.sceneId, asset: `lighting:${scene.lighting}`, declared: 'lighting' });
+    }
+    const seen = new Set();
+    for (const c of seg.cuts || []) for (const id of c.characters || []) if (!seen.has(id)) { seen.add(id); imgs.push({ entity: id, asset: 'sheet', declared: 'character' }); }
+    for (const c of seg.cuts || []) for (const id of c.props || []) if (!seen.has(id)) { seen.add(id); imgs.push({ entity: id, asset: 'sheet', declared: 'prop' }); }
+  }
+  imgs.forEach((r, i) => (r.label = `@image${i + 1}`));
+  const speakers = [];
+  for (const c of seg.cuts || []) for (const id of c.characters || []) if (id !== 'VO' && !speakers.includes(id)) speakers.push(id);
+  const voices = speakers.map((id, i) => ({ entity: id, label: `@audio${i + 1}`, kind: 'audio' }));
+  return { imgs, voices };
+}
+
+const roleOfEntity = (id, asset, declared) => {
+  if (declared) return declared;
+  if (/^lighting:/.test(String(asset))) return 'lighting';
+  if (/^S/i.test(id)) return 'scene';
+  if (/^P/i.test(id)) return 'prop';
+  if (/^C/i.test(id)) return 'character';
+  return 'unknown';
+};
+
+const plan = planOf(seg);
+for (const r of plan.imgs) r.role = roleOfEntity(r.entity, r.asset, r.declared);
+
+// ── identity tokens: how a legend can prove WHICH person a picture is ──────
+const CN_STOP = new Set(['男的', '女的', '一個', '一个', 'the', 'and', 'with', 'her', 'his', 'man', 'woman']);
+function tokensOf(ch) {
+  const t = new Set();
+  const add = (s) => {
+    for (const w of String(s || '').toLowerCase().split(/[^a-z0-9\u4e00-\u9fff]+/)) if (w.length > 2 && !CN_STOP.has(w)) t.add(w);
+    for (const run of String(s || '').match(/[\u4e00-\u9fff]+/g) || []) {
+      if (run.length <= 3) t.add(run);
+      for (let i = 0; i + 1 < run.length && run.length > 3; i++) t.add(run.slice(i, i + 2));
+      for (let i = 0; i + 2 < run.length; i++) t.add(run.slice(i, i + 3));
+    }
+  };
+  add(ch.identityEn);
+  add(ch.profile?.appearance);
+  add(ch.profile?.identity);
+  const g = ch.profile?.gender;
+  if (g) t.add(g === '男' ? '男' : '女');
+  return t;
+}
+
+// ── prompt parsing ─────────────────────────────────────────────────────────
+const SHOT_RE = /(?:^|\n)[ \t]*(?:镜头|鏡頭|Shot)[ \t]*(\d+)/g;
+const shotMarks = [];
+for (let m; (m = SHOT_RE.exec(prompt)); ) shotMarks.push({ k: Number(m[1]), at: m.index + (m[0].startsWith('\n') ? 1 : 0) });
+const firstShotAt = shotMarks.length ? shotMarks[0].at : prompt.length;
+const head = prompt.slice(0, firstShotAt);              // legend + global blocks
+const shots = shotMarks.map((s, i) => ({
+  k: s.k,
+  text: prompt.slice(s.at, i + 1 < shotMarks.length ? shotMarks[i + 1].at : prompt.length),
+}));
+const legend = head.split('\n').filter((l) => /@image\d+/.test(l)).join('\n');
+
+const IMG_RE = /@image(\d+)/g;
+const labelPositions = new Map();       // label -> { first, count, span }
+const legendLines = head.split('\n').filter((l) => /^@image\d+\b/.test(l.trim()));
+{
+  const all = [...head.matchAll(IMG_RE)];
+  for (let i = 0; i < all.length; i++) {
+    const label = `@image${all[i][1]}`;
+    const from = all[i].index;
+    const to = i + 1 < all.length ? all[i + 1].index : head.length;
+    const prev = labelPositions.get(label);
+    const mine = legendLines.filter((l) => l.trim().startsWith(label));
+    const lines = mine.length;
+    const span = lines ? mine[0].trim().slice(label.length) : head.slice(from, to);
+    labelPositions.set(label, { first: from, count: (prev?.count || 0) + 1, lines, span });
+  }
+}
+
+const countIn = (text, label) => (text.match(new RegExp(`${label}(?!\\d)`, 'g')) || []).length;
+
+// ── checks ─────────────────────────────────────────────────────────────────
+const issues = [];
+const add = (level, code, msg) => issues.push({ level, code, msg });
+const ROLE_EN = /\b(scene|environment|character|person|prop|lighting|light reference|moment|action reference|style)\b/i;
+const ROLE_ZH = /(場景|场景|環境|环境|人物|角色|道具|光影|光線|光线|瞬間|瞬间|撞擊|撞击|動作參考|动作参考|風格|风格|車廂|车厢)/;
+const BIND_EN = /\b(match|as in|from|same as|use)\b/i;
+const BIND_ZH = /(照|參照|参照|跟|用|依|係|是)/;
+const NEG_EN = /\b(not|only|do not|don't)\b/i;
+const NEG_ZH = /(唔照|不要照|唔好照|只照|只取|唔要|不要|只係|只是|只)/;
+const hasCJK = (t) => /[\u4e00-\u9fff]/.test(String(t || ''));
+const ROLE_WORDS = { test: (t) => (hasCJK(t) ? ROLE_ZH : ROLE_EN).test(t) };
+const BIND_WORDS = { test: (t) => (hasCJK(t) ? BIND_ZH : BIND_EN).test(t) };
+const NEG_WORDS = { test: (t) => (hasCJK(t) ? NEG_ZH : NEG_EN).test(t) };
+
+// R1 legend exists
+if (!/@image\d+/.test(head)) add('error', 'R1-legend-missing', 'prompt 首段冇 @imageN legend：模型唔知邊張圖係咩');
+
+// R2 every planned image is in the legend, exactly once
+for (const r of plan.imgs) {
+  const p = labelPositions.get(r.label);
+  if (!p) add('error', 'R2-legend-missing-label', `${r.label}（${r.role}：${r.entity}）已接但 legend 冇提`);
+  else if (p.lines > 1) add('warn', 'R2-legend-duplicate', `${r.label} 有 ${p.lines} 條 legend 行 — 一張圖一行，重覆會令模型重新解讀角色`);
+  else if (p.lines === 0) add('warn', 'R2-legend-inline', `${r.label} 夾在一句裡面，冇自己一行 — 一張圖一行，「@image2、@image3 是人物圖」係模型撈亂嘅溫床`);
+}
+
+// R3 prompt uses labels the plan does not have (extra hand-connected refs)
+const maxPlanned = plan.imgs.length;
+for (const label of labelPositions.keys()) {
+  const n = Number(label.slice(6));
+  if (n > maxPlanned) add('warn', 'R3-extra-ref', `${label} 唔在項目派生嘅引用次序（只有 ${maxPlanned} 張圖）— 額外接嘅圖要寫入 segment.refOverride，否則重跑 prompt 就對唔上`);
+}
+
+// R4 role words per label + R5 negative instruction on non-character refs
+for (const r of plan.imgs) {
+  const p = labelPositions.get(r.label);
+  if (!p) continue;
+  if (!ROLE_WORDS.test(p.span)) add('warn', 'R4-role-unsaid', `${r.label} 冇講明係咩角色（場景／人物／道具／光影／瞬間）— 模型只會當「一張參考圖」`);
+  if (r.role !== 'character' && !NEG_WORDS.test(p.span)) add('warn', 'R5-no-negative', `${r.label}（${r.role}）冇寫「只照…唔照…」：非人物圖一定要講明唔好抄佢嘅構圖／人物，唔係模型會連人一齊抄`);
+}
+
+// R6 legend must let the model tell one person from another
+const segChars = [];
+for (const c of seg.cuts || []) for (const id of c.characters || []) if (id !== 'VO' && !segChars.includes(id)) segChars.push(id);
+const charImgs = plan.imgs.filter((r) => r.role === 'character');
+if (segChars.length >= 2 && charImgs.length >= 2) {
+  const total = new Map();
+  for (const id of segChars) for (const t of tokensOf(charById(id) || {})) total.set(t, (total.get(t) || 0) + 1);
+  const assigned = new Map();
+  for (const r of charImgs) {
+    const span = (labelPositions.get(r.label)?.span || '').toLowerCase();
+    let best = null, bestScore = 0;
+    for (const id of segChars) {
+      if (assigned.has(id)) continue;
+      let score = 0;
+      for (const t of tokensOf(charById(id) || {})) if (span.includes(t) && (total.get(t) || 0) <= 1) score += 2;
+      const g = charById(id)?.profile?.gender;
+      if (g && span.includes(g) && segChars.filter((x) => charById(x)?.profile?.gender === g).length === 1) score += 1;
+      if (score > bestScore) { best = id; bestScore = score; }
+    }
+    if (best && bestScore >= 2) assigned.set(best, r.label);
+  }
+  if (assigned.size < segChars.length) {
+    const missing = segChars.filter((id) => !assigned.has(id)).map((id) => charById(id)?.name || id);
+    add('error', 'R6-identity-unresolvable', `legend 冇寫到分辨特徵，無法對應邊張圖係邊個人（對唔到：${missing.join('、')}）。` +
+      `人物圖必須逐張寫「係邊個」——性別、年齡、髮型、衣著各一句；寫成「@image2、@image3 是人物參考圖」模型就會互換面孔`);
+  }
+  // two people described in one legend sentence, with no per-picture role
+  for (const line of legendLines) {
+    const labels = line.match(/@image\d+/g) || [];
+    const people = labels.filter((l) => charImgs.some((r) => r.label === l));
+    if (people.length >= 2 && !/not its|never its|只照|不要照/i.test(line.replace(/@image\d+/g, ''))) {
+      add('warn', 'R6b-shared-clause', `legend 一句寫咗幾個人（${people.join('、')}）又冇逐個講明 — 一張圖一句，否則模型會互換`);
+    }
+  }
+}
+
+// R7 per-shot binding: person / prop / scene
+const cuts = seg.cuts || [];
+if (shots.length !== cuts.length) add('warn', 'R0-shot-count', `prompt 有 ${shots.length} 個鏡頭，分鏡有 ${cuts.length} 個 cut — 冇法逐鏡核對（人名／道具綁定會漏）`);
+const deriveOrder = (() => {
+  // when the legend is ambiguous, fall back to the engine's derivation order to still give per-shot advice
+  const m = new Map();
+  for (const r of plan.imgs) if (r.role === 'character') m.set(r.entity, r.label);
+  return m;
+})();
+const unboundChar = new Map();   // entity -> {label, shots: []}
+for (let i = 0; i < cuts.length && i < shots.length; i++) {
+  const c = cuts[i];
+  const s = shots[i];
+  for (const id of c.characters || []) {
+    const label = deriveOrder.get(id);
+    if (!label) continue;
+    if (!countIn(s.text, label)) {
+      if (!unboundChar.has(id)) unboundChar.set(id, { label, shots: [] });
+      unboundChar.get(id).shots.push(s.k);
+    }
+  }
+  for (const id of c.props || []) {
+    const label = plan.imgs.find((r) => r.entity === id)?.label;
+    if (label && !countIn(s.text, label)) add('error', 'R7-prop', `鏡頭 ${s.k}：道具 ${propById(id)?.name || id} 在畫面內，但鏡頭文冇 ${label} — 道具外觀會由模型自創`);
+  }
+  if (i === 0) {
+    const sceneLabel = plan.imgs.find((r) => r.role === 'scene')?.label;
+    if (sceneLabel && !countIn(head + '\n' + shots.map((x) => x.text).join('\n'), sceneLabel) && !countIn(prompt.slice(firstShotAt), sceneLabel)) {
+      add('warn', 'R7-scene', `${sceneLabel} 只在 legend 出現，冇任何鏡頭講明照佢嘅空間／材質 — 每個 segment 至少講一次「整段都在 ${sceneLabel} 的空間裡」`);
+    }
+  }
+}
+
+for (const [id, bad] of unboundChar) {
+  add('error', 'R7-char', `${charById(id)?.name || id} 在鏡頭 ${bad.shots.join('、')} 畫面內，但鏡頭文冇 ${bad.label}` +
+    ` — 模型會沿用上一鏡嘅樣或者自己編。每個有佢嘅鏡頭都要寫一次 ${bad.label}`);
+}
+// R8 every image ref is actually used somewhere outside the legend (people are covered by R7)
+const bodyText = prompt.slice(firstShotAt);
+for (const r of plan.imgs) {
+  if (r.role === 'character' || r.role === 'scene') continue;
+  if (!countIn(bodyText, r.label)) add('error', 'R8-unused-ref', `${r.label}（${r.role}：${r.entity}）接咗但鏡頭文一次都冇用 — 等於冇參考`);
+}
+
+// R9 dialogue still carries its voice label (mirrors the engine's own rule)
+const dialogues = [];
+for (const c of cuts) {
+  for (let b = (c.beats?.[0] || 1) - 1; b <= (c.beats?.[1] || 1) - 1; b++) {
+    const beat = (scene.flow || [])[b];
+    if (beat?.speaker) dialogues.push({ speaker: beat.speaker, line: beat.line });
+  }
+}
+const norm = (t) => String(t).replace(/[…⋯]/g, '...').replace(/[“”「」『』"]/g, '').replace(/\s+/g, '');
+const flat = norm(prompt);
+for (const l of dialogues) {
+  const needle = norm(l.line);
+  const at = flat.indexOf(needle);
+  if (at < 0) { add('error', 'R9-dialogue-missing', `台詞「${l.line}」逐字唔見（引擎逐字對，標點差異都算唔同）`); continue; }
+  const around = flat.slice(Math.max(0, at - 160), at + needle.length + 60);
+  if (!/@audio\d/.test(around)) add('error', 'R9-voice-unbound', `台詞「${l.line}」附近冇 @audioN — 模型會用預設聲`);
+}
+
+// R10 hygiene: no names, subtitles off
+for (const ch of project.characters || []) if (ch.name && prompt.includes(ch.name)) add('warn', 'R10-name-leak', `prompt 出現角色名「${ch.name}」（引擎 prompt-hygiene 會當錯；描述性名字可以留，但要預期被擋）`);
+if (!params.subtitles && !/(全程无字幕|no subtitles)/i.test(prompt)) add('warn', 'R10-subtitles', 'prompt 冇寫「無字幕／no subtitles」');
+
+// R11 ambiguity sweep: a whole sentence that names several pictures without saying what each is for
+for (const s of shots) {
+  for (const sentence of s.text.split(/[。；;\n]/)) {
+    const body = sentence.replace(/^\s*(?:Shot|镜头|鏡頭)\s*\d+\s*\([^)]*\)\s*:?/, '');
+    const labels = body.match(/@image\d+/g) || [];
+    if (labels.length >= 2 && !BIND_WORDS.test(body.replace(/@image\d+/g, '')) && body.length > 40) {
+      add('warn', 'R11-ambiguous-clause', `鏡頭 ${s.k}「${body.trim().slice(0, 44)}…」一句有 ${labels.length} 張圖但冇逐張講用途 — 拆開寫`);
+    }
+  }
+}
+
+// ── report ─────────────────────────────────────────────────────────────────
+const errors = issues.filter((x) => x.level === 'error');
+const warns = issues.filter((x) => x.level === 'warn');
+if (AS_JSON) {
+  console.log(JSON.stringify({ segment: SEGMENT, shots: shots.length, cuts: cuts.length, plan: plan.imgs, issues }, null, 2));
+} else {
+  console.log(`\n== prompt-audit · ${SEGMENT} ==`);
+  console.log(`refs (derived order): ${plan.imgs.map((r) => `${r.label}=${r.role}:${r.entity}`).join('  ')}`);
+  console.log(`shots in prompt: ${shots.length} · cuts in storyboard: ${cuts.length} · images declared: ${labelPositions.size}\n`);
+  for (const x of errors) console.log(`❌ [${x.code}] ${x.msg}`);
+  for (const x of warns) console.log(`⚠️  [${x.code}] ${x.msg}`);
+  if (!issues.length) console.log('✅ 冇發現引用綁定問題');
+  console.log(`\n${errors.length} error(s), ${warns.length} warning(s)`);
+}
+process.exit(errors.length ? 1 : 0);