@wji.ai/cli 0.9.1 → 0.9.4

package/skill/short-drama/reference/reference-binding.md

@@ -0,0 +1,84 @@
+# Reference binding — 參考圖綁定
+
+A video prompt is **one string** read by a model that has never seen the storyboard, the project
+or the canvas. It cannot tell `@image2` from `@image3`, cannot tell the room from a prop, and
+defaults to whatever it saw in the previous shot for anything the shot does not name. So the
+picture plan has to be written twice: once as a legend that says **what each picture is**, and once
+inside every shot that depends on it.
+
+Gate: `ref-binding` (stage `storyboard`, sheets mode). Proofreader for text that is not in the
+project file: `prompt-audit.mjs` (shipped in `reference/`).
+
+## The six roles — and what may not be copied from each
+
+| Role | Legend says | Take | Explicitly NOT take |
+| --- | --- | --- | --- |
+| 場景 environment | `@image1` 係邊個場景 | 佈局、材質、磨損、比例 | 人物（場景圖本身冇人） |
+| 光影 lighting | `@image5` 係光影參考 | 光比、反差、色溫、光源位置 | 構圖、人物、衣著 |
+| 人物 character | `@image2` 係邊個人 | 面容、髮型、衣著 | 其他鏡頭嘅表情／姿勢 |
+| 道具 prop | `@image4` 係邊件道具 | 外形、材質、尺寸、標記 | 枱面、背景、手 |
+| 瞬間 moment | `@image6` 係「撞擊一瞬間」 | 打光方向、反差、剪影形狀 | 背景細節、其他鏡頭嘅走位 |
+| 風格 style | `@image7` 係風格板 | 質感、顆粒、調色 | 題材、人物、構圖 |
+
+Two consequences the model actually obeys:
+
+- **一張圖一行。** 兩張人物圖寫在同一句（`@image2、@image3 是人物參考圖`）＝ 兩張圖冇分別 ＝ 面孔互換。
+- **唔係人物嘅圖，一定要寫「唔照」。** 光影圖／瞬間圖如果唔寫「只照光比，不要照它的構圖與人物」，
+  模型會連圖裡面嘅人一齊抄落去。
+
+## Legend — good and bad
+
+```
+✗ @image2、@image3 是人物參考圖，照他們的面容、髮型與衣著
+✗ @image4 is the character sheet of one person in this segment — match face, hair and clothes exactly
+```
+
+```
+✓ @image2 是女主角三視圖（二十五六歲，齊蔭 bob 頭、oversized 黑 tee、白波鞋）：只照面容、髮型、衣著
+✓ @image3 是男主角三視圖（二十七八歲，豎起黑髮、小鬍子、白 tee）：只照面容、髮型、衣著
+✓ @image1 是車廂場景（右軚電動車、近黑、中控屏微光）：只照空間、材質、磨損，場景圖裡沒有人
+✓ @image4 是白衣女鬼服裝與妝容參考：濕透白袍、長濕黑髮、赤腳
+✓ @image5 是車廂光影參考：只照光比、反差、色溫、光源位置，不要照它的構圖與人物
+✓ @image6 是「車撞女鬼一瞬間」：只照打光方向、反差與剪影形狀，不要照它的背景
+✓ @image7 是風格板：只照質感、顆粒、調色
+```
+
+Every character line must carry at least two facts from the character bible (age / gender / hair /
+clothes), because that is the only thing that makes two sheets distinguishable.
+
+## Shot binding — the shot names what it depends on
+
+```
+✓ 鏡頭 1（…）：@image2 坐直指住屏幕，@image3 單手扶軚伸手去點…
+✓ 鏡頭 4（…）：@image4 走出馬路正中間（照 @image6 的剪影與打光）…
+✓ 鏡頭 8（…）：@image4 的白臉貼喺玻璃外；屏幕上的導航界面照 @image5 的 UI
+✓ 整段都在 @image1 的空間裡：佈局、材質與舊痕不變，不要重新設計場景。
+```
+
+- 每個有角色／道具嘅鏡頭，**逐個 label 寫一次**；唔係寫喺 legend 就當夠。
+- 場景 label 每段講一次就夠（「整段都在 @image1 的空間裡」）。
+- Handheld / 行車感 / 鏡頭語言呢類**全段共通**嘅指令，用獨立一段寫，唔好混入逐鏡描述。
+- Hand-written production prompts: run `prompt-audit.mjs` before generation; it reports unbound
+  people and props per shot, indistinguishable legend lines, unused pictures, and refs that are not
+  in the derived plan (so the canvas will renumber them).
+
+```bash
+node reference/prompt-audit.mjs --project <project> --ep 1 --segment E01-02 --prompt prompt.txt
+```
+
+## Ref budget
+
+- **≤ 4 pictures per segment** unless every extra picture has a role from the table above.
+- More than 3 people in one shot → every person in that shot is named; a group sheet still needs
+  the individual labels if faces must be right.
+- The engine caps a segment by model (`minimax-h3-ref` 9, `seedance-2.5-ref` 30, `wan-3.0-ref` 10);
+  the gate `ref-binding` now also fails a sheets segment over the model's cap.
+
+## Checklist before generating
+
+1. 每張圖有一行 legend，寫明係邊個／邊件／邊度／邊種參考。
+2. 兩張人物圖唔會讀起來一樣（年齡、性別、髮型、衣著至少兩樣）。
+3. 非人物圖有「唔照」句。
+4. 場景每段講一次；每個出場人物、每件道具喺自己嘅鏡頭有 label。
+5. 額外接嘅圖（光影／瞬間／風格）寫入 `segment.refOverride`，否則 canvas 一接線全部編號會移。
+6. `$SD validate --stage storyboard` 綠；或者 `prompt-audit.mjs` 零 error。

package/skill/short-drama/reference/schema.md

@@ -65,7 +65,7 @@ All commands accept either the directory or the `project.json` path.
 | `maxVideoRefs` / `maxAudioRefs` | per mode | how many clip / voice references the model takes; set them lower to tighten, never higher |
 | `lookPhraseZh` | empty | the same look phrase in the drama's language, gated in a **Chinese** prompt (`wan`) where the English `lookPhrase` would be both out of place and extra length for the content filter to judge |
 | `speechLeadSeconds` | 0.5 | added to a cut's dialogue seconds before the fit check — the mouth opens after the cut starts; also where a line starts in the `full` soundtrack prompt |
-| `audioTrack` | `post` | where the soundtrack is made. **`post`** (the default): the clips carry dialogue, ambience and in-picture sfx but **no music** (`No music.` / `non_diegetic_music: N/A` / `无配乐`) and no `post: true` sfx — music and sound design cannot carry across separately generated segments, so they are made once over the **edited episode**, on its timecode (`prompt score E01`, then `mix-over-video.md`); no per-segment audio node. `bed` (ambience + sfx + timed music, no voices) / `full` (the bed plus every line in the connected `@Audio1..3` voices — at most three speakers): a Seed Audio node per segment, music in the clip prompt. `none`: no soundtrack work |
+| `audioTrack` | `post` | where the soundtrack is made. **`post`** (the default): the clips carry dialogue, ambience and in-picture sfx but **no music** (`No music.` / `non_diegetic_music: N/A` / `无配乐`) and no `post: true` sfx — music and sound design cannot carry across separately generated segments, so they are made once over the **edited episode**, on its timecode (`prompt score E01`, then `../short-drama-sound/reference/mix-over-video.md`); no per-segment audio node. `bed` (ambience + sfx + timed music, no voices) / `full` (the bed plus every line in the connected `@Audio1..3` voices — at most three speakers): a Seed Audio node per segment, music in the clip prompt. `none`: no soundtrack work |
 | `foregroundLevel` | `medium`; `high` for 唯美 / 美感优先 (`composition: art-first`, `cinematography: sweet-romance` or `costume-idol`, a genre naming 唯美 or 美感) | how often a cut shoots past a foreground, measured across the episode: `low` up to 25 % of the cuts, `medium` 25–50 %, `high` 50 % or more (gate `photography`; `--foreground` at init). Which cuts: `stage-storyboard.md` |
 | `scoreFps` | 24 | frame rate `prompt score` writes SMPTE `HH:MM:SS:FF` at (`--fps` overrides); set it to the edit's real fps |
 | `subtitles` | false | `true` drops the subtitles-off sentence requirement from the prompts |
@@ -87,6 +87,7 @@ All commands accept either the directory or the `project.json` path.
 | `turnSeconds` | 30 | the longest stretch allowed between turn points — the hook, every beat marked `turn: true`, the 爽点 and the cliff |
 | `maxConsecutiveLines` | 3 | line beats in a row without an action beat between them |
 | `minActionRatio` | 0.35 | action beats ÷ line beats over an episode with ≥ 5 lines (谈话+行动) |
+| `dialogueReview` | `required` | `required`: every episode's lines are reviewed against the `w-film-dialogue-editor` guide of their 语体 before the storyboard (gate `dialogue-review`); `off`: no review — only when the user asks |
 | `maxLineWords` / `maxNameCallsPerScene` | 20 / 2 | English registers only: words per line, and how often another character is addressed by name in one scene (advisory) |
 | `charsPerSecond` / `actionSeconds` | 4.5 / 2.5 | script duration math: a line takes `nonWhitespaceChars ÷ charsPerSecond` s, an action beat takes `actionSeconds` |
 | `tolerance` | 0.15 | ±15 % for episode duration gates |
@@ -203,7 +204,7 @@ id or list) and `gag` (`setup` / `build` / `punchline`) with `runningGag`.
 ```json
 {
   "id": "S01", "name": "便利店店内", "primary": true, "reusePlan": "",
-  "summary": "设计意图…", "anchors": [{ "name": "闪烁的关东煮灯箱", "desc": "…" }],
+  "summary": "设计意图…", "condition": "lived-in", "conditionReason": "这家店每晚都在营业，是用旧了而不是残破", "anchors": [{ "name": "闪烁的关东煮灯箱", "desc": "…" }],
   "lighting": [{ "state": "凌晨冷白光", "prompt": "…" }],
   "image": { "prompt": "…", "negativePrompt": "…", "sheet": "…", "tags": [] },
   "variantOf": "", "changes": "",
@@ -217,13 +218,20 @@ paintable and checkable), `lighting` (≥1, prompts in `imagePromptLang`, derive
 episodes it appears in), `image.*` (`imagePromptLang`, **empty scene — negative prompt
 must ban people**), `variantOf` + `changes` for variants of a parent scene.
 `wji.nodes.lighting` / `wji.assets.lighting` are keyed by lighting `state`.
+`condition` is `new`, `lived-in` or `worn`, and `conditionReason` says why the story gives it
+that condition — the same two fields are required on every prop. Realism is materials, light and
+detail, not wear. A luxury interior, a set just built, a prop the organisers have just issued,
+and every national or regional flag are `new` and accurate. Anchors follow the condition: do not
+add peeling paint, rust or fading by default. Only `worn` may ban `brand-new` or `showroom-clean`
+in `negativePrompt`. A `new` asset fails when an anchor or prompt says worn, rust, faded, chipped,
+torn, frayed, peeling, stain, 殘舊, 生鏽, 褪色 or 剝落.
 
 ### props[]
 
 ```json
 {
   "id": "P01", "name": "旧监控硬盘", "function": "全剧悬念核心：里面到底录了什么",
-  "beatIds": ["B02"], "scale": "handheld", "summary": "…",
+  "beatIds": ["B02"], "scale": "handheld", "summary": "…", "condition": "worn", "conditionReason": "三年前的硬盘，标签撕掉、接口生锈",
   "anchors": [{ "name": "…", "desc": "…" }], "states": [{ "state": "封在防静电袋里", "prompt": "…" }],
   "relatedScenes": ["S01"], "carriedBy": ["林小夏"],
   "image": { "prompt": "…", "negativePrompt": "…", "sheet": "…", "tags": [] },
@@ -295,6 +303,7 @@ prompt bans hands**). At most `thresholds.maxProps` (8).
 | `flow[].turn` | `true` on a beat that brings new information, an obstacle, a counter or a choice; the `rhythm` gate measures hook → turns → 爽点 → cliff against `turnSeconds` |
 | `scenes[].sceneId` / `lighting` / `characters` / `props` | must exist in `project.json`; `lighting` must be a state of that scene |
 | `flow[]` | alternating **action beats** `{action}` (narrative, one thing per beat, **common actions only**, no quotes) and **line beats** `{speaker, line, delivery?}`; `speaker` ∈ scene `characters` or `VO` (an off-screen voice: it gets no `@audio` reference, the prompt calls it off-screen and keeps every visible mouth closed while it speaks); `line` ≤ `maxLineChars` non-whitespace characters; `delivery` names nobody when the video prompt carries it (a Chinese `wan` prompt, or an English delivery) — gate `beat-flow` |
+| `dialogueReview` | the 对白校对 record (gate `dialogue-review`, required unless `params.dialogueReview: off`), seeded by `seed dialogue`: `{ level: 轻修 \| 中修 \| 重修, guides: ["lang-zh-hk.md", …] (every speaker's 语体 guide of `w-film-dialogue-editor`), diagnosis: { goal, conflict, subtext, voice, rhythm }, rows: [{ scene, beat, speaker, original, issue, revised, reason, note, delivery? }] (one per line beat, 1-based positions; `revised` must equal the script's line — `apply dialogue` writes it; `note` a playable action), selfCheck: [{ item, ok, note }] (the guide's 自检清单, ≥ 3) }` |
 
 Duration: `Σ lines (chars ÷ charsPerSecond) + Σ actions × actionSeconds`.
 
@@ -322,7 +331,7 @@ Duration: `Σ lines (chars ÷ charsPerSecond) + Σ actions × actionSeconds`.
 | `motion` | `{ trigger, action, result }` — cue, one visible state chain, landing (or `hold`); English unless the prompt language is Chinese; no names, no inferred states, ≤ 3 hand verbs — gate `motion` |
 | `cameraNote` | `{ motivation, trigger, path, endFraming }` on any move (`motivation` ∈ `reveal / pressure / alignment / relationship / transition / stillness`), `{ hold }` on a `Static Shot` — gate `motion` |
 | `frame` | English prompt for the moment: composition, this moment's positions and poses, the size phrase, the style's `framePhrase` (`photorealistic cinematic film still` / `cinematic film still` / `hand-painted anime film still`), **and verbatim** the camera phrase (`shot on an ARRI ALEXA Mini LF with a Zeiss Supreme Prime 75mm at T1.5`), the light-direction phrase and the foreground, and under a `cinematography` preset also its `framePhrase` and the `light.source`; **no names** — identity comes from the sheet references |
-| `prompt` | the video prompt, shape by `params.video` (below); gate `prompt-hygiene` (no `…`, no ids / paths / notes, subtitles off unless `params.subtitles`, H3 lips-closed clause, sheets-mode `@imageK` per shot) |
+| `prompt` | the video prompt, shape by `params.video` (below); gate `prompt-hygiene` (no `…`, no ids / paths / notes, subtitles off unless `params.subtitles`, H3 lips-closed clause, sheets-mode `@imageK` per shot) and gate `ref-binding` (one legend line per picture saying **what** it is — a person by age / hair / clothes from the bible, a room by its name, a prop by its name; every non-person picture carries a *take only … / not its composition or its people* clause; every prop and person is bound by its label again inside its own shot; the room is named once per segment; no connected picture unused; sheets-mode picture cap) |
 | `segments[].axis` | `{ between: [C01, C02], workingSide, screenLeft, screenRight }`, required when two or more people appear in the segment |
 | `segments[].grammar` / `exceptions` | `grammar` (a string or an array) says how the segment is cut — `drama` · `action` · `faceslap` · `comedy`; missing, it is `params.grammar`; the old `action: true` means `action`. Each grammar's rules apply (gates `grammar`, `action`, `faceslap`, `comedy`; advisory `action-advice`, `grammar-advice`) and the prompt carries each grammar's line. `exceptions: [{ rule, reason }]` waives one rule for this segment when the user asks — rules `open-wide` `one-move` `impact-reaction` `no-graphic` `reveal-insert` `reaction-chain` `punchline-reaction` `setup-first` (the old `actionExceptions` still works) |
 | `cuts[].idle` | sheets mode: `{ <id>: "shifts her weight" }`, a small motion in place for a person other than the one carrying the action, or `"still"` (breathing and blinking only). In a cut with two or more visible people, all but one need one (gate `idle-motion`) |
@@ -330,7 +339,7 @@ Duration: `Σ lines (chars ÷ charsPerSecond) + Σ actions × actionSeconds`.
 | `cuts[].ots` | the person whose shoulder is near the lens in an over-the-shoulder; `characters` holds both people; counts as a foreground; that person needs no `eyeline` |
 | `segments[].sound` | `{ ambience, sfx: [{ cut, event, post? }], music }` — ambience identical across a scene's segments; `music` is `N/A`, an intent string, or `{ intent, cues: [{ at: <cut number \| "10.5s">, do: enter \| swell \| duck \| drop \| silence \| out, note }] }` — gate `sound`. An sfx with `post: true` is **sound design outside the story world** (stinger, riser, whoosh, impact hit, reverse cymbal) — never in the clip prompt, made over the edit. In `audioTrack: post` the `music` is the **spotting intent** for the episode score: it never enters the clip, and `prompt score E01` lays it, the post sfx, the lines and the scene beds on the episode timeline. In `bed` / `full`, `prompt audio E01-01` derives the segment's Seed Audio soundtrack prompt from it (timecoded `[start s:end s]` ambience / sfx / music, plus the lines with `--full`) — gate `audio-track` |
 | `wji.assets.edit` / `wji.assets.mix` (episode) | `audioTrack: post` only — the locked edit of the episode (clips cut together, no music) and the final mixed episode (edit + score + sound design). `manifest` asks for the edit once every clip is bound, then for the mix |
-| `segments[].refOverride[]` | optional — the Director pins this segment's **picture** references instead of the derived set (environment, people, props): `[{ entity, asset?: "sheet" \| "headshot" \| "lighting:<state>", faceOf?, why?, use?, zh? }]` in the order they will be connected (`@image1…`). `faceOf: "C01"` says this picture supplies that character's face, so their lines bind to it even when a body sheet of theirs comes later. Voices and the previous clip are still appended after the pictures. An empty array means "derived" |
+| `segments[].refOverride[]` | optional — the Director pins this segment's **picture** references instead of the derived set (environment, people, props): `[{ entity, asset?: "sheet" \| "headshot" \| "lighting:<state>", role?, faceOf?, why?, take?, takeZh?, notTake?, use?, zh? }]` in the order they will be connected (`@image1…`). `role` is `environment` · `lighting` · `character` · `prop` · `moment` · `style`; `take` / `notTake` write the legend line for a reference that is not simply the place, a person or the prop — e.g. `{ role: "lighting", take: "only its light ratios, contrast and colour temperature", notTake: "its composition and its people" }` renders as *take only what it is for, not its composition and its people* (gate `ref-binding` requires the negative clause on every non-person picture). `use` / `zh` override the whole line. `faceOf: "C01"` says this picture supplies that character's face, so their lines bind to it even when a body sheet of theirs comes later. Voices and the previous clip are still appended after the pictures. An empty array means "derived" |
 | `storyboard.visualPlans[]` | `{ sceneIndex, turn, alignment, pressure, progression, rhythm, landing, sound }` for the 爽点 scene and the cliff scene (seeded) — gate `visual-plan` |
 | `props[].textPolicy` | optional `{ mode: readable \| symbolic \| blank, text }` — gate `text-policy` |
 

package/skill/short-drama/reference/stage-characters.md

@@ -40,7 +40,7 @@ relationship carries a `secret` — that is where the reversals come from.
 
 **`bible.speech` is the source of the voice.** `voice.pace` / `videoDirection` and every line's
 `delivery` are written from `rhythm`, `silence` and `forbiddenWords`; `seed script` prints them in
-the `seedNote` and `render dialogue` puts them beside each line for the polish pass.
+the `seedNote` and `render dialogue` puts them beside each line for the dialogue review.
 
 ## Image prompts (English, always)
 

package/skill/short-drama/reference/stage-script.md

@@ -19,6 +19,20 @@ budget. **Before writing: `render guideline`, then the scene bibles (anchors, li
 the speakers' `render character` — the bible's `speech` is where the lines come from** (see
 `dialogue-craft.md`). Delete `seedNote` when the scenes are written.
 
+`seedNote.dialogue` points at the user's dialogue skill, **`w-film-dialogue-editor`** (installed
+beside this one): the revision framework, the language guide of every 语体 in the episode
+(`lang-zh-hk.md`, `lang-zh-cn.md`, `lang-zh-tw.md`, `lang-zh-hk-mandarin.md`, `lang-zh-period.md`,
+`lang-en.md`) and, per speaker, the corpus query for someone of their age:
+
+```bash
+node <skills>/w-film-dialogue-editor/scripts/ref.mjs --lang zh-hk list --age young --limit 30
+node <skills>/w-film-dialogue-editor/scripts/ref.mjs --lang zh-hk search 搞掂       # one word
+```
+
+Read the guide before writing the first line and pull phrasing from the corpus — particles, slang,
+swearing levels and trade talk come from there, never from memory. The corpus is a vocabulary of
+real usage, not lines to paste: a line still grows from the person and the moment.
+
 ## Hard rules (all gated)
 
 1. **The duration budget is the wall.** Lines cost `chars ÷ 4.5` s, action beats 2.5 s; the
@@ -95,13 +109,46 @@ scene rather than hopping four.
 | omniscience | lines explain everything | move information into action and delivery |
 | flat ending | ends calm | last beat becomes the cliff, even just a whistle |
 
-## Polish (`render dialogue --ep N`)
+## Dialogue review — required after the script (`dialogue-review`)
+
+A script is not done when its structural gates are green: every line is then reviewed against the
+语体 guide, the way `w-film-dialogue-editor` reviews a scene, and the review is kept in the episode.
+The storyboard stage cannot start until it is (`validate --stage storyboard` runs the script gates
+too). `params.dialogueReview: off` skips it — only when the user asks.
+
+```bash
+node {baseDir}/scripts/short-drama.mjs seed  <project> dialogue --ep 1   # dialogueReview: one row per line + the guides to read
+node {baseDir}/scripts/short-drama.mjs apply <project> dialogue --ep 1   # write the revised lines (and deliveries) into the flow
+node {baseDir}/scripts/short-drama.mjs validate <project> --stage script --ep 1
+node {baseDir}/scripts/short-drama.mjs render <project> dialogue --ep 1  # the filled 对白校对 table
+```
 
-When the user asks for a polish, when the dialogue is the visible weakness, or when the register
-gate keeps noting the same scene: `render dialogue --ep 1` prints the 逐句校对表 (原句 / 问题 / 修改句 /
-修改理由 / 导演／演员提示) with each speaker's speech rules in the last column. Fill it per
-`dialogue-craft.md` §3, apply the changes to the flow, re-validate, and append the filled table to
-the screenplay dataset under `## 对白校对`.
+1. **Read** the framework (`references/dialogue_revision_framework.md`) and the guide of every 语体
+   `seed dialogue` prints; query the corpus for each speaker it lists.
+2. **Diagnose the episode first** (`diagnosis`: `goal` 场面目标, `conflict` 主要冲突, `subtext` 潜台词核心,
+   `voice` 角色声音问题, `rhythm` 节奏问题) and pick `level` 轻修 / 中修 / 重修. A scene with no goal,
+   obstacle or turn is a script problem — fix the beats, not the words.
+3. **Every row**: `issue` (the problem, or `无：<why it stands>`), `revised` (the new line, or the
+   same line), `reason` when it changed, `note` — the 导演／演员提示 as a playable action
+   ("你不是解释，你是在逼对方认错"), never a mood (自然一点 / 生气 are refused). A row may set
+   `delivery` to replace the beat's parenthetical.
+4. **`selfCheck`**: the guide's 自检清单, item by item (`{ item, ok, note }`, at least three); an
+   item that fails says why it stands or the lines are fixed. `lang-zh-hk.md` has no section called
+   自检: for 粤语 the checklist is its §九 (书面中文 → 港式粤语 and the 三个硬指标) plus the skill's
+   `SKILL.md` 品质检查.
+5. `apply dialogue`, then re-validate — a shorter or longer line moves the duration and the dialogue
+   fit downstream.
+
+Editing a line later reopens it: the gate compares every row's `revised` with the script, so a
+changed line fails until it is reviewed again. `seed dialogue` re-run keeps the rows of unchanged
+lines and opens only the moved ones (`--force` resets them all). Cutting or adding a line is done in
+the flow first, then re-seed.
+
+The gate checks: a review exists, the level, all five diagnosis fields, the guide of every speaker's
+语体 listed in `guides`, one row per line and none for a beat that is not a line, the revision
+applied, an issue on every row, a reason on every change, a playable note on every row, and the
+self-check. `render dialogue` prints the filled table — append it to the screenplay dataset under
+`## 对白校对`; `render script`'s 创作自检 gains the row 对白是否已按语体 guide 逐句校对.
 
 ## Sync
 
@@ -115,7 +162,7 @@ wji fs set <id> --merge content=@/tmp/E01-script.md --merge viewMode=content --m
 ## Gates
 
 `script-shape` · `beat-flow` · `duration` · `hook-cliff` · `beats-claimed` · `script-design` ·
-`rhythm` · `filmable` · `dialogue-register` (advisory).
+`rhythm` · `filmable` · `dialogue-review` · `dialogue-register` (advisory).
 
 `render script` writes the reference's episode format — `△（from-to秒）` action lines with the
 seconds derived, `角色（delivery）：line`, `【爽点标记】` after the 爽点, `【钩子标记】` at the end — and

package/skill/short-drama/reference/stage-storyboard.md

@@ -23,7 +23,7 @@ the script and `validate` names the cuts that broke.
 ## Start from the seed
 
 ```bash
-node {baseDir}/scripts/short-drama.mjs seed <project> storyboard --ep 1
+$SD seed <project> storyboard --ep 1
 ```
 
 Three worksheets, all deleted once the segments are written:
@@ -92,6 +92,16 @@ into the prompt.
   notes, the subtitles-off sentence in the prompt's language unless `params.subtitles`, on H3 the
   other person on screen during a line has `lips remain completely closed`, `<d>` holds only the
   line, and in sheets mode every person on screen is bound by `@imageK` inside their own shot.
+- **Reference binding** (gate `ref-binding`) — the picture plan has to survive into the delivery
+  text. Every picture gets **one** legend line saying what it is: a person by age / hair / clothes
+  from the character bible (two people described in the same words is how faces get swapped), 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*). Every person and every prop is bound by its label again **inside
+  its own shot**, the room is named once per segment (`Every shot stays inside @image1 …`), no
+  connected picture may go unused, and a sheets segment may not carry more pictures than the model
+  takes. Wording patterns and the six roles: `reference/reference-binding.md`; proofread text that
+  is not yet in the project file with `reference/prompt-audit.mjs`.
 - **`params.speechLeadSeconds`** (0.5) is added to the dialogue-fit check: a 4.0 s line needs a
   5 s cut.
 - A prop may carry **`textPolicy { mode: readable | symbolic | blank, text }`** (gate
@@ -152,13 +162,13 @@ preset's banned directions (e.g. `low-key` / `costume-dark`: `frontal` / `ambien
 `sweet-romance` / `costume-idol`: `underlight` / `top`), at least half the cuts with people in a
 segment use its preferred ones, the stop stays within the preset's limit, and a period or costume
 preset refuses a source its era did not have. Stage the person against the source first, then pick the camera side; the
-tables are in `cinematography.md` and the preset's guide.
+tables are in `../../short-drama-core/reference/cinematography.md` and the preset's guide.
 
 **Each segment is cut in its grammar** (`segment.grammar`, else `params.grammar`; see
-`grammar.md`): `action` segments open on a wide, take one move per cut and answer every impact
-(`action.md`); `faceslap` segments show the proof and a cascade of reactions
-(`grammar-faceslap.md`); `comedy` segments set up every punchline and give it laugh room
-(`grammar-comedy.md`). Each grammar adds its line to the prompt (`Action:` / `Payoff:` /
+`../../short-drama-core/reference/grammar.md`): `action` segments open on a wide, take one move per cut and answer every impact
+(`../../short-drama-core/reference/action.md`); `faceslap` segments show the proof and a cascade of reactions
+(`../../short-drama-core/reference/grammar-faceslap.md`); `comedy` segments set up every punchline and give it laugh room
+(`../../short-drama-core/reference/grammar-comedy.md`). Each grammar adds its line to the prompt (`Action:` / `Payoff:` /
 `Comedy:`).
 
 ### Foreground
@@ -205,7 +215,7 @@ default and more than two camera moves per segment is showing off.
 Get the skeleton:
 
 ```bash
-node {baseDir}/scripts/short-drama.mjs prompt <project> segment E01-01
+$SD prompt <project> segment E01-01
 ```
 
 It prints the parts that are derived and gated verbatim — the alignment line, `[Shot k] At
@@ -371,12 +381,6 @@ segment is rejected, retry, then trim — do not go hunting for a forbidden word
 To bisect anyway, use a scratch video node at `-d 2`, 480P: free when rejected, ~30 credits when it
 passes, and sample each fragment several times because one pass proves nothing.
 
-**Last resort: the uncensored twin.** `wan-3.0-ref-uncensored` (MuleRouter, no green-net) takes the
-same prompt and the same 10 images + 5 videos + 5 audios, at ~1.3× the price; note that it bills the
-reference-video seconds as well as the output. Switch a stuck segment to it only with the
-Producer's approval, and only after retry-then-trim has failed. `short-drama.mjs` does not know it
-as a mode, so set it as the node's model directly.
-
 ## Sheets mode — no storyboard images (the default)
 
 `params.references: sheets` drops stage 5's image generation entirely: the video node references
@@ -438,9 +442,10 @@ score. So by default (`params.audioTrack: post`) **the clips carry no music**: t
 says `No music.` (H3 `non_diegetic_music: N/A`, Wan `无配乐`), keeps the dialogue, the ambience and
 the in-picture sfx, and drops every sfx marked `post: true` (stingers, risers, whooshes, impact
 hits — sound design outside the story world). The storyboard still writes `sound.music` (intent and
-cues) and the `post: true` hits: they are the **spotting notes**. When every clip is in, cut the
-episode as in `edit-tools.md` (FilmCraft, or ffmpeg if `filmcraft-cli` is missing; the `.fcproj`
-stays in the project), and score that locked picture as one piece on its timecode:
+cues) and the `post: true` hits: they are the **spotting notes**. When every clip is in, the Operator
+cuts the episode as in `../../short-drama-core/reference/edit-tools.md` (FilmCraft, or ffmpeg if
+`filmcraft-cli` is missing; the `.fcproj` stays in the project), and the Sound Designer scores that
+locked picture as one piece on its timecode:
 
 ```bash
 $SD prompt <project> score E01            # planned spotting sheet: SMPTE rows + ≤120 s Seed Audio sections
@@ -450,7 +455,7 @@ The sheet lays the segments end to end — scene beds (BGE), every line (where t
 segment's music cues at their episode time, the `post` sound design, the sfx already in the clips
 (never doubled) and the hook / turns / 爽点 / cliff as the 起承转合 anchors. It is the *planned*
 timeline: the Sound Designer conforms it to the edit's real cuts
-(`mix-over-video.md`) before any prompt is written. `manifest`
+(`../../short-drama-sound/reference/mix-over-video.md`) before any prompt is written. `manifest`
 then asks for the episode's `wji.assets.edit` and `wji.assets.mix`.
 
 ### Per-segment soundtrack node (`audioTrack: bed` / `full`)
@@ -463,8 +468,8 @@ music goes into the clip prompt too; the prompt is derived — not written — f
 the cut timing:
 
 ```bash
-node {baseDir}/scripts/short-drama.mjs prompt <project> audio E01-05            # bed (default)
-node {baseDir}/scripts/short-drama.mjs prompt <project> audio E01-02 --full     # + the lines, in the @AudioN voices
+$SD prompt <project> audio E01-05            # bed (default)
+$SD prompt <project> audio E01-02 --full     # + the lines, in the @AudioN voices
 ```
 
 ```
@@ -526,14 +531,14 @@ recorded in the episode's `wji.displayNodes` as `"C01:voice"`, exactly like the
 ```bash
 wji fs mkdir <shot folder> E01                                                   # bind --ep 1 wji.shotFolderId
 wji dataset new <shot/E01> "E01-01" -t shot -d "<scene · seconds>"               # bind --ep 1 storyboard.segments.E01-01.wji.datasetId
-node {baseDir}/scripts/short-drama.mjs render <project> segment E01-01 > /tmp/E01-01.md
+$SD render <project> segment E01-01 > /tmp/E01-01.md
 wji fs set <id> --merge content=@/tmp/E01-01.md --merge viewMode=content --merge video_prompt=@/tmp/E01-01-prompt.txt
-node {baseDir}/scripts/short-drama.mjs bind <project> --ep 1 storyboard.segments.E01-01.wji.nodes.keyframes.0 <nodeId>
-node {baseDir}/scripts/short-drama.mjs manifest <project> --ep 1                 # must be clean before the episode is "done"
+$SD bind <project> --ep 1 storyboard.segments.E01-01.wji.nodes.keyframes.0 <nodeId>
+$SD manifest <project> --ep 1                 # must be clean before the episode is "done"
 ```
 
 ## Gates
 
 `segments` · `timing` (durations, the speech lead-in **and** the reference caps) · `cut-content` ·
-`photography` · `video-prompt` · `shot-purpose` · `framing-axis` · `continuity` · `motion` · `sound`
-· `visual-plan` · `text-policy` · `prompt-hygiene` · `audio-track` · `storyboard-advice` (advisory).
+`photography` · `video-prompt` · `shot-purpose` · `boundary-contrast` · `framing-axis` · `continuity` · `motion` · `sound`
+· `visual-plan` · `text-policy` · `prompt-hygiene` · `ref-binding` · `audio-track` · `boundary-contrast-episode` (advisory) · `storyboard-advice` (advisory).

package/skill/short-drama/scripts/selftest.mjs

@@ -15,7 +15,7 @@ import {
   renderSegment, runGates, seedScript, seedStoryboard, seedanceAnchors, slug, timestamp, writeJsonFile,
   wanLegendLine, wanSheetLegendLine, wanShotHeader, segmentReferences, imageReferences, audioReferences, voiceSamplePrompt, voiceSampleSeconds, nonWhitespace,
   speechDirection, cameraPhrase, lightPhrase, idlePhrase, idleSentence, IDLE_MAX_CHARS, eyelinePhrase, eyelineValid, onlyInFramePhrase, overShoulderPhrase, sidesLine, foregroundPhrase, foregroundBand, foregroundForLook, focalOf, LIGHT_DIRECTIONS, identityOf, isEnglish,
-  renderGuideline, renderDialogue, scriptSelfCheck, beatOffsets, turnPoints, parseAge, talkAndAction, normalizeRegister, REGISTERS, REGISTER_ACCENT,
+  renderGuideline, renderDialogue, seedDialogue, applyDialogue, dialogueRefs, DIALOGUE_GUIDES, corpusAge, scriptSelfCheck, beatOffsets, turnPoints, parseAge, talkAndAction, normalizeRegister, REGISTERS, REGISTER_ACCENT,
   cutStart, segmentCharacters, keyScenes, shotTexts, motionSentence, cameraClause, PURPOSE_KINDS, ANGLES, HEIGHTS, NO_SUBTITLES, LIPS_CLOSED,
   audioPrompt, musicOf, cueSeconds, AUDIO_MODEL, AUDIO_MAX_CHARS, AUDIO_MAX_SECONDS, scoreSheet, smpte,
   VIDEO_MODELS, VIDEO_MODES, DEFAULT_PARAMS, videoModel, videoResolutionFor, nearestResolution, segmentPlan, videoReferences,
@@ -262,6 +262,35 @@ const breaks = (gateId, project, episode, needle) => {
   const p5 = clone(PROJECT); p5.props[0].states = [];
   breaks('prop-design', p5, null, 'state');
 }
+{
+  const missing = clone(PROJECT);
+  delete missing.scenes[0].condition;
+  delete missing.scenes[0].conditionReason;
+  breaks('scene-design', missing, null, 'condition must be new, lived-in or worn');
+  const fresh = clone(PROJECT);
+  fresh.scenes[0].condition = 'new';
+  fresh.scenes[0].conditionReason = 'the organisers finished this set this morning';
+  breaks('scene-design', fresh, null, 'condition is new but');
+  const clean = clone(PROJECT);
+  clean.scenes[1].condition = 'new';
+  clean.scenes[1].conditionReason = 'a pavilion built for this episode and struck the next day';
+  const scrub = (text) => String(text || '').replace(/\b(worn|rust|rusty|rusted|faded|chipped|torn|frayed|peeling|stain|stained|stains)\b/gi, 'plain').replace(/殘舊|残旧|生鏽|生锈|褪色|剝落|剥落/g, '');
+  for (const a of clean.scenes[1].anchors) { a.name = scrub(a.name); a.desc = scrub(a.desc); }
+  clean.scenes[1].image.prompt = scrub(clean.scenes[1].image.prompt);
+  clean.scenes[1].image.sheet = scrub(clean.scenes[1].image.sheet);
+  for (const l of clean.scenes[1].lighting) l.prompt = scrub(l.prompt);
+  passes('scene-design', clean, null);
+  const ban = clone(PROJECT);
+  ban.scenes[0].image.negativePrompt += ', brand-new, showroom-clean';
+  breaks('scene-design', ban, null, 'only condition worn may ban');
+  const propMissing = clone(PROJECT);
+  delete propMissing.props[0].conditionReason;
+  breaks('prop-design', propMissing, null, 'conditionReason');
+  const propNew = clone(PROJECT);
+  propNew.props[0].condition = 'new';
+  propNew.props[0].conditionReason = 'a drive the organisers just issued';
+  breaks('prop-design', propNew, null, 'condition is new but');
+}
 
 // ── script gates ────────────────────────────────────────────────────────────
 {
@@ -479,7 +508,8 @@ const breaks = (gateId, project, episode, needle) => {
   const results = runGates(makeContext(PROJECT, { 1: e }));
   const adv = results.find((r) => r.id === 'dialogue-register');
   ok(adv.advisory && adv.failures.length === 0 && adv.warnings.length > 0, 'advisory findings come back as warnings, not failures');
-  ok(results.filter((r) => !r.advisory && r.stage !== 'storyboard').every((r) => !r.failures.length), 'a register note never fails validate (the storyboard is stale on purpose here)');
+  ok(results.filter((r) => !r.advisory && r.stage !== 'storyboard' && r.id !== 'dialogue-review').every((r) => !r.failures.length), 'a register note never fails validate (the storyboard is stale on purpose here)');
+  ok(results.find((r) => r.id === 'dialogue-review').failures.some((m) => m.includes('differs from the reviewed one')), 'an edited line is no longer reviewed — dialogue-review asks for it again');
   const p2 = clone(PROJECT); p2.params.register = '粤语';
   ok(failures('dialogue-register', p2, E01).some((m) => m.includes('书面中文硬转')), 'a 粤语 drama with written-Chinese lines is noted');
   const p3 = clone(PROJECT); p3.params.register = '古装';
@@ -1681,19 +1711,23 @@ const breaks = (gateId, project, episode, needle) => {
   const script = renderScript(PROJECT, E01);
   ok(script.includes('穿雨衣的男人（') && script.includes('VO'), 'script markdown names speakers and VO');
   ok(script.includes('△（0-3秒）') && script.includes('【爽点标记】') && script.includes('【钩子标记】'), 'script markdown is in the 短剧 episode format with derived time marks');
-  ok(script.includes('## 本集功能') && script.includes('## 本集任务') && script.includes('## 创作自检') && script.includes('16/16'), 'script markdown carries the episode design and a full self-check');
+  ok(script.includes('## 本集功能') && script.includes('## 本集任务') && script.includes('## 创作自检') && script.includes('18/18') && script.includes('对白是否已按语体 guide 逐句校对'), 'script markdown carries the episode design and a full self-check');
   const guide = renderGuideline(PROJECT);
   ok(guide.includes('创作指引') && guide.includes('## 9. 本项目绝对要注意的重点') && !guide.includes('{{'), 'guideline markdown is complete with no placeholders');
   ok(guide.includes(PROJECT.outline.brief.logline), 'guideline carries the logline');
-  const dlg = renderDialogue(PROJECT, E01);
+  const blank = clone(E01); delete blank.dialogueReview;
+  const dlg = renderDialogue(PROJECT, blank);
   eq((dlg.match(/^\| [^|]*：/gm) || []).length, 13, 'dialogue worksheet has one row per line');
   ok(dlg.includes('禁用：其实、我觉得'), 'dialogue worksheet carries the speaker\'s speech rules');
+  ok(dlg.includes('lang-zh-cn.md') && dlg.includes('ref.mjs --lang zh-cn list --age elder'), 'dialogue worksheet points each speaker at their guide and corpus query');
+  const filled = renderDialogue(PROJECT, E01);
+  ok(filled.includes('| 修改幅度 | 轻修 |') && filled.includes('（保留）') && filled.includes('## 自检'), 'a filled review renders as the 对白校对 table');
   const c1 = renderCharacter(PROJECT, 'C01');
   ok(c1.includes('## 人物圣经') && c1.includes('### 过去的关键选择') && c1.includes('## 人物关系网') && c1.includes('### 写作备忘'), 'character markdown has the bible sections');
   const out = renderOutline(PROJECT);
   ok(out.includes('## 创作简报') && out.includes('## 全剧结构') && out.includes('## 分集目录'), 'outline markdown has brief, structure and episode index');
   const check = scriptSelfCheck(PROJECT, E01);
-  eq(check.score, 16, 'fixture episode scores 16/16');
+  eq(check.score, 18, 'fixture episode scores 18/18');
   const segMd = renderSegment(PROJECT, E01, 'E01-02');
   ok(segMd.includes('| 00:02.000 |') && segMd.includes('integrated_multimodal_description'), 'segment markdown has cut table and prompt');
   ok(segMd.includes('### SHOT 1 · establish') && segMd.includes('| 唯一动作 |') && segMd.includes('| 视觉依据 |') && segMd.includes('轴线：'), 'segment markdown has the SHOT blocks, the axis and the sound line');
@@ -1701,6 +1735,86 @@ const breaks = (gateId, project, episode, needle) => {
   assert.throws(() => renderCharacter(PROJECT, 'C99'), /no character/, 'render refuses unknown ids');
 }
 
+// ── dialogue review (w-film-dialogue-editor) ────────────────────────────────
+{
+  passes('dialogue-review', PROJECT, E01);
+  ok(GATES.find((g) => g.id === 'dialogue-review').advisory === false, 'dialogue-review blocks');
+  eq(DIALOGUE_GUIDES['粤语'].lang, 'zh-hk', '粤语 reads the zh-hk guide and corpus');
+  eq(DIALOGUE_GUIDES['古装'].corpus, false, '古装 has a guide but no corpus');
+  eq(corpusAge(24), 'young', 'a 24-year-old queries the young band');
+  eq(corpusAge(61), 'elder', 'a 61-year-old queries the elder band');
+  const none = clone(E01); delete none.dialogueReview;
+  breaks('dialogue-review', PROJECT, none, 'seed <project> dialogue');
+  const pOff = clone(PROJECT); pOff.params.dialogueReview = 'off';
+  passes('dialogue-review', pOff, none);
+  const lvl = clone(E01); lvl.dialogueReview.level = '';
+  breaks('dialogue-review', PROJECT, lvl, 'level');
+  const diag = clone(E01); diag.dialogueReview.diagnosis.subtext = '';
+  breaks('dialogue-review', PROJECT, diag, '潜台词核心');
+  const guide = clone(E01); guide.dialogueReview.guides = [];
+  breaks('dialogue-review', PROJECT, guide, 'lang-zh-cn.md');
+  const yue = clone(PROJECT); yue.characters.find((c) => c.id === 'C02').bible.speech.register = '粤语';
+  breaks('dialogue-review', yue, E01, 'lang-zh-hk.md');
+  const miss = clone(E01); miss.dialogueReview.rows.pop();
+  breaks('dialogue-review', PROJECT, miss, 'no review row');
+  const stale = clone(E01); stale.scenes[1].flow[2].line = '又轮到你值夜班。';
+  breaks('dialogue-review', PROJECT, stale, 'differs from the reviewed one');
+  const noIssue = clone(E01); noIssue.dialogueReview.rows[0].issue = '';
+  breaks('dialogue-review', PROJECT, noIssue, 'issue is empty');
+  const noWhy = clone(E01); noWhy.dialogueReview.rows[0].revised = '帮我收着，明天有人拿。'; noWhy.scenes[0].flow[1].line = '帮我收着，明天有人拿。';
+  breaks('dialogue-review', PROJECT, noWhy, 'without a reason');
+  const mood = clone(E01); mood.dialogueReview.rows[0].note = '自然一点';
+  breaks('dialogue-review', PROJECT, mood, 'a mood, not an action');
+  const emo = clone(E01); emo.dialogueReview.rows[1].note = '生气';
+  breaks('dialogue-review', PROJECT, emo, 'a mood, not an action');
+  const check = clone(E01); check.dialogueReview.selfCheck = check.dialogueReview.selfCheck.slice(0, 2);
+  breaks('dialogue-review', PROJECT, check, '自检清单');
+  const checkNo = clone(E01); checkNo.dialogueReview.selfCheck[0] = { item: '读出声', ok: false, note: '' };
+  breaks('dialogue-review', PROJECT, checkNo, 'failed with no note');
+  const ghost = clone(E01); ghost.dialogueReview.rows.push({ scene: 1, beat: 1, speaker: 'C01', original: 'x', issue: 'x', revised: 'x', reason: '', note: '逼' });
+  breaks('dialogue-review', PROJECT, ghost, 'no longer a line');
+
+  // seed keeps filled rows for unchanged lines and opens only the edited one
+  const edited = clone(E01); edited.scenes[1].flow[2].line = '又轮到你值夜班。';
+  const reseeded = seedDialogue(PROJECT, edited, { skillDir: null });
+  eq(reseeded.rows.length, 13, 're-seed keeps one row per line');
+  eq(reseeded.rows.filter((r) => !r.issue).length, 1, 're-seed reopens only the edited line');
+  eq(reseeded.level, '轻修', 're-seed keeps the level and diagnosis');
+  eq(seedDialogue(PROJECT, edited, { skillDir: null, force: true }).rows.filter((r) => !r.issue).length, 13, '--force resets every row');
+  ok(!('refs' in reseeded), 'the stored review carries no machine paths');
+  const refs = dialogueRefs(PROJECT, E01, { skillDir: '/x' });
+  ok(refs.speakers.every((x) => x.guide === '/x/references/lang-zh-cn.md'), 'every speaker gets the zh-cn guide');
+  ok(refs.speakers.find((x) => x.id === 'C02').query.includes('--age elder'), "the sixty-year-old's corpus query is the elder band");
+  ok(!refs.speakers.some((x) => x.id === 'VO'), 'VO is not a speaker with a guide');
+
+  // apply writes revised lines and deliveries back, then the gate is green again
+  const app = clone(E01); app.dialogueReview.rows[0].revised = '帮我收着，明天有人拿。'; app.dialogueReview.rows[0].reason = '更短，一口气说完'; app.dialogueReview.rows[0].delivery = '压着嗓子，塞完就走';
+  breaks('dialogue-review', PROJECT, app, 'differs from the reviewed one');
+  eq(applyDialogue(app), 2, 'apply writes the line and the delivery');
+  eq(app.scenes[0].flow[1].line, '帮我收着，明天有人拿。', 'the revised line is in the script');
+  passes('dialogue-review', PROJECT, app);
+  const bad = clone(E01); bad.dialogueReview.rows[0].revised = '';
+  assert.throws(() => applyDialogue(bad), /revised is empty/, 'apply refuses an empty line'); passed += 1;
+  ok(scriptSelfCheck(PROJECT, none).rows.some((r) => r[0].includes('逐句校对') && r[1] === false), 'the 创作自检 says 否 without a review');
+
+  // seed / apply through the CLI
+  const dir = mkdtempSync(join(tmpdir(), 'sd-dlg-'));
+  try {
+    writeFileSync(join(dir, 'project.json'), JSON.stringify(PROJECT));
+    spawnSync('mkdir', ['-p', join(dir, 'episodes')]);
+    writeFileSync(join(dir, 'episodes/E01.json'), JSON.stringify(none));
+    const engine = join(here, 'short-drama.mjs');
+    const r1 = spawnSync(process.execPath, [engine, 'seed', dir, 'dialogue', '--ep', '1'], { encoding: 'utf8' });
+    eq(r1.status, 0, `seed dialogue runs: ${r1.stderr}`);
+    ok(r1.stdout.includes('13 lines (13 to review)') && r1.stdout.includes('lang-zh-cn.md'), 'seed dialogue prints the guide per speaker');
+    const r2 = spawnSync(process.execPath, [engine, 'apply', dir, 'dialogue', '--ep', '1'], { encoding: 'utf8' });
+    eq(r2.status, 0, `apply dialogue runs: ${r2.stderr}`);
+    ok(r2.stdout.includes('0 change(s)'), 'apply on an untouched review changes nothing');
+  } finally {
+    rmSync(dir, { recursive: true, force: true });
+  }
+}
+
 // ── manifest & bind ─────────────────────────────────────────────────────────
 {
   const items = manifest(PROJECT, { 1: E01 });
@@ -2414,10 +2528,41 @@ const breaks = (gateId, project, episode, needle) => {
   eq(slug('E01-01 便利店 / 夜班'), 'e01-01-便利店-夜班', 'slug keeps CJK, drops separators');
 }
 
+// ── ref-binding (参考图绑定) ─────────────────────────────────────────────────
+{
+  // The picture plan has to survive into the delivery text: a legend line that names WHO/WHAT each
+  // picture is, a "not its composition / its people" clause on everything that is not a person, and
+  // the label named again inside every shot that depends on it.
+  const p = clone(PROJECT);
+  p.params.video = 'seedance';
+  p.params.references = 'sheets';
+  p.params.videoDuration = 'auto';
+  const e = clone(E01);
+  const seg = e.storyboard.segments[3];
+  const refs = segmentReferences(p, e, seg).filter((r) => r.kind === 'image');
+  const legend = refs.map((r) => `${r.label} is ${r.use}.`).join('\n');
+  const shots = seg.cuts
+    .map((c, k) => `Shot ${k + 1} (${k}.0\u2013${k + 1}.0 s, medium shot, Static Shot): ${refs.map((r) => r.label).join(' and ')} hold.`)
+    .join('\n');
+  const good = `${legend}\n\n${shots}`;
+  const bound = clone(e);
+  bound.storyboard.segments[3].prompt = good;
+  eq(failures('ref-binding', p, bound).filter((m) => m.startsWith('E01-04')).length, 0, 'ref-binding passes a fully bound prompt');
+  const generic = clone(e);
+  generic.storyboard.segments[3].prompt = good.replace(/@image2 is the character sheet of [^\n]*/, '@image2 is the character sheet of one person in this segment \u2014 match face, hair and clothes exactly.');
+  breaks('ref-binding', p, generic, 'never says WHICH person');
+  const noNegative = clone(e);
+  noNegative.storyboard.segments[3].prompt = good.replace(/@image1 is the environment[^\n]*/, '@image1 is the environment \u2014 match its layout, materials and wear exactly.');
+  breaks('ref-binding', p, noNegative, 'never says what NOT');
+  const unused = clone(e);
+  unused.storyboard.segments[3].prompt = `${legend}\n\n${shots.replaceAll('@image2', 'a person')}`;
+  breaks('ref-binding', p, unused, 'used in no shot');
+}
+
 // ── gate inventory (crew engine) ────────────────────────────────────────────
 {
-  eq(GATES.length, 69, 'gate count');
-  eq(GATES.filter((g) => !g.advisory).length, 63, 'hard gate count');
+  eq(GATES.length, 71, 'gate count');
+  eq(GATES.filter((g) => !g.advisory).length, 65, 'hard gate count');
   eq(GATES.filter((g) => g.advisory).length, 6, 'advisory gate count');
   ok(!GATES.find((g) => g.id === 'boundary-contrast').advisory, 'boundary-contrast is hard');
   ok(GATES.find((g) => g.id === 'boundary-contrast-episode').advisory, 'boundary-contrast-episode is advisory');
@@ -2539,8 +2684,8 @@ const breaks = (gateId, project, episode, needle) => {
   const results = runGates(makeContext(PROJECT, { 1: E01 }));
   const rep = formatGates(results);
   eq(Object.keys(rep).sort().join(','), 'advisory,failed,hard,lines,notes,summary,text', 'formatGates shape');
-  eq(rep.summary, '✓ 63 gate(s) passed, 6 advisory (1 note)', 'fixture summary');
-  eq(rep.failed, 0, 'nothing failed'); eq(rep.hard, 63, 'hard count'); eq(rep.advisory, 6, 'advisory count');
+  eq(rep.summary, '✓ 65 gate(s) passed, 6 advisory (1 note)', 'fixture summary');
+  eq(rep.failed, 0, 'nothing failed'); eq(rep.hard, 65, 'hard count'); eq(rep.advisory, 6, 'advisory count');
   // The fixture now raises one advisory (E01-03 carries the 爽点 and is not cut as faceslap),
   // exactly as it does in the wji-cli copy — advisory lines print, failures still must not.
   eq(rep.lines.filter((l) => l.includes('✗')).length, 0, 'a green run prints no failed gates');