@lingjingai/scriptctl 0.35.0 → 0.36.0

package/dist/llm/tasks/ingest/roster-merge.js

@@ -0,0 +1,235 @@
+// Video roster merge (§10.2 pass3, ported from cezhi-verify 03b). Cross-episode
+// identity resolution driven by VISUAL ANCHORS — the same character transcribed
+// independently per episode gets different names (本名 vs 职务/外貌 vs 听错), and
+// only anchor + role continuity can re-unify them. Four stages, all VERBATIM from
+// the tuned scaffold prompts:
+//   A recall  — one global pass, recall-first, over-cluster candidate groups
+//   B refine  — per candidate group, split mis-merges + pick canonical
+//   C canon   — pick a canonical name for a confirmed cluster (本名优先)
+//   D remerge — global recompose, merge-only, recover B's conservative splits
+// Structured output via forced tool calls; the tool-input JSON schemas carry the
+// scaffold's field descriptions verbatim (they are tuned prompt content).
+function joinEvidence(xs, max = 4) {
+    const v = xs.map((s) => s.trim()).filter(Boolean).slice(0, max);
+    return v.length ? v.join(" ／ ") : "无";
+}
+function evidenceBlock(m) {
+    return `- 「${m.name}」 ep[${m.episodes.join(",")}]\n    锚点: ${joinEvidence(m.anchors)}\n    作用: ${joinEvidence(m.roles, 3)}\n    外观: ${joinEvidence(m.appearances, 3)}`;
+}
+// ===== Stage A: recall (high recall candidate grouping) =====
+export const RECALL_SYSTEM = "你是长篇剧集人物花名册的『候选粗聚』器，是两阶段归并的第一步。" +
+    "同一角色跨集常被记成不同名字（听到本名 vs 按职务/外貌称呼、听错写成近音字、某集认错人）。" +
+    "你会拿到完整花名册，每个名字带跨集锚点、外观、剧情作用、出现集号。" +
+    "任务：把**可能是同一个人**的名字聚成候选组。" +
+    "**这一步召回优先**：后面有专门的精判环节会逐组复核、把不该合的拆出去，所以你现在宁可多合、别漏——" +
+    "只要两个名字在视觉锚点或剧情身份上**有合理的同人可能**，就放进同一候选组，哪怕不完全确定。" +
+    "唯一的硬边界：锚点直接冲突（性别不同、年龄段明显不符）的绝不放同组。" +
+    "canonical 这一步随便选组内一个名字即可（后面会重定）。每组至少 2 个名字才有意义。";
+// MergeDecisionSchema (recall reuses it) — verbatim field descriptions.
+const MERGE_DECISION_SCHEMA = {
+    type: "object",
+    additionalProperties: false,
+    required: ["groups"],
+    properties: {
+        groups: {
+            type: "array",
+            description: "每组=一个真实人物+其所有别名。只输出确实需要合并（至少 1 个别名）的组；独一无二的人物不必输出",
+            items: {
+                type: "object",
+                additionalProperties: false,
+                required: ["canonical", "aliases", "reason"],
+                properties: {
+                    canonical: { type: "string", description: "这一组同人的规范名。优先用有名有姓的本名，其次用信息最完整的称呼；避免用纯职务/敬称/临时描述作为规范名（除非整组都没有本名）" },
+                    aliases: { type: "array", items: { type: "string" }, description: "与 canonical 确认为同一个人的其他名字（不含 canonical）。只放视觉锚点确实吻合的" },
+                    reason: { type: "string", description: "判定同人的依据：视觉锚点/剧情作用/出现集，一句话" },
+                },
+            },
+        },
+    },
+};
+export function mergePrompt(actors) {
+    const lines = actors.map(evidenceBlock).join("\n");
+    return `下面是整部剧的完整人物花名册（${actors.length} 个名字）。找出所有『其实是同一个人』的名字并分组：
+
+${lines}
+
+要求：
+- 只输出需要合并的组（每组至少 1 个别名）；独一无二的人物不用输出。
+- 每组选 canonical（优先本名，不用职务/敬称），其余同人放 aliases。
+- 判据是视觉锚点吻合 + 剧情作用连续；名字字面是否相同不重要。
+- 锚点确实吻合就合，别因别名多而犹豫；真拿不准（锚点冲突或信息太少）才保守不合。
+- 尤其：两个名字都只出现在很少几集、锚点又都很泛时，证据不足，倾向不合。
+严格按 schema 返回。`;
+}
+// ===== Stage B: refine (per-group high-precision split) =====
+export const REFINE_SYSTEM = "你是人物同人判定的『精判』器，是两阶段归并的第二步。" +
+    "上一步用宽松口径把一批**可能同人**的名字聚成了一个候选组。你的任务：只看这一个候选组，把它**拆成若干真正是同一个人的子簇**。" +
+    "判据**同等看重两类**：①视觉锚点（发型/发色/脸型五官/体型/年龄段/性别/疤痣眼镜胡须配饰等标志特征）；②剧情身份（所属队伍/组织/阵营/职务定位、与主角的关系、出场场合）。" +
+    "**关键：外貌像 ≠ 同人。** 一部剧里常有多个外貌高度相似、只能靠身份区分的角色（例如同类职务的一批人：几个平头运动服的男教练、一批花白头发西装的中老年高管、几个同款球衣的球员）——" +
+    "这些人视觉锚点几乎一样，光看外貌必然误合。**必须先看剧情身份：所属不同（不同队/不同组织/不同阵营/不同场合）、或角色定位矛盾，就是不同的人，哪怕长相一模一样也绝不合。**" +
+    "**核心心态：拆和合两个方向的错误一样有害——既不能把不同人误合，也不能把同一个人拆散。** " +
+    "拆的唯一正当理由是『存在真实矛盾』（身份冲突或锚点冲突）；**没有矛盾就必须合，别因为名字字面不同就拆。** " +
+    "候选组本就是同一角色跨集被记成不同名字的高发结果——一个人在不同集常被同时叫本名、职务、编号、外貌描述、临时称呼（例如同一个『某号+某发型』的配角，可能分别被记成本名、『X号队员』、『某发型队员』、『队长』等），别名多、叫法杂是正常现象，不是拆的理由。" +
+    "原则：" +
+    "① **身份一致 + 锚点吻合 → 必须合，这是硬性要求**：只要几个名字的所属（同一支队伍/组织）、角色定位、以及具体视觉锚点（如同一组『编号球衣+发型+配饰』特征）都对得上，就是同一个人，**哪怕一个是编号称呼、一个是职务称呼、一个是本名、一个是外貌描述，也必须合进一簇。名字字面不同绝不是拆的理由。**" +
+    "② **只在真实矛盾时才拆**：（a）身份冲突——分属对立的不同队伍/组织/阵营、角色定位对立（一个主角亲信一个对手阵营）、活动场合完全不同；（b）锚点冲突——性别不同、年龄段明显不符、发型/脸型明显对不上（如一个大背头一个寸头）。**有明确的这类矛盾才拆，没有就合。**" +
+    "③ 身份冲突时优先级高于外貌像：不同队/不同阵营的两个人，即使发型脸型服装都像也必须拆（防止把对立两方的相似角色误合）。" +
+    "④ 拆一个名字前先自问：它和簇里其他人，有哪一条身份或锚点是**明确冲突**的？答不上来（只是名字不同、或描述简单）就**不许拆**。" +
+    "⑤ 同名不一定同人：字面相同但身份/锚点明确冲突，才是两个人。" +
+    "⑥ canonical 强制本名优先（见 schema）。";
+const REFINE_DECISION_SCHEMA = {
+    type: "object",
+    additionalProperties: false,
+    required: ["clusters"],
+    properties: {
+        clusters: {
+            type: "array",
+            description: "把输入候选组拆成若干真正同人的子簇。不该合的名字各自单独成簇（members 只含自己）。",
+            items: {
+                type: "object",
+                additionalProperties: false,
+                required: ["canonical", "members", "reason"],
+                properties: {
+                    canonical: { type: "string", description: "这个子簇的规范名。**强制本名优先**：只要簇内有一个是像本名的（有名有姓/像人名），就必须选它，不要选职务('X总''董事长''助理')、外貌('黑衣女子')、临时称呼。全簇都没本名时才退而用信息最全的称呼" },
+                    members: { type: "array", items: { type: "string" }, description: "确认是同一个人的名字（含 canonical 本身）。至少 1 个" },
+                    reason: { type: "string", description: "为什么这些是同人、以及（若有）为什么把某些名字排除出去，一句话" },
+                },
+            },
+        },
+    },
+};
+export function refinePrompt(members) {
+    const lines = members.map(evidenceBlock).join("\n");
+    return `上一步把下面这 ${members.length} 个名字聚成了一个「可能同人」候选组。请复核，把它拆成真正是同一个人的子簇：
+
+${lines}
+
+要求：
+- **判同人前先对身份**：看每个名字的「作用」——所属队伍/组织/阵营、职务定位、与主角关系。身份一致再看外貌是否吻合。
+- **身份不同就拆，哪怕外貌一样**：分属对立两方、角色定位矛盾、活动场合完全不同的，是不同的人，即使发型脸型服装都像也必须拆开。（尤其一批同类职务、同款服装的相似角色，必须靠身份区分，别因为外貌像就合。）
+- 身份一致 + 外貌吻合的若干称呼（哪怕措辞略有出入、一个本名一个职务、只出现一集），属于同人，合。
+- 每个子簇选 canonical：**本名优先**，有本名就用本名，别用职务/外貌/临时称呼。
+- 覆盖输入的每一个名字：每个名字必须恰好出现在一个子簇里。
+严格按 schema 返回。`;
+}
+// ===== Stage C: canonical selection for a confirmed cluster =====
+export const CANON_SYSTEM = "你要为『同一个人物的一组不同称呼』挑一个规范名(canonical)。这些名字已确认是同一个人。" +
+    "挑选优先级：① 有名有姓、像真实姓名的（如「张伟」「李文博」）最优先；" +
+    "② 其次是含姓的称呼（如「张总」「李董」）；③ 最不优先的是纯职务/身份/外貌/占位描述（如「助理」「秘书」「黑衣女子」「前上司」「男员工」）和英文/编号名。" +
+    "在同一优先级里，选出现集数多、信息更完整的。" +
+    "只能从给定列表里原样挑一个，不要创造新名字。";
+const CANON_SCHEMA = {
+    type: "object",
+    additionalProperties: false,
+    required: ["canonical"],
+    properties: {
+        canonical: { type: "string", description: "从给定名字里挑一个当规范名。**必须原样是列表中的某个名字**，不要改写、不要合成新名。" },
+    },
+};
+export function canonPrompt(members) {
+    const lines = members.map((m) => `- ${m.name}（出现 ${m.episodes.length} 集）`).join("\n");
+    return `这些名字是同一个人的不同称呼，挑一个当规范名（本名优先，其次含姓称呼，最后才用职务/描述名）：\n\n${lines}\n\n只从上面原样选一个，按 schema 返回。`;
+}
+// ===== Stage D: global remerge (merge-only) =====
+export const REMERGE_SYSTEM = "你是人物花名册的『全局复合』器，是归并的最后一步。前面的步骤已把名字聚成了一批规范名，每个规范名代表一个人。" +
+    "但前面的精判偏保守（宁可拆开也别误合），可能把**同一个人**拆成了几个独立规范名——" +
+    "例如同一个角色，在不同集分别被记成本名、编号称呼、职务、外貌描述，前面没能全部并到一起，就散成了好几个条目。" +
+    "你的**唯一任务**：通读全部规范名 + 它们的视觉锚点、剧情作用、出现集，找出其中**其实是同一个人**的，把它们合并。" +
+    "**只做合并，绝不拆分**：你不能把任何一个规范名拆开，只能指出'这几个规范名是同一个人、应合成一个'。" +
+    "判据：①所属/身份一致（同一支队伍/组织、同一角色定位、同一关系）；②视觉锚点吻合（同一组具体特征：发型+配饰+服装编号等）。两者都对得上才合。" +
+    "铁律：" +
+    "① **只合确实同人的**：所属身份一致 + 锚点吻合，才判同人合并。典型可合信号：同一批具体特征（如同号球衣+同发型+同配饰）反复出现，只是被记成了不同称呼（本名/编号/职务/外貌）。" +
+    "② **身份或锚点冲突的绝不合**：分属对立的不同队伍/组织/阵营、角色定位相反、或外貌锚点直接冲突（性别不同、一个背头一个寸头、年龄段明显不符）——这些是不同的人，即使名字或球衣号相近也不能合。" +
+    "③ 拿不准就**不合**：只输出你有把握是同一人的合并；宁可漏合，不可错合。" +
+    "④ into 与 names 里的每个名字，都必须原样是给定列表中的规范名，不要造新名。";
+const REMERGE_SCHEMA = {
+    type: "object",
+    additionalProperties: false,
+    required: ["merges"],
+    properties: {
+        merges: {
+            type: "array",
+            description: "需要再合并的组。只输出**确实是同一个人被拆散**的组；拿不准就不要合。没有要合的返回空数组。",
+            items: {
+                type: "object",
+                additionalProperties: false,
+                required: ["into", "names", "reason"],
+                properties: {
+                    into: { type: "string", description: "合并后保留的规范名（本名优先；从这一组里挑最像本名的，或出现集数最多的）。必须原样是给定列表中的某个名字。" },
+                    names: { type: "array", items: { type: "string" }, description: "要合并到 into 的其他规范名（不含 into 自身），至少 1 个。必须都是给定列表中的名字。" },
+                    reason: { type: "string", description: "为什么这些是同一个人：所属/身份 + 视觉锚点如何吻合，一句话。" },
+                },
+            },
+        },
+    },
+};
+export function remergePrompt(entries) {
+    const lines = entries.map((e) => {
+        const anc = joinEvidence(e.anchors, 3);
+        const role = joinEvidence(e.roles, 2);
+        return `- 「${e.name}」 ep[${e.episodes.join(",")}]  锚点: ${anc}  作用: ${role}`;
+    }).join("\n");
+    return `下面是全剧当前的 ${entries.length} 个规范名（每个本应代表一个人）。请找出其中**其实是同一个人、被拆成了多个规范名**的，合并它们：
+
+${lines}
+
+要求：
+- 只合并**确实是同一个人**的规范名：所属/身份一致 + 视觉锚点吻合，才合。
+- 分属不同队伍/组织/阵营、角色定位对立、或外貌锚点冲突（如一个背头一个寸头）的，绝不合。
+- 每组指定 into（合并后保留的名，本名优先）和要并入的其他规范名。
+- 拿不准就不要合。没有要合的返回空 merges 数组。
+严格按 schema 返回。`;
+}
+// ===== run helpers (forced tool call → structured object) =====
+const asList = (v) => (Array.isArray(v) ? v : []);
+const str = (v) => (typeof v === "string" ? v : "");
+const strList = (v) => asList(v).map(str).filter(Boolean);
+function tool(name, description, schema) {
+    return { name, description, inputSchema: schema };
+}
+export async function runRosterRecall(executor, entries, signal) {
+    const res = await executor.generateTool("ingest.roster-recall", {
+        system: RECALL_SYSTEM,
+        user: mergePrompt(entries),
+        tool: tool("submit_groups", "提交同人候选分组", MERGE_DECISION_SCHEMA),
+        ...(signal ? { signal } : {}),
+    });
+    const input = res.input;
+    return asList(input?.["groups"]).map((g) => {
+        const o = g;
+        return { canonical: str(o["canonical"]), aliases: strList(o["aliases"]), reason: str(o["reason"]) };
+    });
+}
+export async function runRosterRefine(executor, members, signal) {
+    const res = await executor.generateTool("ingest.roster-refine", {
+        system: REFINE_SYSTEM,
+        user: refinePrompt(members),
+        tool: tool("submit_clusters", "提交同人子簇拆分", REFINE_DECISION_SCHEMA),
+        ...(signal ? { signal } : {}),
+    });
+    const input = res.input;
+    return asList(input?.["clusters"]).map((c) => strList(c["members"]));
+}
+export async function runRosterCanon(executor, members, signal) {
+    const res = await executor.generateTool("ingest.roster-canon", {
+        system: CANON_SYSTEM,
+        user: canonPrompt(members),
+        tool: tool("submit_canonical", "提交规范名", CANON_SCHEMA),
+        ...(signal ? { signal } : {}),
+    });
+    return str(res.input?.["canonical"]);
+}
+export async function runRosterRemerge(executor, entries, signal) {
+    const res = await executor.generateTool("ingest.roster-remerge", {
+        system: REMERGE_SYSTEM,
+        user: remergePrompt(entries),
+        tool: tool("submit_merges", "提交需再合并的规范名组", REMERGE_SCHEMA),
+        ...(signal ? { signal } : {}),
+    });
+    const input = res.input;
+    return asList(input?.["merges"]).map((m) => {
+        const o = m;
+        return { into: str(o["into"]), names: strList(o["names"]), reason: str(o["reason"]) };
+    });
+}
+//# sourceMappingURL=roster-merge.js.map

package/dist/llm/tasks/ingest/roster-merge.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"roster-merge.js","sourceRoot":"","sources":["../../../../src/llm/tasks/ingest/roster-merge.ts"],"names":[],"mappings":"AAAA,gFAAgF;AAChF,gFAAgF;AAChF,0EAA0E;AAC1E,kFAAkF;AAClF,8BAA8B;AAC9B,6EAA6E;AAC7E,uEAAuE;AACvE,qEAAqE;AACrE,8EAA8E;AAC9E,iFAAiF;AACjF,0EAA0E;AAY1E,SAAS,YAAY,CAAC,EAAY,EAAE,GAAG,GAAG,CAAC;IACzC,MAAM,CAAC,GAAG,EAAE,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC;IAChE,OAAO,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC;AACxC,CAAC;AAED,SAAS,aAAa,CAAC,CAAmB;IACxC,OAAO,MAAM,CAAC,CAAC,IAAI,QAAQ,CAAC,CAAC,QAAQ,CAAC,IAAI,CAAC,GAAG,CAAC,cAAc,YAAY,CAAC,CAAC,CAAC,OAAO,CAAC,aAAa,YAAY,CAAC,CAAC,CAAC,KAAK,EAAE,CAAC,CAAC,aAAa,YAAY,CAAC,CAAC,CAAC,WAAW,EAAE,CAAC,CAAC,EAAE,CAAC;AACzK,CAAC;AAED,+DAA+D;AAE/D,MAAM,CAAC,MAAM,aAAa,GACxB,iCAAiC;IACjC,iDAAiD;IACjD,mCAAmC;IACnC,0BAA0B;IAC1B,qDAAqD;IACrD,iDAAiD;IACjD,oCAAoC;IACpC,iDAAiD,CAAC;AAEpD,wEAAwE;AACxE,MAAM,qBAAqB,GAAe;IACxC,IAAI,EAAE,QAAQ;IACd,oBAAoB,EAAE,KAAK;IAC3B,QAAQ,EAAE,CAAC,QAAQ,CAAC;IACpB,UAAU,EAAE;QACV,MAAM,EAAE;YACN,IAAI,EAAE,OAAO;YACb,WAAW,EAAE,mDAAmD;YAChE,KAAK,EAAE;gBACL,IAAI,EAAE,QAAQ;gBACd,oBAAoB,EAAE,KAAK;gBAC3B,QAAQ,EAAE,CAAC,WAAW,EAAE,SAAS,EAAE,QAAQ,CAAC;gBAC5C,UAAU,EAAE;oBACV,SAAS,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE,WAAW,EAAE,iEAAiE,EAAE;oBAC7G,OAAO,EAAE,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE,EAAE,WAAW,EAAE,oDAAoD,EAAE;oBACxH,MAAM,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE,WAAW,EAAE,2BAA2B,EAAE;iBACrE;aACF;SACF;KACF;CACF,CAAC;AAEF,MAAM,UAAU,WAAW,CAAC,MAA0B;IACpD,MAAM,KAAK,GAAG,MAAM,CAAC,GAAG,CAAC,aAAa,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IACnD,OAAO,kBAAkB,MAAM,CAAC,MAAM;;EAEtC,KAAK;;;;;;;;eAQQ,CAAC;AAChB,CAAC;AAED,+DAA+D;AAE/D,MAAM,CAAC,MAAM,aAAa,GACxB,4BAA4B;IAC5B,oEAAoE;IACpE,4FAA4F;IAC5F,+FAA+F;IAC/F,yFAAyF;IACzF,iDAAiD;IACjD,yDAAyD;IACzD,gIAAgI;IAChI,KAAK;IACL,4JAA4J;IAC5J,0IAA0I;IAC1I,8DAA8D;IAC9D,qEAAqE;IACrE,iCAAiC;IACjC,+BAA+B,CAAC;AAElC,MAAM,sBAAsB,GAAe;IACzC,IAAI,EAAE,QAAQ;IACd,oBAAoB,EAAE,KAAK;IAC3B,QAAQ,EAAE,CAAC,UAAU,CAAC;IACtB,UAAU,EAAE;QACV,QAAQ,EAAE;YACR,IAAI,EAAE,OAAO;YACb,WAAW,EAAE,+CAA+C;YAC5D,KAAK,EAAE;gBACL,IAAI,EAAE,QAAQ;gBACd,oBAAoB,EAAE,KAAK;gBAC3B,QAAQ,EAAE,CAAC,WAAW,EAAE,SAAS,EAAE,QAAQ,CAAC;gBAC5C,UAAU,EAAE;oBACV,SAAS,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE,WAAW,EAAE,0GAA0G,EAAE;oBACtJ,OAAO,EAAE,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE,EAAE,WAAW,EAAE,mCAAmC,EAAE;oBACvG,MAAM,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE,WAAW,EAAE,iCAAiC,EAAE;iBAC3E;aACF;SACF;KACF;CACF,CAAC;AAEF,MAAM,UAAU,YAAY,CAAC,OAA2B;IACtD,MAAM,KAAK,GAAG,OAAO,CAAC,GAAG,CAAC,aAAa,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IACpD,OAAO,WAAW,OAAO,CAAC,MAAM;;EAEhC,KAAK;;;;;;;;eAQQ,CAAC;AAChB,CAAC;AAED,mEAAmE;AAEnE,MAAM,CAAC,MAAM,YAAY,GACvB,kDAAkD;IAClD,qCAAqC;IACrC,iFAAiF;IACjF,wBAAwB;IACxB,wBAAwB,CAAC;AAE3B,MAAM,YAAY,GAAe;IAC/B,IAAI,EAAE,QAAQ;IACd,oBAAoB,EAAE,KAAK;IAC3B,QAAQ,EAAE,CAAC,WAAW,CAAC;IACvB,UAAU,EAAE;QACV,SAAS,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE,WAAW,EAAE,8CAA8C,EAAE;KAC3F;CACF,CAAC;AAEF,MAAM,UAAU,WAAW,CAAC,OAAoD;IAC9E,MAAM,KAAK,GAAG,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,KAAK,CAAC,CAAC,IAAI,OAAO,CAAC,CAAC,QAAQ,CAAC,MAAM,KAAK,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IACtF,OAAO,sDAAsD,KAAK,4BAA4B,CAAC;AACjG,CAAC;AAED,mDAAmD;AAEnD,MAAM,CAAC,MAAM,cAAc,GACzB,wDAAwD;IACxD,6CAA6C;IAC7C,wDAAwD;IACxD,+DAA+D;IAC/D,sDAAsD;IACtD,0EAA0E;IAC1E,KAAK;IACL,+FAA+F;IAC/F,oGAAoG;IACpG,wCAAwC;IACxC,8CAA8C,CAAC;AAEjD,MAAM,cAAc,GAAe;IACjC,IAAI,EAAE,QAAQ;IACd,oBAAoB,EAAE,KAAK;IAC3B,QAAQ,EAAE,CAAC,QAAQ,CAAC;IACpB,UAAU,EAAE;QACV,MAAM,EAAE;YACN,IAAI,EAAE,OAAO;YACb,WAAW,EAAE,iDAAiD;YAC9D,KAAK,EAAE;gBACL,IAAI,EAAE,QAAQ;gBACd,oBAAoB,EAAE,KAAK;gBAC3B,QAAQ,EAAE,CAAC,MAAM,EAAE,OAAO,EAAE,QAAQ,CAAC;gBACrC,UAAU,EAAE;oBACV,IAAI,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE,WAAW,EAAE,uDAAuD,EAAE;oBAC9F,KAAK,EAAE,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE,EAAE,WAAW,EAAE,mDAAmD,EAAE;oBACrH,MAAM,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE,WAAW,EAAE,kCAAkC,EAAE;iBAC5E;aACF;SACF;KACF;CACF,CAAC;AAEF,MAAM,UAAU,aAAa,CAAC,OAAwF;IACpH,MAAM,KAAK,GAAG,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE;QAC9B,MAAM,GAAG,GAAG,YAAY,CAAC,CAAC,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC;QACvC,MAAM,IAAI,GAAG,YAAY,CAAC,CAAC,CAAC,KAAK,EAAE,CAAC,CAAC,CAAC;QACtC,OAAO,MAAM,CAAC,CAAC,IAAI,QAAQ,CAAC,CAAC,QAAQ,CAAC,IAAI,CAAC,GAAG,CAAC,UAAU,GAAG,SAAS,IAAI,EAAE,CAAC;IAC9E,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IACd,OAAO,YAAY,OAAO,CAAC,MAAM;;EAEjC,KAAK;;;;;;;eAOQ,CAAC;AAChB,CAAC;AAED,iEAAiE;AAEjE,MAAM,MAAM,GAAG,CAAC,CAAU,EAAa,EAAE,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;AACtE,MAAM,GAAG,GAAG,CAAC,CAAU,EAAU,EAAE,CAAC,CAAC,OAAO,CAAC,KAAK,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;AACrE,MAAM,OAAO,GAAG,CAAC,CAAU,EAAY,EAAE,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC;AAE7E,SAAS,IAAI,CAAC,IAAY,EAAE,WAAmB,EAAE,MAAkB;IACjE,OAAO,EAAE,IAAI,EAAE,WAAW,EAAE,WAAW,EAAE,MAAM,EAAE,CAAC;AACpD,CAAC;AAGD,MAAM,CAAC,KAAK,UAAU,eAAe,CAAC,QAAyB,EAAE,OAA2B,EAAE,MAAoB;IAChH,MAAM,GAAG,GAAG,MAAM,QAAQ,CAAC,YAAY,CAAC,sBAAsB,EAAE;QAC9D,MAAM,EAAE,aAAa;QACrB,IAAI,EAAE,WAAW,CAAC,OAAO,CAAC;QAC1B,IAAI,EAAE,IAAI,CAAC,eAAe,EAAE,UAAU,EAAE,qBAAqB,CAAC;QAC9D,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KAC9B,CAAC,CAAC;IACH,MAAM,KAAK,GAAG,GAAG,CAAC,KAAgC,CAAC;IACnD,OAAO,MAAM,CAAC,KAAK,EAAE,CAAC,QAAQ,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE;QACzC,MAAM,CAAC,GAAG,CAA4B,CAAC;QACvC,OAAO,EAAE,SAAS,EAAE,GAAG,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,EAAE,OAAO,EAAE,OAAO,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,EAAE,MAAM,EAAE,GAAG,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,EAAE,CAAC;IACtG,CAAC,CAAC,CAAC;AACL,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,eAAe,CAAC,QAAyB,EAAE,OAA2B,EAAE,MAAoB;IAChH,MAAM,GAAG,GAAG,MAAM,QAAQ,CAAC,YAAY,CAAC,sBAAsB,EAAE;QAC9D,MAAM,EAAE,aAAa;QACrB,IAAI,EAAE,YAAY,CAAC,OAAO,CAAC;QAC3B,IAAI,EAAE,IAAI,CAAC,iBAAiB,EAAE,UAAU,EAAE,sBAAsB,CAAC;QACjE,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KAC9B,CAAC,CAAC;IACH,MAAM,KAAK,GAAG,GAAG,CAAC,KAAgC,CAAC;IACnD,OAAO,MAAM,CAAC,KAAK,EAAE,CAAC,UAAU,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,OAAO,CAAE,CAA6B,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC;AACpG,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,cAAc,CAAC,QAAyB,EAAE,OAAoD,EAAE,MAAoB;IACxI,MAAM,GAAG,GAAG,MAAM,QAAQ,CAAC,YAAY,CAAC,qBAAqB,EAAE;QAC7D,MAAM,EAAE,YAAY;QACpB,IAAI,EAAE,WAAW,CAAC,OAAO,CAAC;QAC1B,IAAI,EAAE,IAAI,CAAC,kBAAkB,EAAE,OAAO,EAAE,YAAY,CAAC;QACrD,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KAC9B,CAAC,CAAC;IACH,OAAO,GAAG,CAAE,GAAG,CAAC,KAAiC,EAAE,CAAC,WAAW,CAAC,CAAC,CAAC;AACpE,CAAC;AAGD,MAAM,CAAC,KAAK,UAAU,gBAAgB,CAAC,QAAyB,EAAE,OAAwF,EAAE,MAAoB;IAC9K,MAAM,GAAG,GAAG,MAAM,QAAQ,CAAC,YAAY,CAAC,uBAAuB,EAAE;QAC/D,MAAM,EAAE,cAAc;QACtB,IAAI,EAAE,aAAa,CAAC,OAAO,CAAC;QAC5B,IAAI,EAAE,IAAI,CAAC,eAAe,EAAE,aAAa,EAAE,cAAc,CAAC;QAC1D,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KAC9B,CAAC,CAAC;IACH,MAAM,KAAK,GAAG,GAAG,CAAC,KAAgC,CAAC;IACnD,OAAO,MAAM,CAAC,KAAK,EAAE,CAAC,QAAQ,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE;QACzC,MAAM,CAAC,GAAG,CAA4B,CAAC;QACvC,OAAO,EAAE,IAAI,EAAE,GAAG,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,EAAE,KAAK,EAAE,OAAO,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,EAAE,MAAM,EAAE,GAAG,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,EAAE,CAAC;IACxF,CAAC,CAAC,CAAC;AACL,CAAC"}

package/dist/llm/tasks/ingest/state-merge.d.ts

@@ -0,0 +1,10 @@
+import type { LlmTaskExecutor } from "../../types.js";
+export declare const STATE_MERGE_SYSTEM: string;
+export declare function stateMergePrompt(personName: string, states: Array<{
+    name: string;
+    count: number;
+}>): string;
+export declare function runStateMerge(executor: LlmTaskExecutor, person: string, states: Array<{
+    name: string;
+    count: number;
+}>, signal?: AbortSignal): Promise<Record<string, string>>;

package/dist/llm/tasks/ingest/state-merge.js

@@ -0,0 +1,84 @@
+// Video state merge (§10.2 pass3c, ported from cezhi-verify 03c). Per character,
+// collapse the many free-form outfit wordings (Gemini describes the same outfit
+// differently each episode) into canonical states. Distinct from govern's
+// state-reconcile: this dedups the wording fragmentation before convergence;
+// govern later selects default + drops transient. Prompt is VERBATIM from the
+// tuned scaffold; the tool-input schema carries its field descriptions verbatim.
+export const STATE_MERGE_SYSTEM = "你是剧集人物『造型状态』的归并器。同一个角色在不同镜头/不同集里穿的**同一套衣服**，" +
+    "常被逐集自由描述成很多不同措辞——有的写全（连颜色、款式、编号都写上），有的只写个大概（只有『白衣』『球服』这种笼统类别），" +
+    "还有的把颜色/款式说岔（把某个颜色写成邻近色、把一种服装误认成另一种）。" +
+    "你的任务：只看**一个人物**的全部状态措辞，把其中**其实是同一套装扮**的合并成一组，每组选一个规范状态名。" +
+    "核心判据：**是不是同一套服装造型**（同一件/同一套衣服），而不是颜色字面像不像。" +
+    "原则：" +
+    "① 同一套的不同繁简说法要合：同一套衣服，有的措辞带全颜色+款式+编号、有的只写大类——只要都指这人那一套，就合成一组，canonical 取信息最完整的。" +
+    "② 明显的口误/误认要并进对应造型：把某类服装误写成相近的另一类、把某颜色写成邻近色调，只要指的是同一套，就并进去。" +
+    "③ **不同造型绝不合**：不同颜色/不同主色的两套（如一套浅色、一套深色）、正装 vs 便装、正常 vs 受伤/特殊状态，是不同造型，各自成组。" +
+    "④ 拿不准是不是同一套时，宁可**不合**、让它单独成组——错合（把两套并成一套）比漏合更糟。" +
+    "⑤ canonical 必须原样是组内出现过的措辞，不要合成新词、不要改字。" +
+    "⑥ 覆盖每一个输入措辞：每个措辞必须恰好属于一个组（哪怕它独自成组）。";
+const STATE_MERGE_SCHEMA = {
+    type: "object",
+    additionalProperties: false,
+    required: ["groups"],
+    properties: {
+        groups: {
+            type: "array",
+            description: "把输入的状态措辞分组。每组=同一套造型的不同说法。只指同一套装扮的才归一组；不同颜色/不同款式/不同场合的造型各自单独成组。",
+            items: {
+                type: "object",
+                additionalProperties: false,
+                required: ["canonical", "members", "reason"],
+                properties: {
+                    canonical: { type: "string", description: "这一组的规范状态名。从组内措辞里挑**信息最完整、最清晰**的那个：同时含颜色+款式+编号/标识细节的，优于只有笼统类别的（信息全的一版优于「白衣」「运动服」这种大概说法）。必须原样是组内出现过的某个措辞，不要新造。" },
+                    members: { type: "array", items: { type: "string" }, description: "确认是**同一套服装造型**的所有措辞（含 canonical 自身）。只放真正指同一套的。" },
+                    reason: { type: "string", description: "为什么这些是同一套造型，一句话（如：都是同一套主色球衣，只是措辞繁简不同）。" },
+                },
+            },
+        },
+    },
+};
+export function stateMergePrompt(personName, states) {
+    const lines = states.map((s) => `- 「${s.name}」（出现 ${s.count} 场）`).join("\n");
+    return `下面是角色「${personName}」在全剧里被记录到的所有造型状态措辞。请把其中**其实是同一套装扮**的合并分组：
+
+${lines}
+
+要求：
+- 判据是「是不是同一套服装造型」，不是颜色字面像不像。
+- 同一套衣服的繁简/口误/误认说法合成一组，canonical 取信息最完整清晰的那个措辞（原样照抄，别新造）。
+- 不同颜色/款式/场合的造型（如一套浅色装 vs 一套深色装 vs 正装 vs 受伤/特殊状态）各自单独成组。
+- 拿不准就让它单独成组，别硬合。
+- 覆盖每一个措辞：每个措辞恰好属于一个组。
+严格按 schema 返回。`;
+}
+// Returns { wording -> canonical } for the merged (≥2-member) groups only,
+// mirroring the scaffold mergePerson: canonical from LLM if valid, else the
+// highest-count wording; hallucinated members dropped.
+export async function runStateMerge(executor, person, states, signal) {
+    const valid = new Set(states.map((s) => s.name));
+    const cnt = new Map(states.map((s) => [s.name, s.count]));
+    const res = await executor.generateTool("ingest.state-merge", {
+        system: STATE_MERGE_SYSTEM,
+        user: stateMergePrompt(person, states),
+        tool: { name: "submit_state_groups", description: "提交造型状态归并分组", inputSchema: STATE_MERGE_SCHEMA },
+        ...(signal ? { signal } : {}),
+    });
+    const groups = Array.isArray(res.input?.["groups"]) ? res.input["groups"] : [];
+    const map = {};
+    for (const g of groups) {
+        const o = g;
+        const membersRaw = Array.isArray(o["members"]) ? o["members"] : [];
+        const mem = membersRaw.map((m) => (typeof m === "string" ? m : "")).filter((m, i, a) => valid.has(m) && a.indexOf(m) === i);
+        if (mem.length < 2)
+            continue;
+        const canonRaw = typeof o["canonical"] === "string" ? o["canonical"] : "";
+        let canon = valid.has(canonRaw) && mem.includes(canonRaw) ? canonRaw : "";
+        if (!canon)
+            canon = mem.reduce((b, s) => ((cnt.get(s) ?? 0) > (cnt.get(b) ?? 0) ? s : b), mem[0]);
+        for (const s of mem)
+            if (s !== canon)
+                map[s] = canon;
+    }
+    return map;
+}
+//# sourceMappingURL=state-merge.js.map

package/dist/llm/tasks/ingest/state-merge.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"state-merge.js","sourceRoot":"","sources":["../../../../src/llm/tasks/ingest/state-merge.ts"],"names":[],"mappings":"AAAA,iFAAiF;AACjF,gFAAgF;AAChF,0EAA0E;AAC1E,6EAA6E;AAC7E,8EAA8E;AAC9E,iFAAiF;AAIjF,MAAM,CAAC,MAAM,kBAAkB,GAC7B,8CAA8C;IAC9C,gEAAgE;IAChE,sCAAsC;IACtC,0DAA0D;IAC1D,4CAA4C;IAC5C,KAAK;IACL,gFAAgF;IAChF,4DAA4D;IAC5D,2EAA2E;IAC3E,iDAAiD;IACjD,wCAAwC;IACxC,qCAAqC,CAAC;AAExC,MAAM,kBAAkB,GAAe;IACrC,IAAI,EAAE,QAAQ;IACd,oBAAoB,EAAE,KAAK;IAC3B,QAAQ,EAAE,CAAC,QAAQ,CAAC;IACpB,UAAU,EAAE;QACV,MAAM,EAAE;YACN,IAAI,EAAE,OAAO;YACb,WAAW,EAAE,gEAAgE;YAC7E,KAAK,EAAE;gBACL,IAAI,EAAE,QAAQ;gBACd,oBAAoB,EAAE,KAAK;gBAC3B,QAAQ,EAAE,CAAC,WAAW,EAAE,SAAS,EAAE,QAAQ,CAAC;gBAC5C,UAAU,EAAE;oBACV,SAAS,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE,WAAW,EAAE,8GAA8G,EAAE;oBAC1J,OAAO,EAAE,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE,EAAE,WAAW,EAAE,gDAAgD,EAAE;oBACpH,MAAM,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE,WAAW,EAAE,wCAAwC,EAAE;iBAClF;aACF;SACF;KACF;CACF,CAAC;AAEF,MAAM,UAAU,gBAAgB,CAAC,UAAkB,EAAE,MAA8C;IACjG,MAAM,KAAK,GAAG,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,MAAM,CAAC,CAAC,IAAI,QAAQ,CAAC,CAAC,KAAK,KAAK,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IAC7E,OAAO,SAAS,UAAU;;EAE1B,KAAK;;;;;;;;eAQQ,CAAC;AAChB,CAAC;AAED,2EAA2E;AAC3E,4EAA4E;AAC5E,uDAAuD;AACvD,MAAM,CAAC,KAAK,UAAU,aAAa,CACjC,QAAyB,EACzB,MAAc,EACd,MAA8C,EAC9C,MAAoB;IAEpB,MAAM,KAAK,GAAG,IAAI,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC;IACjD,MAAM,GAAG,GAAG,IAAI,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;IAC1D,MAAM,GAAG,GAAG,MAAM,QAAQ,CAAC,YAAY,CAAC,oBAAoB,EAAE;QAC5D,MAAM,EAAE,kBAAkB;QAC1B,IAAI,EAAE,gBAAgB,CAAC,MAAM,EAAE,MAAM,CAAC;QACtC,IAAI,EAAE,EAAE,IAAI,EAAE,qBAAqB,EAAE,WAAW,EAAE,YAAY,EAAE,WAAW,EAAE,kBAAkB,EAAE;QACjG,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KAC9B,CAAC,CAAC;IACH,MAAM,MAAM,GAAG,KAAK,CAAC,OAAO,CAAE,GAAG,CAAC,KAAiC,EAAE,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAG,GAAG,CAAC,KAAiC,CAAC,QAAQ,CAAe,CAAC,CAAC,CAAC,EAAE,CAAC;IACxJ,MAAM,GAAG,GAA2B,EAAE,CAAC;IACvC,KAAK,MAAM,CAAC,IAAI,MAAM,EAAE,CAAC;QACvB,MAAM,CAAC,GAAG,CAA4B,CAAC;QACvC,MAAM,UAAU,GAAG,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,CAAE,CAAC,CAAC,SAAS,CAAe,CAAC,CAAC,CAAC,EAAE,CAAC;QAClF,MAAM,GAAG,GAAG,UAAU,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,OAAO,CAAC,KAAK,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC;QAC5H,IAAI,GAAG,CAAC,MAAM,GAAG,CAAC;YAAE,SAAS;QAC7B,MAAM,QAAQ,GAAG,OAAO,CAAC,CAAC,WAAW,CAAC,KAAK,QAAQ,CAAC,CAAC,CAAE,CAAC,CAAC,WAAW,CAAY,CAAC,CAAC,CAAC,EAAE,CAAC;QACtF,IAAI,KAAK,GAAG,KAAK,CAAC,GAAG,CAAC,QAAQ,CAAC,IAAI,GAAG,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,EAAE,CAAC;QAC1E,IAAI,CAAC,KAAK;YAAE,KAAK,GAAG,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC,CAAE,CAAC,CAAC;QACnG,KAAK,MAAM,CAAC,IAAI,GAAG;YAAE,IAAI,CAAC,KAAK,KAAK;gBAAE,GAAG,CAAC,CAAC,CAAC,GAAG,KAAK,CAAC;IACvD,CAAC;IACD,OAAO,GAAG,CAAC;AACb,CAAC"}

package/dist/llm/tasks/ingest/state-reconcile.d.ts

@@ -0,0 +1,8 @@
+import type { StateGroup } from "../../../domain/ingest/state-reconcile.js";
+import type { LlmTaskExecutor } from "../../types.js";
+export declare const STATE_RECONCILE_SYSTEM: string;
+export declare function parseStateGroups(raw: string, validIds: Set<string>): StateGroup[];
+export declare function runStateReconcile(executor: LlmTaskExecutor, states: Array<{
+    id: string;
+    name: string;
+}>, signal?: AbortSignal): Promise<StateGroup[]>;

package/dist/llm/tasks/ingest/state-reconcile.js

@@ -0,0 +1,44 @@
+// ingest.state-reconcile task (§9.8). Per canonical owner, merge state names
+// that describe the same stable visual appearance. No roster recall — the whole
+// owner's states go to the model at once. Emotion/pose/transient states should
+// already be absent; the model only groups equivalent durable looks.
+import { generateOptional } from "./asset-govern.js";
+export const STATE_RECONCILE_SYSTEM = [
+    "你在给某个角色/地点/道具**筛选并归并稳定外观造型**。给你它名下全部候选状态（每行 <id> | <状态名>）。",
+    "目标是得到一小组**互不相同、值得单独出图的外观档位**。",
+    "只保留可视**外观造型**（穿什么、发型妆造、形态）：",
+    "- 保留：整套着装造型（婚服/喜服/常服/黑衣/睡衣/粗布衣裳）、发型、持续性妆造、非人形态/兽化/神化、巨大化、**持续性伤势造型**（满身血污、胸口缠绷带这种跨镜头持续的受伤外观）。",
+    "- 删除（绝不列出）：情绪与神态（悲痛/冰冷/狂喜/惨白/濒死/身姿挺拔）、姿势与动作（坐轮椅/站立/苏醒/昏迷/睡着/被绑/转身/跪地）、位置情境（门外/在牢内/台上）、镜头、瞬时或过程描述（眼眶发红/满嘴油光/嘴角沾米粒/从X变Y）、纯剧情态。",
+    "**大胆合并**：描述同一套外观的不同写法要并成一组（各种“受伤/血污/绷带/刀伤”并成一个受伤造型；同一套衣服的不同细节写法并一组）。只有**真正不同的造型**才分开。宁可少而清，别把近义拆成一堆。",
+    "canonical 名只能从输入的状态名里选，挑最能代表该造型的那个。",
+    "输出：每个保留的外观组一行 `<canonical名>: id1, id2, ...`（单个也要一行）；**没列出的 id 一律删除**。",
+    "如果没有任何值得保留的稳定外观，就输出一行 `无`。不要输出别的文字。",
+].join("\n");
+// Each output line is a KEPT durable-state group (canonical + ≥1 member id).
+// Ids not present in any group are dropped by the caller.
+export function parseStateGroups(raw, validIds) {
+    const groups = [];
+    for (const line of raw.split("\n")) {
+        const t = line.trim();
+        if (t === "" || t === "无" || t.startsWith("#"))
+            continue;
+        const m = /^(.+?)\s*[:：]\s*(.+)$/.exec(t);
+        if (!m)
+            continue;
+        const ids = [...new Set(m[2].split(/[,，、]/).map((s) => s.trim()).filter((s) => validIds.has(s)))];
+        if (ids.length >= 1)
+            groups.push({ canonical: m[1].trim(), state_ids: ids });
+    }
+    return groups;
+}
+export async function runStateReconcile(executor, states, signal) {
+    if (states.length < 1)
+        return [];
+    const text = await generateOptional(executor, "ingest.state-reconcile", {
+        system: STATE_RECONCILE_SYSTEM,
+        user: states.map((s) => `${s.id} | ${s.name}`).join("\n"),
+        ...(signal ? { signal } : {}),
+    });
+    return parseStateGroups(text, new Set(states.map((s) => s.id)));
+}
+//# sourceMappingURL=state-reconcile.js.map

package/dist/llm/tasks/ingest/state-reconcile.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"state-reconcile.js","sourceRoot":"","sources":["../../../../src/llm/tasks/ingest/state-reconcile.ts"],"names":[],"mappings":"AAAA,6EAA6E;AAC7E,gFAAgF;AAChF,+EAA+E;AAC/E,qEAAqE;AAIrE,OAAO,EAAE,gBAAgB,EAAE,MAAM,mBAAmB,CAAC;AAErD,MAAM,CAAC,MAAM,sBAAsB,GAAG;IACpC,4DAA4D;IAC5D,+BAA+B;IAC/B,6BAA6B;IAC7B,+FAA+F;IAC/F,8HAA8H;IAC9H,sGAAsG;IACtG,qCAAqC;IACrC,uEAAuE;IACvE,qCAAqC;CACtC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAEb,6EAA6E;AAC7E,0DAA0D;AAC1D,MAAM,UAAU,gBAAgB,CAAC,GAAW,EAAE,QAAqB;IACjE,MAAM,MAAM,GAAiB,EAAE,CAAC;IAChC,KAAK,MAAM,IAAI,IAAI,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC;QACnC,MAAM,CAAC,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC;QACtB,IAAI,CAAC,KAAK,EAAE,IAAI,CAAC,KAAK,GAAG,IAAI,CAAC,CAAC,UAAU,CAAC,GAAG,CAAC;YAAE,SAAS;QACzD,MAAM,CAAC,GAAG,uBAAuB,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;QAC1C,IAAI,CAAC,CAAC;YAAE,SAAS;QACjB,MAAM,GAAG,GAAG,CAAC,GAAG,IAAI,GAAG,CAAC,CAAC,CAAC,CAAC,CAAE,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QACnG,IAAI,GAAG,CAAC,MAAM,IAAI,CAAC;YAAE,MAAM,CAAC,IAAI,CAAC,EAAE,SAAS,EAAE,CAAC,CAAC,CAAC,CAAE,CAAC,IAAI,EAAE,EAAE,SAAS,EAAE,GAAG,EAAE,CAAC,CAAC;IAChF,CAAC;IACD,OAAO,MAAM,CAAC;AAChB,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,iBAAiB,CAAC,QAAyB,EAAE,MAA2C,EAAE,MAAoB;IAClI,IAAI,MAAM,CAAC,MAAM,GAAG,CAAC;QAAE,OAAO,EAAE,CAAC;IACjC,MAAM,IAAI,GAAG,MAAM,gBAAgB,CAAC,QAAQ,EAAE,wBAAwB,EAAE;QACtE,MAAM,EAAE,sBAAsB;QAC9B,IAAI,EAAE,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,GAAG,CAAC,CAAC,EAAE,MAAM,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC;QACzD,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KAC9B,CAAC,CAAC;IACH,OAAO,gBAAgB,CAAC,IAAI,EAAE,IAAI,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;AAClE,CAAC"}

package/dist/llm/tasks/ingest/text-normalize.d.ts

@@ -0,0 +1,10 @@
+import type { Chunk } from "../../../domain/ingest/chunker.js";
+import type { IngestRecord } from "../../../domain/ingest/records.js";
+import type { LlmTaskExecutor } from "../../types.js";
+export declare const TEXT_NORMALIZE_SYSTEM: string;
+export declare function buildNormalizeUser(chunk: Chunk): string;
+export declare function parseNormalizeOutput(markdown: string, chunk: Chunk): IngestRecord[];
+export declare function runTextNormalize(executor: LlmTaskExecutor, chunk: Chunk, signal?: AbortSignal): Promise<{
+    markdown: string;
+    records: IngestRecord[];
+}>;

package/dist/llm/tasks/ingest/text-normalize.js

@@ -0,0 +1,89 @@
+// ingest.text-normalize task (§3.3/§pass2). The cheap model rewrites one source
+// chunk as a lean markdown document; a deterministic parser (markdown-parser.ts)
+// turns that into records (§7) with inline entity tags (§7.6). Presence/coverage
+// is chunk-granular: a non-empty chunk must yield ≥1 record. No repair prompt and
+// no downgrade-to-raw on failure (§12.3).
+import { CliError, EXIT_RUNTIME } from "../../../common.js";
+import { parseNormalizeMarkdown, MarkdownParseError } from "../../../domain/ingest/markdown-parser.js";
+export const TEXT_NORMALIZE_SYSTEM = [
+    "你是剧本规范化器。把给定的源文本块改写成一份精简 markdown，严格按源文本发生顺序。",
+    "只输出该 markdown 文档：不要代码围栏，前后不要任何多余文字或解释。",
+    "",
+    "严格保持源文本顺序，绝不重排：块开头若在任何『集/场』标记之前就有正文（上一块没写完的戏），",
+    "必须把这些正文**原位放在输出最前面**（不加任何头），`# 第N集`/`## 场景` 只在它在源文本里真正出现的位置输出。",
+    "绝不要把靠后出现的 `# 第N集` 提到开头。",
+    "",
+    "结构（只有源文本里真的出现对应边界时才写，绝不臆造；块可能从一场戏中间开始，那就直接从正文写起）：",
+    "# 第N集 标题        （出现集边界时；标题可省）",
+    "## 场景            （出现新场景/新地点/新时空时）",
+    "- 时空: 内 日       （紧跟场景头，可省；内|外 日|夜）",
+    "- 地点: 地点名       （紧跟场景头，可省；地点是场级）",
+    "- 人物: 甲【状态】, 乙, 丙   （本场在场的具体单人角色；【】内写整套外观状态，可省）",
+    "- 道具: 关键道具1, 道具2    （本场需要跨镜头连续性的关键道具；通用/一次性物件不列）",
+    "",
+    "正文每行一条，按发生顺序。动作/环境描写就是**裸行**（不加任何前缀）；只有对白/内心独白/转场/批注需要标记：",
+    "动作/环境：直接写一行纯文本，不加前缀（这是默认）",
+    "[dlg|说话人] 对白        [dlg|说话人|情绪] 带情绪的对白",
+    "[think|说话人] 内心独白/画外音（含 OS/旁白）",
+    "[trans] 转场",
+    "[note|author] 作者批注   [note|format] 排版说明",
+    "",
+    "铁律——逐行转换，绝不原样保留源标记。输入往往已是半结构化剧本，把每一行都转成上面的结构：",
+    "  源 `场3-2 日/内 客厅`            → `## 场景` + `- 时空: 内 日` + `- 地点: 客厅`",
+    "  源 `人物：甲【白衣,昏迷】，乙`      → `- 人物: 甲【白衣,昏迷】, 乙`",
+    "  源 `道具：婚书，玉佩`             → `- 道具: 婚书, 玉佩`",
+    "  源 `△ 甲推门而入。`              → `甲推门而入。`（裸行，动作不加前缀）",
+    "  源 `甲：你来了`                  → `[dlg|甲] 你来了`",
+    "  源 `甲（冷笑）：你来了`           → `[dlg|甲|冷笑] 你来了`",
+    "  源 `甲OS：（恐慌）他没死吧？`       → `[think|甲|恐慌] 他没死吧？`",
+    "  源 `【字幕：回门日】`             → `字幕：回门日`（画面文字是动作，裸行）",
+    "出现残留的 △ 前缀、裸『名：台词』（该转成 [dlg]）就是错的。",
+    "",
+    "正文口径：**纯文本、逐字保留**，不加任何标注，不润色、不扩写、不纠错、不加戏。人物/道具的在场只写在场景头的 `- 人物`/`- 道具`，不要写进正文。",
+    "",
+    "- 人物 只列**具体单人角色**；家族/机构/群体/泛称（顾家、大房、林家、众人、随从们、丫鬟们…）绝不列入 - 人物（群体喊话用 group 说话人或写进 act 旁白）。",
+    "- 说话人：有名字/有台词/露脸的具体人物一律用角色名作 actor 说话人（如 [dlg|传旨太监]），全剧同一人保持一致；kind 枚举 system|broadcast|offscreen|group 只给真正无实体的声音（画外广播、无法归到具体人的群体喊话）。同一个人不要一会儿 system 一会儿 actor。",
+    "- 画面上出现的文字（字幕/标题卡/人物介绍条/时间地点卡）是画面内容，用 > [act] 记录并把文字原样写进正文；与某句对白完全重复的冗余字幕并入对应 [dlg]。",
+    "- note 只用于真正无法归入剧本、也不属于画面/对白的内容；能归入 act/dlg 的绝不用 note。",
+    "源文本块的实质内容不得丢失。",
+].join("\n");
+export function buildNormalizeUser(chunk) {
+    // Do NOT expose the chunk index — the model conflates a chunk number with an
+    // episode number (`#15` → `# 第15集`). Chunk source spans are assigned by the
+    // parser, not the model.
+    return ["【待规范化的原文片段（可能从一场戏中间开始）】", "", chunk.text].join("\n");
+}
+function runtimeError(message, received) {
+    return new CliError("INGEST NORMALIZE FAILED", message, {
+        exitCode: EXIT_RUNTIME,
+        errorCode: "INGEST_NORMALIZE_FAILED",
+        received,
+        nextSteps: ["Inspect the chunk error sidecar and rerun `scriptctl ingest`."],
+    });
+}
+// Parse the model's markdown for one chunk and enforce presence: a chunk whose
+// source text has real content must produce at least one record.
+export function parseNormalizeOutput(markdown, chunk) {
+    let records;
+    try {
+        records = parseNormalizeMarkdown(markdown, chunk);
+    }
+    catch (err) {
+        if (err instanceof MarkdownParseError)
+            throw runtimeError(`chunk ${chunk.index}: ${err.message}`, [err.message]);
+        throw err;
+    }
+    if (records.length === 0 && chunk.text.trim() !== "") {
+        throw runtimeError(`chunk ${chunk.index}: non-empty chunk produced no records`, ["records: 0"]);
+    }
+    return records;
+}
+export async function runTextNormalize(executor, chunk, signal) {
+    const response = await executor.generateText("ingest.text-normalize", {
+        system: TEXT_NORMALIZE_SYSTEM,
+        user: buildNormalizeUser(chunk),
+        ...(signal ? { signal } : {}),
+    });
+    return { markdown: response.text, records: parseNormalizeOutput(response.text, chunk) };
+}
+//# sourceMappingURL=text-normalize.js.map

package/dist/llm/tasks/ingest/text-normalize.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"text-normalize.js","sourceRoot":"","sources":["../../../../src/llm/tasks/ingest/text-normalize.ts"],"names":[],"mappings":"AAAA,gFAAgF;AAChF,iFAAiF;AACjF,iFAAiF;AACjF,kFAAkF;AAClF,0CAA0C;AAE1C,OAAO,EAAE,QAAQ,EAAE,YAAY,EAAE,MAAM,oBAAoB,CAAC;AAE5D,OAAO,EAAE,sBAAsB,EAAE,kBAAkB,EAAE,MAAM,2CAA2C,CAAC;AAIvG,MAAM,CAAC,MAAM,qBAAqB,GAAG;IACnC,+CAA+C;IAC/C,wCAAwC;IACxC,EAAE;IACF,gDAAgD;IAChD,gEAAgE;IAChE,yBAAyB;IACzB,EAAE;IACF,mDAAmD;IACnD,+BAA+B;IAC/B,mCAAmC;IACnC,oCAAoC;IACpC,kCAAkC;IAClC,iDAAiD;IACjD,kDAAkD;IAClD,EAAE;IACF,2DAA2D;IAC3D,2BAA2B;IAC3B,yCAAyC;IACzC,+BAA+B;IAC/B,YAAY;IACZ,yCAAyC;IACzC,EAAE;IACF,+CAA+C;IAC/C,mEAAmE;IACnE,gDAAgD;IAChD,6CAA6C;IAC7C,mDAAmD;IACnD,8CAA8C;IAC9C,8CAA8C;IAC9C,kDAAkD;IAClD,mDAAmD;IACnD,oCAAoC;IACpC,EAAE;IACF,gFAAgF;IAChF,EAAE;IACF,0FAA0F;IAC1F,mKAAmK;IACnK,sFAAsF;IACtF,wDAAwD;IACxD,gBAAgB;CACjB,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAEb,MAAM,UAAU,kBAAkB,CAAC,KAAY;IAC7C,6EAA6E;IAC7E,4EAA4E;IAC5E,yBAAyB;IACzB,OAAO,CAAC,yBAAyB,EAAE,EAAE,EAAE,KAAK,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAChE,CAAC;AAED,SAAS,YAAY,CAAC,OAAe,EAAE,QAAkB;IACvD,OAAO,IAAI,QAAQ,CAAC,yBAAyB,EAAE,OAAO,EAAE;QACtD,QAAQ,EAAE,YAAY;QACtB,SAAS,EAAE,yBAAyB;QACpC,QAAQ;QACR,SAAS,EAAE,CAAC,+DAA+D,CAAC;KAC7E,CAAC,CAAC;AACL,CAAC;AAED,+EAA+E;AAC/E,iEAAiE;AACjE,MAAM,UAAU,oBAAoB,CAAC,QAAgB,EAAE,KAAY;IACjE,IAAI,OAAuB,CAAC;IAC5B,IAAI,CAAC;QACH,OAAO,GAAG,sBAAsB,CAAC,QAAQ,EAAE,KAAK,CAAC,CAAC;IACpD,CAAC;IAAC,OAAO,GAAG,EAAE,CAAC;QACb,IAAI,GAAG,YAAY,kBAAkB;YAAE,MAAM,YAAY,CAAC,SAAS,KAAK,CAAC,KAAK,KAAK,GAAG,CAAC,OAAO,EAAE,EAAE,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC;QACjH,MAAM,GAAG,CAAC;IACZ,CAAC;IACD,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC,IAAI,KAAK,CAAC,IAAI,CAAC,IAAI,EAAE,KAAK,EAAE,EAAE,CAAC;QACrD,MAAM,YAAY,CAAC,SAAS,KAAK,CAAC,KAAK,uCAAuC,EAAE,CAAC,YAAY,CAAC,CAAC,CAAC;IAClG,CAAC;IACD,OAAO,OAAO,CAAC;AACjB,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,gBAAgB,CAAC,QAAyB,EAAE,KAAY,EAAE,MAAoB;IAClG,MAAM,QAAQ,GAAG,MAAM,QAAQ,CAAC,YAAY,CAAC,uBAAuB,EAAE;QACpE,MAAM,EAAE,qBAAqB;QAC7B,IAAI,EAAE,kBAAkB,CAAC,KAAK,CAAC;QAC/B,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KAC9B,CAAC,CAAC;IACH,OAAO,EAAE,QAAQ,EAAE,QAAQ,CAAC,IAAI,EAAE,OAAO,EAAE,oBAAoB,CAAC,QAAQ,CAAC,IAAI,EAAE,KAAK,CAAC,EAAE,CAAC;AAC1F,CAAC"}

package/dist/llm/tasks/ingest/video-correct.d.ts

@@ -0,0 +1,30 @@
+import type { TaskVideoPort, VideoFileHandle } from "../../types.js";
+export declare const VIDEO_CORRECT_SYSTEM: string;
+export declare function videoCorrectPrompt(episodeNo: number, contextMd: string, pass1Md: string): string;
+export interface VideoCorrection {
+    name_fixes: Array<{
+        from: string;
+        to: string;
+        scope: string;
+        reason: string;
+    }>;
+    same_person: Array<{
+        names: string[];
+        reason: string;
+    }>;
+    speaker_fixes: Array<{
+        ts: string;
+        from: string;
+        to: string;
+        reason: string;
+    }>;
+    dialogue_fixes: Array<{
+        ts: string;
+        speaker: string;
+        wrong: string;
+        right: string;
+        reason: string;
+    }>;
+    notes: string[];
+}
+export declare function runVideoCorrect(port: TaskVideoPort, file: VideoFileHandle, episodeNo: number, contextMd: string, pass1Md: string, fps: number, signal?: AbortSignal): Promise<VideoCorrection>;