Complete durable canvas groups and add isolated P40 video upscaling

.gitignore

@@ -112,3 +112,15 @@ public/vendor/zoo-modeling-app/
 .cache/smart-edit/
 
 services/video-render/dist/
+
+# Downloaded and normalized subtitle fonts are runtime assets.
+.cache/smart-edit-fonts/
+services/smart-edit/__pycache__/
+
+# Local verification artefacts (also excluded from Tailwind's automatic source scan).
+.qa/
+.next-video-*/
+
+services/video-upscale/dist/
+services/video-upscale/__pycache__/
+services/video-upscale/vendor/__pycache__/

app/api/canvas/[id]/route.ts

@@ -163,6 +163,8 @@ export async function PUT(
           .values({
             id,
             userId: session.user.id,
+            createdAt: new Date(),
+            updatedAt: new Date(),
             title: title || getDefaultWorkspaceTitle(workspaceType || "canvas"),
             workspaceType: workspaceType || "canvas",
             canvasData: {

app/api/canvas/generation-tasks/route.ts

@@ -1,3 +1,4 @@
+import { videoUpscaleSchema } from "@/lib/video-canvas/video-upscale-contract";
 import { smartEditRequestSchema } from "@/lib/video-canvas/smart-edit/contract";
 import { audioRequestSchema } from "@/lib/video-canvas/audio-contract";
 import { canResumeCanvasTask } from "@/lib/video-canvas/task-recovery";
@@ -43,6 +44,7 @@ export async function POST(request: Request) {
       throw Object.assign(Error("任务数据过大，请使用已上传的素材链接"), { status: 413 });
     const parsed = canvasTaskRequestSchema.safeParse(JSON.parse(text));
     if (!parsed.success) throw Object.assign(Error("任务参数不合法"), { status: 400 });
+    if(parsed.data.workflow)throw Object.assign(Error("编组任务必须使用整组提交接口"),{status:400});
     if (
       parsed.data.kind === "image-operation" &&
       !imageOperationSchema.safeParse(parsed.data.params).success
@@ -61,6 +63,7 @@ export async function POST(request: Request) {
       !smartEditRequestSchema.safeParse(parsed.data.params).success
     )
       throw Object.assign(Error("剪辑参数无效"), { status: 400 });
+    if (parsed.data.kind === "video-upscale" && !videoUpscaleSchema.safeParse(parsed.data.params).success) throw Object.assign(Error("视频高清参数无效"), {status:400});
     const userId = await owner(parsed.data.projectId);
     const task = await createCanvasGenerationTask(userId, parsed.data);
     kick(userId, parsed.data.projectId);

app/api/canvas/group-runs/route.ts

@@ -0,0 +1,39 @@
+import { auth } from "@/app/(auth)/auth";
+import { after, NextResponse } from "next/server";
+import { canvasTaskRequestSchema } from "@/lib/video-canvas/task-contract";
+import {
+  createCanvasGroupTasks,
+  ownsTaskProject,
+} from "@/lib/video-canvas/task-store";
+import { runCanvasGenerationQueue } from "@/lib/video-canvas/task-worker";
+export const runtime = "nodejs",
+  maxDuration = 900;
+export async function POST(request: Request) {
+  try {
+    const session = await auth();
+    if (!session?.user?.id)
+      return NextResponse.json({ error: "未登录" }, { status: 401 });
+    const text = await request.text();
+    if (Buffer.byteLength(text) > 1048576)
+      return NextResponse.json({ error: "编组数据过大" }, { status: 413 });
+    const parsed = canvasTaskRequestSchema
+      .array()
+      .min(1)
+      .max(40)
+      .safeParse(JSON.parse(text));
+    if (!parsed.success)
+      return NextResponse.json({ error: "编组参数无效" }, { status: 400 });
+    const projectId = parsed.data[0].projectId,
+      userId = session.user.id;
+    if (!(await ownsTaskProject(userId, projectId)))
+      return NextResponse.json({ error: "画布不存在" }, { status: 404 });
+    const tasks = await createCanvasGroupTasks(userId, parsed.data);
+    after(() => runCanvasGenerationQueue(userId, projectId));
+    return NextResponse.json({ success: true, tasks }, { status: 202 });
+  } catch (e) {
+    return NextResponse.json(
+      { error: e instanceof Error ? e.message : "编组执行失败" },
+      { status: Number((e as any)?.status) || 400 }
+    );
+  }
+}

app/api/canvas/group-templates/route.ts

@@ -0,0 +1,40 @@
+import { auth } from "@/app/(auth)/auth";
+import { db } from "@/lib/db";
+import { canvasProject } from "@/lib/db/schema";
+import { eq, and, sql, desc } from "drizzle-orm";
+import { NextResponse } from "next/server";
+export async function GET() {
+  const session = await auth();
+  if (!session?.user?.id)
+    return NextResponse.json({ error: "未登录" }, { status: 401 });
+  const rows = await db
+    .select({
+      projectId: canvasProject.id,
+      title: canvasProject.title,
+      templates: sql<any[]>`${canvasProject.canvasData}->'groupTemplates'`,
+    })
+    .from(canvasProject)
+    .where(
+      and(
+        eq(canvasProject.userId, session.user.id),
+        eq(canvasProject.workspaceType, "video"),
+        sql`${canvasProject.canvasData} ? 'groupTemplates'`
+      )
+    )
+    .orderBy(desc(canvasProject.updatedAt))
+    .limit(100);
+  return NextResponse.json(
+    {
+      templates: rows.flatMap((row) =>
+        Array.isArray(row.templates)
+          ? row.templates.map((template) => ({
+              ...template,
+              projectId: row.projectId,
+              projectTitle: row.title,
+            }))
+          : []
+      ),
+    },
+    { headers: { "Cache-Control": "no-store" } }
+  );
+}

app/api/canvas/route.ts

@@ -100,6 +100,8 @@ export async function POST(request: NextRequest) {
       .insert(canvasProject)
       .values({
         userId: session.user.id,
+        createdAt: new Date(),
+        updatedAt: new Date(),
         title: title || getDefaultWorkspaceTitle(workspaceType),
         description,
         workspaceType,

app/api/canvas/smart-edit/capabilities/route.ts

@@ -10,7 +10,14 @@ export async function GET() {
   const session = await auth();
   if (!session?.user?.id) return NextResponse.json({ error: "未登录" }, { status: 401 });
   try {
-    let capabilities = { scenes: false, beats: false, asr: false, ffmpeg: false, ffprobe: false };
+    let capabilities = {
+      separation: false,
+      scenes: false,
+      beats: false,
+      asr: false,
+      ffmpeg: false,
+      ffprobe: false,
+    };
     try {
       const { stdout } = await promisify(execFile)(
         process.env.SMART_EDIT_PYTHON || "python3",
@@ -35,7 +42,9 @@ export async function GET() {
       }
     }
     const models = (await getEnabledAiModelsWithProvider(false))
-      .filter((item) => isCanvasVisionModel({ ...item.model, providerType: item.provider.type }))
+      .filter((item) =>
+        isCanvasVisionModel({ ...item.model, providerType: item.provider.type }),
+      )
       .map(({ model }) => ({ id: model.id, name: model.displayName, modelId: model.modelId }));
     return NextResponse.json(
       { capabilities: { ...capabilities, render: renderAvailable }, models },

app/api/canvas/smart-edit/fonts/[id]/route.ts

@@ -0,0 +1,22 @@
+import { auth } from "@/app/(auth)/auth";
+import { captionFont } from "@/lib/video-canvas/smart-edit/fonts";
+import { captionFontFile } from "@/lib/video-canvas/smart-edit/font-files";
+import { readFile } from "node:fs/promises";
+export const runtime = "nodejs";
+export const maxDuration = 180;
+export async function GET(_request: Request, { params }: { params: Promise<{ id: string }> }) {
+  if (!(await auth())?.user?.id) return Response.json({ error: "未登录" }, { status: 401 });
+  const { id } = await params;
+  if (!captionFont(id)) return Response.json({ error: "字体不存在" }, { status: 404 });
+  try {
+    const data = await readFile(await captionFontFile(id));
+    return new Response(new Uint8Array(data), {
+      headers: {
+        "Content-Type": "font/ttf",
+        "Cache-Control": "private, max-age=31536000, immutable",
+      },
+    });
+  } catch {
+    return Response.json({ error: "字体加载失败，请检查字体服务配置后重试" }, { status: 503 });
+  }
+}

app/api/canvas/video-upscale/capabilities/route.ts

@@ -0,0 +1,39 @@
+import { auth } from "@/app/(auth)/auth";
+import { upscaleConfig } from "@/lib/video-canvas/video-upscale-remote";
+import { NextResponse } from "next/server";
+export async function GET() {
+  if (!(await auth())?.user?.id)
+    return NextResponse.json({ error: "未登录" }, { status: 401 });
+  try {
+    const config = upscaleConfig();
+    if (!config)
+      return NextResponse.json({
+        available: false,
+        message: "视频高清服务尚未接通",
+      });
+    const response = await fetch(config.base + "/health", {
+      headers: { authorization: "Bearer " + config.token },
+      signal: AbortSignal.timeout(5000),
+      redirect: "error",
+      cache: "no-store",
+    });
+    if (!response.ok) throw Error("unavailable");
+    const data = await response.json();
+    return NextResponse.json(
+      {
+        available: data.status === "ok",
+        engine: "Real-ESRGAN",
+        resolutions: ["1080P", "2K", "4K"],
+        maxScale: 4,
+        maxSeconds: 120,
+        queued: data.queued,
+      },
+      { headers: { "Cache-Control": "no-store" } }
+    );
+  } catch {
+    return NextResponse.json(
+      { available: false, message: "视频高清服务暂时不可用，请稍后重试" },
+      { headers: { "Cache-Control": "no-store" } }
+    );
+  }
+}

components/canvas/AudioGenerationComposer.tsx

@@ -9,14 +9,16 @@ import {
 type Model = { id: string; modelId: string; name: string; kind: string };
 export function AudioGenerationComposer({
   initial,
+  connectedText = "",
   busy,
   onSubmit,
   onUpload,
   onDraft,
 }: {
   initial?: Record<string, unknown>;
+  connectedText?: string;
   busy: boolean;
-  onSubmit: (params: AudioRequest) => Promise<unknown>;
+  onSubmit: (params: AudioRequest, draft: AudioRequest) => Promise<unknown>;
   onUpload: (file: File) => Promise<void>;
   onDraft: (params: Record<string, unknown>) => void;
 }) {
@@ -73,6 +75,7 @@ export function AudioGenerationComposer({
       if (!model) throw Error("当前模式没有已启用的官方模型");
       const params = audioRequestSchema.parse({
         ...draft,
+        text: [draft.text, connectedText].filter(Boolean).join("\n\n"),
         modelRef: model.id,
         voice,
         format,
@@ -80,8 +83,9 @@ export function AudioGenerationComposer({
         ...(!contract?.instructions ? { instructions: "" } : {}),
       });
       buildAudioRequest(params, model.modelId);
-      patch(params);
+      const savedDraft = { ...params, text: String(draft.text || "") };
-      await onSubmit(params);
+      patch(savedDraft);
+      await onSubmit(params, savedDraft);
     } catch (e) {
       setError(e instanceof Error ? e.message : "提交失败");
     } finally {
@@ -96,6 +100,7 @@ export function AudioGenerationComposer({
     >
       <h3 className="font-semibold">语音与音乐 · 阿里云官方</h3>
       <fieldset disabled={busy || sending} className="mt-3 space-y-3 disabled:opacity-60">
+        {connectedText && <p className="mb-3 whitespace-pre-wrap rounded bg-white/5 p-2 text-xs" aria-label="已连接文本">已连接文本：{connectedText}</p>}
         <div className="grid grid-cols-2 gap-3">
           <label>
             生成类型
@@ -142,7 +147,7 @@ export function AudioGenerationComposer({
         </label>
         {draft.mode === "tts" ? (
           <>
-            <div className="grid grid-cols-2 gap-3">
+              <div className="grid grid-cols-2 gap-3">
               <label>
                 音色
                 <select

components/canvas/CanvasGroupToolbar.tsx

@@ -0,0 +1,125 @@
+"use client";
+import {
+  Download,
+  Play,
+  Ungroup,
+  Share2,
+  Grid3X3,
+  ScanLine,
+} from "lucide-react";
+export function CanvasGroupToolbar({
+  title,
+  color,
+  status,
+  busy,
+  storyboard,
+  canStoryboard,
+  onRename,
+  onColor,
+  onLayout,
+  onRun,
+  onUngroup,
+  onSave,
+  onDownload,
+  onStoryboard,
+  onCancel,
+}: {
+  title: string;
+  color: string;
+  status: string;
+  busy: boolean;
+  storyboard: boolean;
+  canStoryboard: boolean;
+  onRename: (s: string) => void;
+  onColor: (s: string) => void;
+  onLayout: (s: "grid" | "horizontal" | "vertical") => void;
+  onRun: () => void;
+  onUngroup: () => void;
+  onSave: () => void;
+  onDownload: () => void;
+  onStoryboard: () => void;
+  onCancel: () => void;
+}) {
+  return (
+    <div
+      data-canvas-overlay="true"
+      data-group-toolbar="true"
+      className="absolute left-1/2 top-[80px] z-[70] -translate-x-1/2 rounded-[22px] border border-white/10 bg-[#252525] p-4 text-white shadow-xl"
+      onPointerDown={(e) => e.stopPropagation()}
+    >
+      <div className="flex items-center gap-5 whitespace-nowrap text-[16px]">
+        <input
+          aria-label="编组颜色"
+          type="color"
+          value={color}
+          onChange={(e) => onColor(e.target.value)}
+          className="h-7 w-7 cursor-pointer rounded-full border-0 bg-transparent"
+        />
+        <label title="编组布局" className="flex items-center gap-1">
+          <Grid3X3 size={20} />
+          <select
+            aria-label="编组布局"
+            defaultValue=""
+            onChange={(e) => {
+              onLayout(e.target.value as any);
+              e.target.value = "";
+            }}
+            className="w-5 bg-[#252525]"
+          >
+            <option value="" disabled>
+              ⌄
+            </option>
+            <option value="grid">网格布局</option>
+            <option value="horizontal">横向布局</option>
+            <option value="vertical">纵向布局</option>
+          </select>
+        </label>
+        <span className="h-8 w-px bg-white/10" />
+        <button
+          type="button"
+          disabled={busy}
+          onClick={onRun}
+          className="flex items-center gap-2 disabled:opacity-40"
+        >
+          <Play size={22} />
+          整组执行
+        </button>
+        <button onClick={onSave} className="flex items-center gap-2">
+          <Share2 size={22} />
+          添加到工具箱
+        </button>
+        <button
+          type="button"
+          disabled={!canStoryboard}
+          title={
+            canStoryboard ? "" : "分镜组需要全部成员都是已就绪的图片或视频"
+          }
+          onClick={onStoryboard}
+          className="flex items-center gap-2 disabled:opacity-35"
+        >
+          <ScanLine size={22} />
+          {storyboard ? "预览分镜" : "转分镜组"}
+        </button>
+        <button onClick={onUngroup} className="flex items-center gap-2">
+          <Ungroup size={22} />
+          解组
+        </button>
+        <span className="h-8 w-px bg-white/10" />
+        <button title="下载编组模板" onClick={onDownload}>
+          <Download size={22} />
+        </button>
+      </div>
+      <div className="mt-3 flex items-center justify-between gap-4 text-xs text-white/55">
+        <input
+          aria-label="编组名称"
+          value={title}
+          onChange={(e) => onRename(e.target.value)}
+          maxLength={120}
+          className="w-48 border-b border-white/10 bg-transparent py-1"
+        />
+        <span>{status || "Shift 点击多选 · ⌘/Ctrl+G 编组"}</span>
+        {busy && <button onClick={onCancel}>取消未开始任务</button>}
+      </div>
+    </div>
+  );
+}

components/canvas/GroupStoryboardPreview.tsx

@@ -0,0 +1,122 @@
+"use client";
+import { useEffect, useState } from "react";
+import { ArrowLeft, ArrowRight, X } from "lucide-react";
+export type GroupShot = {
+  id: string;
+  title: string;
+  url: string;
+  type: "image" | "video";
+};
+export function GroupStoryboardPreview({
+  shots,
+  onOrder,
+  onClose,
+}: {
+  shots: GroupShot[];
+  onOrder: (ids: string[]) => void;
+  onClose: () => void;
+}) {
+  const [index, setIndex] = useState(0),
+    shot = shots[Math.min(index, shots.length - 1)];
+  useEffect(() => {
+    const key = (e: KeyboardEvent) => {
+      if (e.key === "Escape") onClose();
+    };
+    window.addEventListener("keydown", key);
+    return () => window.removeEventListener("keydown", key);
+  }, [onClose]);
+  if (!shot) return null;
+  const move = (delta: number) => {
+    const next = index + delta;
+    if (next < 0 || next >= shots.length) return;
+    const ids = shots.map((s) => s.id);
+    [ids[index], ids[next]] = [ids[next], ids[index]];
+    onOrder(ids);
+    setIndex(next);
+  };
+  return (
+    <div
+      data-canvas-overlay="true"
+      role="dialog"
+      aria-label="编组分镜预览"
+      className="fixed inset-0 z-[120] flex flex-col bg-black/95 p-8 text-white"
+      onPointerDown={(e) => e.stopPropagation()}
+    >
+      <header className="mb-4 flex items-center justify-between">
+        <h2>
+          分镜 {index + 1}/{shots.length} · {shot.title}
+        </h2>
+        <button aria-label="关闭分镜预览" onClick={onClose}>
+          <X />
+        </button>
+      </header>
+      <div className="flex min-h-0 flex-1 items-center justify-center">
+        {shot.type === "video" ? (
+          <video
+            key={shot.id}
+            src={shot.url}
+            controls
+            autoPlay
+            className="h-full max-w-full object-contain"
+            onEnded={() => setIndex((i) => Math.min(i + 1, shots.length - 1))}
+          />
+        ) : (
+          <img
+            src={shot.url}
+            alt={shot.title}
+            className="h-full max-w-full object-contain"
+          />
+        )}
+      </div>
+      <div className="mt-4 flex items-center justify-center gap-6">
+        <button
+          type="button"
+          disabled={index === 0}
+          onClick={() => setIndex((i) => i - 1)}
+          className="disabled:opacity-30"
+        >
+          <ArrowLeft />
+        </button>
+        <span>
+          {index + 1} / {shots.length}
+        </span>
+        <button
+          type="button"
+          disabled={index === shots.length - 1}
+          onClick={() => setIndex((i) => i + 1)}
+          className="disabled:opacity-30"
+        >
+          <ArrowRight />
+        </button>
+        <button
+          type="button"
+          disabled={index === 0}
+          onClick={() => move(-1)}
+          className="disabled:opacity-30"
+        >
+          镜头前移
+        </button>
+        <button
+          type="button"
+          disabled={index === shots.length - 1}
+          onClick={() => move(1)}
+          className="disabled:opacity-30"
+        >
+          镜头后移
+        </button>
+      </div>
+      <div className="mt-5 flex justify-center gap-3 overflow-x-auto">
+        {shots.map((s, i) => (
+          <button
+            type="button"
+            key={s.id}
+            onClick={() => setIndex(i)}
+            className={`shrink-0 rounded-lg border px-4 py-2 text-sm ${i === index ? "border-white bg-white/20" : "border-white/20"}`}
+          >
+            {i + 1} · {s.title}
+          </button>
+        ))}
+      </div>
+    </div>
+  );
+}

components/canvas/smart-edit/CaptionFonts.tsx

@@ -0,0 +1,226 @@
+"use client";
+import { useEffect, useLayoutEffect, useRef, useState } from "react";
+import { Check, ChevronDown, Search } from "lucide-react";
+import { captionFonts, captionFont } from "@/lib/video-canvas/smart-edit/fonts";
+import type { CaptionStyle } from "@/lib/video-canvas/smart-edit/contract";
+const loaded = new Map<string, Promise<FontFace>>();
+export function loadCaptionFont(id: string) {
+  const font = captionFont(id);
+  if (!font) return Promise.resolve();
+  if (!loaded.has(id)) {
+    const face = new FontFace(
+      font.family,
+      `url(${font.url ? `/api/canvas/smart-edit/fonts/${font.id}` : `/fonts/smart-edit/${font.file}`})`,
+      {
+        weight: "400",
+      },
+    );
+    loaded.set(
+      id,
+      face
+        .load()
+        .then((loadedFace) => {
+          document.fonts.add(loadedFace);
+          return loadedFace;
+        })
+        .catch((error) => {
+          loaded.delete(id);
+          throw error;
+        }),
+    );
+  }
+  return loaded.get(id)!;
+}
+export function CaptionFontPicker({
+  value,
+  onChange,
+  onError,
+}: {
+  value: CaptionStyle["fontFamily"];
+  onChange: (id: NonNullable<CaptionStyle["fontFamily"]>) => void;
+  onError: (message: string) => void;
+}) {
+  const [open, setOpen] = useState(false),
+    [query, setQuery] = useState(""),
+    [category, setCategory] = useState("全部"),
+    [categoriesOpen, setCategoriesOpen] = useState(false),
+    [busy, setBusy] = useState(false),
+    [scrollTop, setScrollTop] = useState(0);
+  const root = useRef<HTMLDivElement>(null);
+  const [placement, setPlacement] = useState({ left: 0, top: 0, width: 300, height: 300 });
+  useLayoutEffect(() => {
+    if (!open) return;
+    const place = () => {
+      const rect = root.current?.getBoundingClientRect();
+      if (!rect) return;
+      const below = window.innerHeight - rect.bottom - 16;
+      const above = rect.top - 16;
+      const up = below < 394 && above > below;
+      const height = Math.max(40, Math.min(300, (up ? above : below) - 94));
+      const width = Math.min(window.innerWidth - 16, rect.width + 32);
+      setPlacement({
+        left: Math.max(8, Math.min(rect.left - 16, window.innerWidth - width - 8)),
+        top: up ? Math.max(8, rect.top - height - 94 - 8) : rect.bottom + 8,
+        width,
+        height,
+      });
+    };
+    place();
+    const scroll = (event: Event) => {
+      if (!root.current?.contains(event.target as Node)) place();
+    };
+    window.addEventListener("resize", place);
+    window.addEventListener("scroll", scroll, true);
+    return () => {
+      window.removeEventListener("resize", place);
+      window.removeEventListener("scroll", scroll, true);
+    };
+  }, [open]);
+  useEffect(() => {
+    const close = (e: globalThis.PointerEvent) => {
+      if (!root.current?.contains(e.target as Node)) setOpen(false);
+    };
+    document.addEventListener("pointerdown", close);
+    return () => document.removeEventListener("pointerdown", close);
+  }, []);
+  const options = captionFonts.filter(
+    (font) =>
+      font.name.toLowerCase().includes(query.toLowerCase()) &&
+      (category === "全部" ||
+        (category === "常用" && font.common) ||
+        (category === "推荐" && font.recommended) ||
+        (category === "Google" && font.google)),
+  );
+  const first = Math.max(0, Math.floor(scrollTop / 40) - 4);
+  useEffect(() => {
+    setScrollTop(0);
+  }, [query, category, open]);
+  useEffect(() => {
+    if (!open) return;
+    let stopped = false,
+      index = 0;
+    const visible = options.slice(first, first + 18);
+    const worker = async () => {
+      while (!stopped && index < visible.length) {
+        const font = visible[index++];
+        try {
+          await loadCaptionFont(font.id);
+        } catch {
+          /* Selection reports any download failure. */
+        }
+      }
+    };
+    void worker();
+    void worker();
+    return () => {
+      stopped = true;
+    };
+  }, [open, first, query, category]);
+  return (
+    <div className="se-font-picker" ref={root}>
+      <button
+        className="se-font"
+        aria-label="字幕字体"
+        aria-expanded={open}
+        onClick={() => setOpen(!open)}
+      >
+        {captionFont(value)?.name || "默认字体"}
+        <ChevronDown size={14} />
+      </button>
+      {open && (
+        <div
+          className="se-font-menu"
+          style={{
+            position: "fixed",
+            left: placement.left,
+            top: placement.top,
+            width: placement.width,
+            right: "auto",
+          }}
+        >
+          <div className="se-search">
+            <Search size={14} />
+            <input
+              aria-label="搜索字体"
+              placeholder="搜索字体"
+              value={query}
+              onChange={(e) => setQuery(e.target.value)}
+            />
+          </div>
+          <div className="se-font-categories">
+            <button
+              type="button"
+              aria-haspopup="menu"
+              aria-expanded={categoriesOpen}
+              onClick={() => setCategoriesOpen(!categoriesOpen)}
+            >
+              {category === "Google" ? "Google 字体" : `${category}字体`}
+              <ChevronDown size={12} />
+            </button>
+            {categoriesOpen && (
+              <div role="menu" aria-label="字体分类">
+                {["全部", "常用", "推荐", "Google"].map((name) => (
+                  <button
+                    type="button"
+                    role="menuitemradio"
+                    aria-checked={category === name}
+                    key={name}
+                    onClick={() => {
+                      setCategory(name);
+                      setCategoriesOpen(false);
+                    }}
+                  >
+                    {name === "Google" ? "Google 字体" : `${name}字体`}
+                    {category === name && <Check size={14} />}
+                  </button>
+                ))}
+              </div>
+            )}
+          </div>
+          <div
+            key={`${query}:${category}`}
+            role="listbox"
+            style={{ height: placement.height }}
+            aria-label={`字体库，共 ${options.length} 项`}
+            onScroll={(e) => setScrollTop(e.currentTarget.scrollTop)}
+          >
+            <div style={{ height: options.length * 40, position: "relative" }}>
+              {options.slice(first, first + 18).map((font, index) => (
+                <button
+                  style={{
+                    position: "absolute",
+                    top: (first + index) * 40,
+                    fontFamily: `"${font.family}", "Source Han Sans SC", sans-serif`,
+                  }}
+                  aria-posinset={first + index + 1}
+                  aria-setsize={options.length}
+                  role="option"
+                  aria-selected={value === font.id}
+                  key={font.id}
+                  disabled={busy}
+                  onClick={async () => {
+                    setBusy(true);
+                    try {
+                      await loadCaptionFont(font.id);
+                      onChange(font.id);
+                      setOpen(false);
+                    } catch {
+                      onError("字体加载失败，请检查网络后重试");
+                    } finally {
+                      setBusy(false);
+                    }
+                  }}
+                >
+                  {font.name}
+                  {value === font.id && <Check size={14} />}
+                </button>
+              ))}
+            </div>
+          </div>
+          {!options.length && <div className="se-muted">没有匹配的字体</div>}
+          {busy && <div role="status">正在加载字体…</div>}
+        </div>
+      )}
+    </div>
+  );
+}

components/canvas/smart-edit/EditPreview.tsx

@@ -1,10 +1,18 @@
 "use client";
-import { useEffect, useRef } from "react";
+import { useEffect, useRef, useState } from "react";
+import { captionFont } from "@/lib/video-canvas/smart-edit/fonts";
+import { loadCaptionFont } from "./CaptionFonts";
+import { usePreviewGain } from "./usePreviewGain";
 import {
-  compiledSubtitles,
+  editableCaptions,
+  defaultCaptionStyle,
+  type CaptionStyle,
   planTimeline,
   SECOND,
   type EditPlan,
+  defaultVideoTransform,
+  type VideoTransform,
+  type TimelineCaption,
 } from "@/lib/video-canvas/smart-edit/contract";
 export function EditPreview({
   plan,
@@ -13,7 +21,25 @@ export function EditPreview({
   onTime,
   onPlaying,
   onError,
+  fit = false,
+  cover = false,
+  externalClock = false,
+  selectedClipId,
+  selectedCaptionId,
+  onSelectClip,
+  onSelectCaption,
+  onTransform,
+  onCaptionPosition,
 }: {
+  fit?: boolean;
+  cover?: boolean;
+  externalClock?: boolean;
+  selectedClipId?: string;
+  selectedCaptionId?: string;
+  onSelectClip?: (id: string) => void;
+  onSelectCaption?: (id: string) => void;
+  onTransform?: (id: string, value: VideoTransform) => void;
+  onCaptionPosition?: (id: string, value: { x: number; y: number }) => void;
   plan: EditPlan;
   time: number;
   playing: boolean;
@@ -23,13 +49,51 @@ export function EditPreview({
 }) {
   const videos = useRef(new Map<string, HTMLVideoElement>()),
     music = useRef<HTMLAudioElement | null>(null),
+    audios = useRef(new Map<string, HTMLAudioElement>()),
     latest = useRef({ time, plan, onTime, onPlaying });
   latest.current = { time, plan, onTime, onPlaying };
-  const timeline = planTimeline(plan),
+  const fontIds = [
-    current = timeline.clips.filter((c) => time * SECOND >= c.start && time * SECOND < c.end),
+    ...new Set(
-    subs = compiledSubtitles(plan).filter((s) => time * SECOND >= s.start && time * SECOND < s.end);
+      editableCaptions(plan)
+        .map((c) => c.style?.fontFamily)
+        .filter(Boolean),
+    ),
+  ].join(",");
   useEffect(() => {
-    if (!playing) return;
+    let canceled = false;
+    void Promise.all(
+      [
+        ...new Set([...fontIds.split(",").filter(Boolean), ...(fontIds ? ["source-han"] : [])]),
+      ].map(loadCaptionFont),
+    ).catch(() => {
+      if (!canceled) onError("字幕字体加载失败，请检查网络后重试");
+    });
+    return () => {
+      canceled = true;
+    };
+  }, [fontIds]);
+  const audioGain = usePreviewGain(plan, onError);
+  const frameRef = useRef<HTMLDivElement>(null);
+  const [frameSize, setFrameSize] = useState({ width: 0, height: 0 });
+  useEffect(() => {
+    const el = frameRef.current;
+    if (!el) return;
+    const observer = new ResizeObserver(([entry]) =>
+      setFrameSize({ width: entry.contentRect.width, height: entry.contentRect.height }),
+    );
+    observer.observe(el);
+    return () => observer.disconnect();
+  }, []);
+  const timeline = planTimeline(plan),
+    visibleTime = Math.min(time * SECOND, timeline.duration - 1),
+    current = timeline.clips.filter(
+      (c) => visibleTime >= c.renderStart && visibleTime < c.renderEnd,
+    ),
+    subs = plan.subtitles
+      ? editableCaptions(plan).filter((s) => time * SECOND >= s.start && time * SECOND < s.end)
+      : [];
+  useEffect(() => {
+    if (!playing || externalClock) return;
     let frame = 0,
       last = 0;
     const tick = (now: number) => {
@@ -57,8 +121,9 @@ export function EditPreview({
       cancelAnimationFrame(frame);
       document.removeEventListener("visibilitychange", hide);
     };
-  }, [playing]);
+  }, [playing, externalClock]);
   useEffect(() => {
+    audioGain.retain([...videos.current.values(), ...audios.current.values()]);
     for (const clip of timeline.clips) {
       const video = videos.current.get(clip.id);
       if (!video) continue;
@@ -67,25 +132,97 @@ export function EditPreview({
         video.pause();
         continue;
       }
-      const target = clip.sourceIn / SECOND + (time - clip.start / SECOND) * clip.speed;
+      const target = Math.max(
+        clip.sourceIn / SECOND,
+        Math.min(
+          clip.sourceOut / SECOND - 1 / 30,
+          clip.sourceIn / SECOND + (time - clip.start / SECOND) * clip.speed,
+        ),
+      );
+      const held = time * SECOND < clip.start || time * SECOND >= clip.end;
       if (Number.isFinite(target) && Math.abs(video.currentTime - target) > 0.1)
         video.currentTime = target;
       video.playbackRate = clip.speed;
       const index = timeline.clips.findIndex((c) => c.id === clip.id),
         next = timeline.clips[index + 1];
       let gain = 1;
-      if (clip.overlap && time * SECOND < clip.start + clip.overlap)
+      if (clip.overlap && time * SECOND < clip.transitionStart + clip.overlap)
-        gain = (time * SECOND - clip.start) / clip.overlap;
+        gain = (time * SECOND - clip.transitionStart) / clip.overlap;
-      if (next?.overlap && time * SECOND > next.start)
+      if (next?.overlap && time * SECOND > next.transitionStart)
-        gain = Math.min(gain, (clip.end - time * SECOND) / next.overlap);
+        gain = Math.min(
-      video.volume = Math.min(1, Math.max(0, clip.gain * gain));
+          gain,
-      if (playing)
+          (next.transitionStart + next.overlap - time * SECOND) / next.overlap,
-        void video.play().catch(() => {
+        );
+      const source = plan.assets.find((a) => a.id === clip.assetId)!.url;
+      if (!audioGain.ready(source)) {
+        video.pause();
+        if (playing) {
+          audioGain.retry();
+          onPlaying(false);
+          onError("正在准备音量预览素材，请稍后点击播放");
+        }
+        continue;
+      }
+      audioGain.volume(
+        video,
+        clip.audioDetached || clip.muted || held ? 0 : clip.gain * gain,
+        playing,
+      );
+      if (playing && !held)
+        void video.play().catch((error) => {
+          if (error?.name === "AbortError") return;
           onPlaying(false);
           onError("预览播放失败，请检查素材链接或点击播放重试");
         });
       else video.pause();
     }
+    for (const clip of timeline.audioClips) {
+      const audio = audios.current.get(clip.id);
+      if (!audio) continue;
+      const relative = time - clip.start / SECOND;
+      if (relative < 0 || time * SECOND >= clip.end) {
+        audio.pause();
+        continue;
+      }
+      const target = clip.sourceIn / SECOND + relative * clip.speed;
+      if (Math.abs(audio.currentTime - target) > 0.1) audio.currentTime = target;
+      audio.playbackRate = clip.speed;
+      const length = clip.duration / SECOND,
+        fadeIn = Math.min(clip.fadeIn, length),
+        fadeOut = Math.min(clip.fadeOut, length);
+      const gain =
+        Math.min(
+          1,
+          fadeIn
+            ? (clip.fadeInStart || 0) + ((1 - (clip.fadeInStart || 0)) * relative) / fadeIn
+            : 1,
+        ) *
+        Math.min(
+          1,
+          fadeOut
+            ? (clip.fadeOutEnd || 0) +
+                ((1 - (clip.fadeOutEnd || 0)) * (length - relative)) / fadeOut
+            : 1,
+        );
+      const source = plan.assets.find((a) => a.id === clip.assetId)!.url;
+      if (!audioGain.ready(source)) {
+        audio.pause();
+        if (playing) {
+          audioGain.retry();
+          onPlaying(false);
+          onError("正在准备音量预览素材，请稍后点击播放");
+        }
+        continue;
+      }
+      audioGain.volume(audio, clip.muted ? 0 : clip.gain * gain, playing);
+      if (playing)
+        void audio.play().catch((error) => {
+          if (error?.name === "AbortError") return;
+          onPlaying(false);
+          onError("独立音轨无法播放");
+        });
+      else audio.pause();
+    }
     const audio = music.current;
     if (audio && plan.music) {
       const target = time + plan.music.start / SECOND;
@@ -96,53 +233,328 @@ export function EditPreview({
       if (Math.abs(audio.currentTime - target) > 0.15) audio.currentTime = target;
       audio.volume = Math.min(1, plan.music.gain);
       if (playing)
-        void audio.play().catch(() => {
+        void audio.play().catch((error) => {
+          if (error?.name === "AbortError") return;
           onPlaying(false);
           onError("背景音乐无法播放");
         });
       else audio.pause();
     }
   }, [time, playing, plan]);
+  const [ratioWidth, ratioHeight] = plan.ratio.split(":").map(Number);
+  const ratio = ratioWidth / ratioHeight;
+  const height = Math.min(frameSize.height || 720, (frameSize.width || 1280) / ratio);
+  const incomingClip = current.find(
+    (c) => c.overlap && visibleTime < c.transitionStart + c.overlap,
+  );
+  const progress = incomingClip
+    ? (visibleTime - incomingClip.transitionStart) / incomingClip.overlap
+    : 1;
+  const [gesture, setGesture] = useState<{
+    clipId?: string;
+    captionId?: string;
+    transform?: VideoTransform;
+    x?: number;
+    y?: number;
+  }>();
+  const gestureCleanup = useRef<() => void>();
+  useEffect(() => () => gestureCleanup.current?.(), []);
+  const moveTransform = (e: React.PointerEvent, id: string, corner?: [number, number]) => {
+    e.preventDefault();
+    e.stopPropagation();
+    onPlaying(false);
+    onSelectClip?.(id);
+    const c = plan.clips.find((c) => c.id === id)!,
+      a = plan.assets.find((a) => a.id === c.assetId)!,
+      original = c.transform || defaultVideoTransform;
+    const frameWidth = height * ratio,
+      baseWidth = Math.min(frameWidth, (height * a.width) / a.height),
+      baseHeight = (baseWidth * a.height) / a.width,
+      startX = e.clientX,
+      startY = e.clientY;
+    let value = original;
+    const clamp = (v: number, lo: number, hi: number) => Math.max(lo, Math.min(hi, v));
+    const move = (ev: PointerEvent) => {
+      const dx = ev.clientX - startX,
+        dy = ev.clientY - startY;
+      if (corner) {
+        const nextScale = clamp(
+            original.scale + ((dx * corner[0]) / baseWidth + (dy * corner[1]) / baseHeight) / 2,
+            0.1,
+            3,
+          ),
+          delta = nextScale - original.scale;
+        value = {
+          scale: nextScale,
+          x: clamp(original.x + (delta * baseWidth * corner[0]) / 2 / frameWidth, -1, 1),
+          y: clamp(original.y + (delta * baseHeight * corner[1]) / 2 / height, -1, 1),
+        };
+      } else
+        value = {
+          ...original,
+          x: clamp(original.x + dx / frameWidth, -1, 1),
+          y: clamp(original.y + dy / height, -1, 1),
+        };
+      setGesture({ clipId: id, transform: value });
+    };
+    const cleanup = () => {
+      window.removeEventListener("pointermove", move);
+      window.removeEventListener("pointerup", end);
+      window.removeEventListener("pointercancel", cancel);
+      gestureCleanup.current = undefined;
+    };
+    const end = () => {
+        cleanup();
+        setGesture(undefined);
+        onTransform?.(id, value);
+      },
+      cancel = () => {
+        cleanup();
+        setGesture(undefined);
+      };
+    gestureCleanup.current?.();
+    gestureCleanup.current = cleanup;
+    window.addEventListener("pointermove", move);
+    window.addEventListener("pointerup", end);
+    window.addEventListener("pointercancel", cancel);
+  };
+  const moveCaption = (e: React.PointerEvent, id: string, style: CaptionStyle) => {
+    e.preventDefault();
+    e.stopPropagation();
+    onPlaying(false);
+    onSelectCaption?.(id);
+    const startX = e.clientX,
+      startY = e.clientY;
+    let value = { x: style.x ?? 0.5, y: style.y ?? 0.95 };
+    const move = (ev: PointerEvent) => {
+      value = {
+        x: Math.max(
+          0,
+          Math.min(1, (style.x ?? 0.5) + (ev.clientX - startX) / (height * ratio)),
+        ),
+        y: Math.max(0, Math.min(1, (style.y ?? 0.95) + (ev.clientY - startY) / height)),
+      };
+      setGesture({ captionId: id, ...value });
+    };
+    const cleanup = () => {
+      window.removeEventListener("pointermove", move);
+      window.removeEventListener("pointerup", end);
+      window.removeEventListener("pointercancel", cancel);
+      gestureCleanup.current = undefined;
+    };
+    const end = () => {
+        cleanup();
+        setGesture(undefined);
+        onCaptionPosition?.(id, value);
+      },
+      cancel = () => {
+        cleanup();
+        setGesture(undefined);
+      };
+    gestureCleanup.current?.();
+    gestureCleanup.current = cleanup;
+    window.addEventListener("pointermove", move);
+    window.addEventListener("pointerup", end);
+    window.addEventListener("pointercancel", cancel);
+  };
+  const selectedClip = current.find((c) => c.id === selectedClipId);
+  const selectedAsset = selectedClip && plan.assets.find((a) => a.id === selectedClip.assetId);
+  const selectedTransform =
+    gesture && gesture.clipId === selectedClipId
+      ? gesture.transform
+      : selectedClip?.transform || defaultVideoTransform;
   return (
     <div
-      className="relative mx-auto w-full max-w-4xl overflow-hidden bg-black"
+      ref={frameRef}
-      style={{ aspectRatio: plan.ratio.replace(":", "/") }}
+      className={
+        fit
+          ? "relative h-full w-full overflow-hidden"
+          : "relative mx-auto w-full max-w-4xl overflow-hidden bg-black"
+      }
+      style={fit ? undefined : { aspectRatio: plan.ratio.replace(":", "/") }}
       aria-label="剪辑实时预览"
     >
-      {timeline.clips.map((clip) => {
+      <div
-        const asset = plan.assets.find((a) => a.id === clip.assetId)!;
+        style={
-        const active = current.some((c) => c.id === clip.id),
+          fit && !cover
-          incoming = clip.overlap && time * SECOND < clip.start + clip.overlap;
+            ? {
-        return (
+                position: "absolute",
-          <video
+                width: height * ratio,
-            key={clip.id}
+                height,
-            ref={(element) => {
+                left: "50%",
-              if (element) videos.current.set(clip.id, element);
+                top: "50%",
-              else videos.current.delete(clip.id);
+                transform: "translate(-50%,-50%)",
+                overflow: "hidden",
+                background: "black",
+              }
+            : { position: "absolute", inset: 0 }
+        }
+      >
+        {timeline.clips.map((clip) => {
+          const asset = plan.assets.find((a) => a.id === clip.assetId)!;
+          const active = current.some((c) => c.id === clip.id),
+            incoming = clip.overlap && visibleTime < clip.transitionStart + clip.overlap;
+          return (
+            <video
+              key={`${clip.id}:${audioGain.source(asset.url)}`}
+              ref={(element) => {
+                if (element) videos.current.set(clip.id, element);
+                else videos.current.delete(clip.id);
+              }}
+              src={audioGain.source(asset.url)}
+              muted={clip.audioDetached || clip.muted || clip.gain === 0}
+              playsInline
+              preload="metadata"
+              className={`absolute inset-0 h-full w-full ${cover ? "object-cover" : "object-contain"}`}
+              style={{
+                visibility: active ? "visible" : "hidden",
+                transform: (() => {
+                  const t =
+                    gesture?.clipId === clip.id
+                      ? gesture.transform!
+                      : clip.transform || defaultVideoTransform;
+                  return `translate(${t.x * 100}%,${t.y * 100}%) scale(${t.scale})`;
+                })(),
+                opacity: incoming
+                  ? clip.transitionType === "fade"
+                    ? progress
+                    : progress < 0.5
+                      ? 0
+                      : 1
+                  : 1,
+              }}
+              onLoadedMetadata={(e) => {
+                const target =
+                  clip.sourceIn / SECOND +
+                  (Math.max(0, visibleTime - clip.start) / SECOND) * clip.speed;
+                e.currentTarget.currentTime = Math.min(clip.sourceOut / SECOND - 0.001, target);
+              }}
+              onError={() => {
+                onPlaying(false);
+                onError(`${asset.name} 预览加载失败`);
+              }}
+            />
+          );
+        })}
+        {timeline.audioClips.map((c) => (
+          <audio
+            key={`${c.id}:${audioGain.source(plan.assets.find((a) => a.id === c.assetId)!.url)}`}
+            ref={(el) => {
+              if (el) audios.current.set(c.id, el);
+              else audios.current.delete(c.id);
             }}
-            src={asset.url}
+            src={audioGain.source(plan.assets.find((a) => a.id === c.assetId)!.url)}
-            playsInline
             preload="metadata"
-            className="absolute inset-0 h-full w-full object-contain"
+          />
+        ))}
+        {plan.music ? <audio ref={music} src={plan.music.url} preload="metadata" /> : null}
+        {incomingClip && incomingClip.transitionType !== "fade" && (
+          <div
             style={{
-              visibility: active ? "visible" : "hidden",
+              position: "absolute",
-              opacity: incoming ? Math.max(0, (time * SECOND - clip.start) / clip.overlap) : 1,
+              inset: 0,
-            }}
+              pointerEvents: "none",
-            onError={() => {
+              background: incomingClip.transitionType === "fadeblack" ? "black" : "white",
-              onPlaying(false);
+              opacity: 1 - Math.abs(2 * progress - 1),
-              onError(`${asset.name} 预览加载失败`);
             }}
           />
-        );
+        )}
-      })}
+        {selectedClip && selectedAsset && selectedTransform && onTransform && (
-      {plan.music ? <audio ref={music} src={plan.music.url} preload="metadata" /> : null}
+          <div
-      <div
+            aria-label="调整画面位置"
-        className="pointer-events-none absolute inset-x-4 bottom-[5%] text-center text-[clamp(12px,2vw,22px)] text-white"
+            onClick={(e) => e.stopPropagation()}
-        style={{ textShadow: "-1px -1px 2px black,1px 1px 2px black" }}
+            onPointerDown={(e) => moveTransform(e, selectedClip.id)}
-      >
+            style={{
-        {subs.map((s, i) => (
+              position: "absolute",
-          <div key={i}>{s.text}</div>
+              left: `${(selectedTransform.x + 0.5) * 100}%`,
-        ))}
+              top: `${(selectedTransform.y + 0.5) * 100}%`,
+              width:
+                Math.min(
+                  height * ratio,
+                  (height * selectedAsset.width) / selectedAsset.height,
+                ) * selectedTransform.scale,
+              height:
+                Math.min(
+                  height,
+                  (height * ratio * selectedAsset.height) / selectedAsset.width,
+                ) * selectedTransform.scale,
+              transform: "translate(-50%,-50%)",
+              border: "2px solid #87c9ff",
+              cursor: "move",
+              zIndex: 3,
+            }}
+          >
+            {(
+              [
+                [-1, -1],
+                [1, -1],
+                [1, 1],
+                [-1, 1],
+              ] as [number, number][]
+            ).map(([x, y]) => (
+              <button
+                key={`${x}:${y}`}
+                type="button"
+                aria-label="调整视频尺寸"
+                onPointerDown={(e) => moveTransform(e, selectedClip.id, [x, y])}
+                style={{
+                  position: "absolute",
+                  width: 12,
+                  height: 12,
+                  left: x < 0 ? 0 : "100%",
+                  top: y < 0 ? 0 : "100%",
+                  transform: "translate(-50%,-50%)",
+                  border: "1px solid white",
+                  background: "#87c9ff",
+                  borderRadius: 2,
+                  cursor: x === y ? "nwse-resize" : "nesw-resize",
+                }}
+              />
+            ))}
+          </div>
+        )}
+        <div className="pointer-events-none absolute inset-0 text-center" style={{ zIndex: 4 }}>
+          {subs.map((s, i) => {
+            const style = (s as { style?: CaptionStyle }).style || defaultCaptionStyle;
+            return (
+              <div
+                key={i}
+                onClick={(e) => e.stopPropagation()}
+                onPointerDown={(e) => {
+                  if (onCaptionPosition && "id" in s) moveCaption(e, s.id as string, style);
+                }}
+                style={{
+                  position: "absolute",
+                  left: `${((gesture && gesture.captionId === (s as TimelineCaption).id ? gesture.x : style.x) ?? 0.5) * 100}%`,
+                  bottom: `${(1 - ((gesture && gesture.captionId === (s as TimelineCaption).id ? gesture.y : style.y) ?? 0.95)) * 100}%`,
+                  transform: "translateX(-50%)",
+                  maxWidth: "95%",
+                  pointerEvents: onCaptionPosition ? "auto" : "none",
+                  cursor: onCaptionPosition ? "move" : undefined,
+                  outline:
+                    selectedCaptionId === (s as TimelineCaption).id
+                      ? "1px solid #87c9ff"
+                      : undefined,
+                  fontFamily: `"${captionFont(style.fontFamily)?.family || "Noto Sans CJK SC"}", "Source Han Sans SC", sans-serif`,
+                  whiteSpace: "pre-wrap",
+                  fontSize: (style.fontSize * height) / (style.referenceHeight || 720),
+                  letterSpacing: (style.spacing * height) / (style.referenceHeight || 720),
+                  color: style.color,
+                  opacity: style.opacity,
+                  WebkitTextStroke: `${((style.outlineWidth * height) / (style.referenceHeight || 720)) * 2}px ${style.outlineColor}${Math.round(
+                    (style.outlineOpacity ?? 1) * 255,
+                  )
+                    .toString(16)
+                    .padStart(2, "0")}`,
+                  paintOrder: "stroke fill",
+                  lineHeight: 1.2,
+                }}
+              >
+                {s.text}
+              </div>
+            );
+          })}
+        </div>
       </div>
     </div>
   );

components/canvas/smart-edit/SmartEditDialog.tsx

@@ -1,6 +1,9 @@
 "use client";
 import { useEffect, useRef, useState } from "react";
 import { X, Upload } from "lucide-react";
+import { hasRecoverableCanvasTask } from "@/lib/video-canvas/task-recovery";
+import { SmartEditEditor, SmartEditEmptyEditor } from "./SmartEditEditor";
+import { assertBeatMusic } from "@/lib/video-canvas/smart-edit/intent";
 import { EditPreview } from "./EditPreview";
 import {
   smartEditRequestSchema,
@@ -19,6 +22,7 @@ export function SmartEditDialog({
   plans: serverPlans,
   planTaskId,
   sources,
+  connectedSourceIds = [],
   audios,
   preferredModelRef,
   tasks,
@@ -26,11 +30,13 @@ export function SmartEditDialog({
   onDraft,
   onPlan,
   onRender,
+  onCancel,
 }: {
   initial?: Record<string, any>;
   plans?: EditPlan[];
   planTaskId?: string;
   sources: Asset[];
+  connectedSourceIds?: string[];
   audios: { url: string; name: string }[];
   preferredModelRef?: string;
   tasks: CanvasGenerationTask[];
@@ -38,14 +44,15 @@ export function SmartEditDialog({
   onDraft: (draft: Record<string, unknown>) => void;
   onPlan: (params: SmartEditRequest) => Promise<unknown>;
   onRender: (plan: EditPlan) => Promise<unknown>;
+  onCancel?: (taskId: string) => Promise<unknown>;
 }) {
   const [draft, setDraft] = useState<Record<string, any>>(() => ({
     assets: [],
-    mode: "montage",
+    mode: "auto",
     prompt: "",
     targetDuration: 20,
     ratio: "16:9",
-    resolution: "720p",
+    resolution: "1080p",
     semantic: true,
     removeSilence: false,
     transcribe: false,
@@ -55,7 +62,9 @@ export function SmartEditDialog({
     preferredModelRef,
     ...initial,
     plans:
-      initial?.planTaskId === planTaskId ? initial?.plans || serverPlans || [] : serverPlans || [],
+      initial?.planTaskId === planTaskId
+        ? initial?.plans || serverPlans || []
+        : serverPlans || [],
     planTaskId,
   }));
   const latest = useRef(draft);
@@ -66,7 +75,12 @@ export function SmartEditDialog({
     setDraft(next);
     onDraft(next);
   };
-  const [selected, setSelected] = useState(0),
+  const [selected, setSelected] = useState(() =>
+      Math.max(
+        0,
+        Math.min((draft.plans?.length || 1) - 1, Number(initial?.selectedPlanIndex) || 0),
+      ),
+    ),
     [clipIndex, setClipIndex] = useState(0),
     [time, setTime] = useState(0),
     [playing, setPlaying] = useState(false),
@@ -75,14 +89,22 @@ export function SmartEditDialog({
     [cap, setCap] = useState<any>(),
     [zoom, setZoom] = useState(1),
     [snap, setSnap] = useState(true);
+  const [configure, setConfigure] = useState(false);
   const applied = useRef(planTaskId);
   useEffect(() => {
     if (planTaskId && planTaskId !== applied.current) {
       applied.current = planTaskId;
-      patch({ plans: serverPlans || [], planTaskId });
+      const processing = tasks.some(
-      setSelected(0);
+        (t) =>
+          t.id === planTaskId &&
+          ["separate", "transcribe"].includes(String(t.request.params.operation)),
+      );
+      const index = processing ? Math.min(selected, (serverPlans?.length || 1) - 1) : 0;
+      patch({ plans: serverPlans || [], planTaskId, selectedPlanIndex: index });
+      setSelected(index);
       setTime(0);
       setPlaying(false);
+      setConfigure(false);
     }
   }, [planTaskId, serverPlans]);
   useEffect(() => {
@@ -108,9 +130,51 @@ export function SmartEditDialog({
   } catch (e) {
     validation = (e as Error).message;
   }
+  const autoPrepared = useRef(false);
+  const startSequence = async () => {
+    const assets = [...new Set(connectedSourceIds)]
+      .map((id) => sources.find((a) => a.id === id))
+      .filter((a): a is Asset => Boolean(a));
+    if (!assets.length) return;
+    autoPrepared.current = true;
+    setBusy(true);
+    setError("");
+    try {
+      const params = smartEditRequestSchema.parse({
+        operation: "plan",
+        assets,
+        mode: "sequence",
+        targetDuration: 300,
+        semantic: false,
+        ratio: draft.ratio,
+        resolution: draft.resolution,
+      });
+      patch({ assets });
+      const preparationRequestId = await onPlan(params);
+      patch({ preparationRequestId });
+    } catch (e) {
+      setError((e as Error).message);
+    } finally {
+      setBusy(false);
+    }
+  };
+  useEffect(() => {
+    if (
+      !plan &&
+      !configure &&
+      !autoPrepared.current &&
+      !draft.preparationRequestId &&
+      !waiting &&
+      connectedSourceIds.length
+    )
+      void startSequence();
+  }, [Boolean(plan), configure, connectedSourceIds.join("|"), waiting]);
   const changePlan = (next: EditPlan) => {
     setPlaying(false);
-    patch({ plans: draft.plans.map((p: EditPlan, i: number) => (i === selected ? next : p)) });
+    patch({
+      selectedPlanIndex: selected,
+      plans: draft.plans.map((p: EditPlan, i: number) => (i === selected ? next : p)),
+    });
   };
   const saveFile = (name: string, text: string) => {
     const url = URL.createObjectURL(new Blob([text], { type: "text/plain;charset=utf-8" }));
@@ -166,13 +230,78 @@ export function SmartEditDialog({
         right = { ...plan.clips[clipIndex], id: crypto.randomUUID(), sourceIn: point };
       changePlan({
         ...plan,
-        clips: [...plan.clips.slice(0, clipIndex), left, right, ...plan.clips.slice(clipIndex + 1)],
+        clips: [
+          ...plan.clips.slice(0, clipIndex),
+          left,
+          right,
+          ...plan.clips.slice(clipIndex + 1),
+        ],
       });
     } else clipEdit(clipIndex, action === "left" ? { sourceIn: point } : { sourceOut: point });
   };
   const button =
       "rounded border border-white/20 px-3 py-2 text-xs hover:bg-white/10 disabled:opacity-40",
     input = "rounded border border-white/15 bg-[#303030] px-2 py-1 text-sm";
+  if (!plan && !configure)
+    return (
+      <SmartEditEmptyEditor
+        onClose={onClose}
+        onConfigure={() => setConfigure(true)}
+        onRetry={() => void startSequence()}
+        loading={busy || waiting || Boolean(draft.preparationRequestId && !last)}
+        message={
+          error ||
+          last?.error ||
+          last?.result?.smartEdit?.stage ||
+          (connectedSourceIds.length ? "读取已连接的视频素材" : "连接视频素材后开始剪辑")
+        }
+      />
+    );
+  const audioTask = [...tasks]
+    .reverse()
+    .find((t) => ["separate", "transcribe"].includes(String(t.request.params.operation)));
+  const audioRecovering =
+    audioTask?.status === "interrupted" && hasRecoverableCanvasTask(audioTask);
+  if (audioTask && (["pending", "running"].includes(audioTask.status) || audioRecovering))
+    return (
+      <SmartEditEmptyEditor
+        onClose={onClose}
+        onConfigure={() => {}}
+        onRetry={() => {}}
+        loading
+        message={
+          audioTask.result?.smartEdit?.stage || "等待音频处理，可退出剪辑，稍后重新打开查看结果"
+        }
+      />
+    );
+  if (plan && timeline && !configure)
+    return (
+      <SmartEditEditor
+        key={`${planTaskId || "draft"}:${selected}`}
+        plan={plan}
+        onChange={changePlan}
+        onClose={onClose}
+        onConfigure={() => setConfigure(true)}
+        onRender={onRender}
+        onCancel={onCancel}
+        canSeparate={Boolean(cap?.capabilities?.separation)}
+        canTranscribe={Boolean(cap?.capabilities?.asr)}
+        processError={
+          audioTask &&
+          (["failed", "cancelled"].includes(audioTask.status) ||
+            (audioTask.status === "interrupted" && !audioRecovering))
+            ? audioTask.error || "音频处理失败，原内容已保留"
+            : undefined
+        }
+        onProcess={(operation, plan, clipId, track, keep) =>
+          onPlan(smartEditRequestSchema.parse({ operation, plan, clipId, track, keep }))
+        }
+        tasks={tasks}
+        canRender={Boolean(cap?.capabilities?.render)}
+        exportState={draft.exportState}
+        onExportState={(exportState) => patch({ exportState })}
+      />
+    );
   return (
     <div
       role="dialog"
@@ -182,6 +311,11 @@ export function SmartEditDialog({
     >
       <header className="flex shrink-0 items-center justify-between border-b border-white/15 px-5 py-3">
         <h2>智能剪辑 · {plan?.title || "从素材到成片"}</h2>
+        {plan && (
+          <button type="button" onClick={() => setConfigure(false)}>
+            返回时间轴
+          </button>
+        )}
         <button type="button" aria-label="关闭智能剪辑" onClick={onClose}>
           <X />
         </button>
@@ -197,6 +331,8 @@ export function SmartEditDialog({
                 value={draft.mode}
                 onChange={(e) => patch({ mode: e.target.value })}
               >
+                <option value="auto">默认模式</option>
+                <option value="sequence">顺序编辑</option>
                 <option value="montage">素材混剪</option>
                 <option value="talk">口播视频</option>
                 <option value="explain">讲解视频</option>
@@ -204,14 +340,15 @@ export function SmartEditDialog({
               </select>
             </label>
             <div>
-              <div className="mb-2 text-sm">素材（{draft.assets.length}/6）</div>
+              <div className="mb-2 text-sm">素材（{draft.assets.length}/12）</div>
               <select
                 aria-label="添加画布视频"
                 className={`${input} w-full`}
                 value=""
                 onChange={(e) => {
                   const asset = sources.find((a) => a.id === e.target.value);
-                  if (asset && draft.assets.length < 6) patch({ assets: [...draft.assets, asset] });
+                  if (asset && draft.assets.length < 12)
+                    patch({ assets: [...draft.assets, asset] });
                 }}
               >
                 <option value="">选择画布已有视频</option>
@@ -234,8 +371,8 @@ export function SmartEditDialog({
                   onChange={async (e) => {
                     const files = Array.from(e.target.files || []);
                     e.target.value = "";
-                    if (files.length + draft.assets.length > 6) {
+                    if (files.length + draft.assets.length > 12) {
-                      setError("最多 6 个素材");
+                      setError("最多 12 个素材");
                       return;
                     }
                     setBusy(true);
@@ -308,7 +445,7 @@ export function SmartEditDialog({
                   className={`${input} w-full`}
                   type="number"
                   min={3}
-                  max={90}
+                  max={300}
                   value={draft.targetDuration}
                   onChange={(e) => patch({ targetDuration: Number(e.target.value) })}
                 />
@@ -321,7 +458,7 @@ export function SmartEditDialog({
                   value={draft.ratio}
                   onChange={(e) => patch({ ratio: e.target.value })}
                 >
-                  {["16:9", "9:16", "1:1"].map((v) => (
+                  {["21:9", "16:9", "9:16", "1:1", "4:3", "3:4"].map((v) => (
                     <option key={v}>{v}</option>
                   ))}
                 </select>
@@ -470,7 +607,9 @@ export function SmartEditDialog({
               </select>
             ) : null}
             {cap &&
-            (!cap.capabilities.scenes || !cap.capabilities.ffmpeg || !cap.capabilities.ffprobe) ? (
+            ((draft.mode !== "sequence" && !cap.capabilities.scenes) ||
+              !cap.capabilities.ffmpeg ||
+              !cap.capabilities.ffprobe) ? (
               <p role="alert" className="text-sm text-amber-200">
                 剪辑处理环境尚未准备好，请配置本地分析服务。
               </p>
@@ -480,7 +619,7 @@ export function SmartEditDialog({
               className={`${button} w-full bg-white/10`}
               disabled={
                 !draft.assets.length ||
-                !cap?.capabilities.scenes ||
+                (draft.mode !== "sequence" && !cap?.capabilities.scenes) ||
                 !cap?.capabilities.ffmpeg ||
                 !cap?.capabilities.ffprobe
               }
@@ -490,6 +629,7 @@ export function SmartEditDialog({
                 setPlaying(false);
                 try {
                   const params = smartEditRequestSchema.parse({ ...draft, operation: "plan" });
+                  if (params.operation === "plan") assertBeatMusic(params);
                   await onPlan(params);
                 } catch (e) {
                   setError((e as Error).message);
@@ -519,7 +659,9 @@ export function SmartEditDialog({
             </p>
           ) : null}
           {cap && !cap.capabilities.render ? (
-            <p className="mb-3 text-sm text-amber-200">视频合成服务暂不可用，剪辑方案仍可保存。</p>
+            <p className="mb-3 text-sm text-amber-200">
+              视频合成服务暂不可用，剪辑方案仍可保存。
+            </p>
           ) : null}
           {plan ? (
             <>
@@ -531,6 +673,7 @@ export function SmartEditDialog({
                     className={`${button} ${selected === i ? "bg-white/15" : ""}`}
                     onClick={() => {
                       setSelected(i);
+                      patch({ selectedPlanIndex: i });
                       setTime(0);
                       setPlaying(false);
                       setClipIndex(0);
@@ -587,7 +730,9 @@ export function SmartEditDialog({
                   <button
                     type="button"
                     className={button}
-                    onClick={() => saveFile("字幕.srt", subtitlesToSrt(compiledSubtitles(plan)))}
+                    onClick={() =>
+                      saveFile("字幕.srt", subtitlesToSrt(compiledSubtitles(plan)))
+                    }
                   >
                     下载字幕
                   </button>
@@ -695,7 +840,9 @@ export function SmartEditDialog({
                       aria-label="剪辑转场"
                       className={input}
                       value={plan.transition}
-                      onChange={(e) => changePlan({ ...plan, transition: Number(e.target.value) })}
+                      onChange={(e) =>
+                        changePlan({ ...plan, transition: Number(e.target.value) })
+                      }
                     >
                       <option value={0}>直接切换</option>
                       <option value={0.25}>叠化 0.25 秒</option>
@@ -774,27 +921,29 @@ export function SmartEditDialog({
                               {i + 1} · {plan.assets.find((a) => a.id === clip.assetId)?.name}
                             </button>
                           </td>
-                          {(["sourceIn", "sourceOut", "speed", "gain"] as const).map((key, j) => (
+                          {(["sourceIn", "sourceOut", "speed", "gain"] as const).map(
-                            <td className="px-2 py-2" key={key}>
+                            (key, j) => (
-                              <input
+                              <td className="px-2 py-2" key={key}>
-                                aria-label={`片段${i + 1}${["入点", "出点", "速度", "音量"][j]}`}
+                                <input
-                                type="number"
+                                  aria-label={`片段${i + 1}${["入点", "出点", "速度", "音量"][j]}`}
-                                className={`${input} w-24`}
+                                  type="number"
-                                min={key === "speed" ? 0.5 : 0}
+                                  className={`${input} w-24`}
-                                max={key === "speed" ? 2 : key === "gain" ? 1 : 180}
+                                  min={key === "speed" ? 0.5 : 0}
-                                step={j < 2 ? 0.1 : 0.1}
+                                  max={key === "speed" ? 2 : key === "gain" ? 1 : 180}
-                                value={j < 2 ? clip[key] / SECOND : clip[key]}
+                                  step={j < 2 ? 0.1 : 0.1}
-                                onChange={(e) =>
+                                  value={j < 2 ? clip[key] / SECOND : clip[key]}
-                                  clipEdit(i, {
+                                  onChange={(e) =>
-                                    [key]:
+                                    clipEdit(i, {
-                                      j < 2
+                                      [key]:
-                                        ? Math.round(Number(e.target.value) * SECOND)
+                                        j < 2
-                                        : Number(e.target.value),
+                                          ? Math.round(Number(e.target.value) * SECOND)
-                                  })
+                                          : Number(e.target.value),
-                                }
+                                    })
-                              />
+                                  }
-                            </td>
+                                />
-                          ))}
+                              </td>
+                            ),
+                          )}
                           <td className="whitespace-nowrap px-2 py-2">
                             {[-1, 1].map((delta) => (
                               <button

components/canvas/smart-edit/SmartEditNode.module.css

@@ -0,0 +1,386 @@
+.card {
+  position: relative;
+  width: 100%;
+  height: 100%;
+  padding: 8px;
+  border: 1px solid #454545;
+  border-radius: 18px;
+  background: #262626;
+  color: #eee;
+  font-family: Arial, "PingFang SC", sans-serif;
+}
+.title {
+  position: absolute;
+  top: -28px;
+  left: 0;
+  display: flex;
+  align-items: center;
+  gap: 6px;
+  font-size: 14px;
+  color: #929292;
+  white-space: nowrap;
+}
+.preview {
+  position: relative;
+  height: calc(100% - 80px);
+  overflow: hidden;
+  border-radius: 10px 10px 0 0;
+  background: #202020;
+}
+.preview > video {
+  width: 100%;
+  height: 100%;
+  object-fit: cover;
+  opacity: 0.35;
+}
+.open {
+  position: absolute;
+  inset: 0;
+  display: flex;
+  flex-direction: column;
+  align-items: center;
+  justify-content: center;
+  gap: 16px;
+}
+.open button {
+  height: 32px;
+  padding: 0 12px;
+  border: 0;
+  border-radius: 8px;
+  background: white;
+  color: #181818;
+  font-size: 16px;
+  cursor: pointer;
+}
+.ruler {
+  height: 26px;
+  display: flex;
+  justify-content: space-around;
+  align-items: center;
+  border-bottom: 1px solid #393939;
+  font-size: 8px;
+  color: #929292;
+}
+.tracks {
+  padding: 8px;
+}
+.tracks > div {
+  display: flex;
+  gap: 3px;
+  height: 20px;
+  margin-bottom: 3px;
+}
+.tracks i {
+  flex: 1;
+  background: #333;
+  border-radius: 3px;
+}
+.composer {
+  position: absolute;
+  z-index: 30;
+  transform: translateX(-50%);
+  width: 660px;
+  max-width: calc(100vw - 32px);
+  padding: 8px;
+  border: 0.5px solid #333;
+  border-radius: 16px;
+  background: #262626;
+  color: #eee;
+  box-shadow: 0 4px 16px #0002;
+  font:
+    14px / 1.5 Arial,
+    "PingFang SC",
+    sans-serif;
+  display: flex;
+  flex-direction: column;
+  gap: 8px;
+}
+.composer button {
+  cursor: pointer;
+  border: 0;
+  color: inherit;
+  font: inherit;
+  background: none;
+  display: flex;
+  align-items: center;
+  justify-content: center;
+  gap: 4px;
+}
+.composer button:disabled {
+  opacity: 0.4;
+  cursor: default;
+}
+.referenceRow {
+  height: 30px;
+  display: flex;
+  justify-content: space-between;
+  padding: 4px 4px 0;
+}
+.referenceRow > button:first-child {
+  font-size: 12px;
+  border-radius: 16px;
+  background: #333;
+  color: #999;
+  padding: 0 6px;
+  height: 26px;
+}
+.referenceRow > button:last-child {
+  color: #999;
+  width: 24px;
+  height: 24px;
+}
+.sources {
+  display: flex;
+  gap: 8px;
+  min-height: 54.8px;
+  flex-wrap: wrap;
+  padding: 0 4px;
+}
+.source {
+  width: 48px;
+  height: 48px;
+  position: relative;
+  border-radius: 8px;
+  background: #191919;
+}
+.source video {
+  width: 100%;
+  height: 100%;
+  object-fit: cover;
+  border-radius: 8px;
+}
+.source b {
+  position: absolute;
+  top: 3px;
+  left: 3px;
+  font-size: 9px;
+  font-weight: 400;
+  background: #0009;
+  border-radius: 8px;
+  padding: 0 3px;
+}
+.source > svg {
+  position: absolute;
+  top: 16px;
+  left: 16px;
+  width: 16px;
+  height: 16px;
+  padding: 3px;
+  border-radius: 50%;
+  background: #0009;
+}
+.source > span {
+  position: absolute;
+  bottom: 2px;
+  left: 3px;
+  font-size: 8px;
+  background: #0009;
+  line-height: 11px;
+}
+.source > button {
+  display: none;
+  position: absolute;
+  top: -4px;
+  right: -4px;
+  width: 15px;
+  height: 15px;
+  border-radius: 50%;
+  background: #666;
+}
+.source:hover > button,
+.source:focus-within > button {
+  display: flex;
+}
+.composer textarea {
+  display: block;
+  width: 100%;
+  resize: none;
+  border: 0;
+  outline: 0;
+  background: transparent;
+  color: #eee;
+  padding: 8px 4px;
+  font:
+    14px / 1.8 Arial,
+    "PingFang SC",
+    sans-serif;
+}
+.composer textarea::placeholder {
+  color: #777;
+}
+.controls {
+  height: 32px;
+  display: flex;
+  gap: 8px;
+}
+.controls .mode,
+.controls .parameters {
+  font-size: 12px;
+  height: 32px;
+  padding: 0 8px;
+  border-radius: 8px;
+  background: #333;
+}
+.controls .mode {
+  background: none;
+}
+.controls .send {
+  margin-left: auto;
+  width: 32px;
+  height: 32px;
+  border-radius: 12px;
+  background: white;
+  color: #222;
+}
+.popup {
+  position: absolute;
+  z-index: 5;
+  bottom: 40px;
+  padding: 12px;
+  background: #262626;
+  border: 0.5px solid #393939;
+  border-radius: 16px;
+  box-shadow: 0 4px 16px #0006;
+  color: #929292;
+}
+.settings {
+  left: 112px;
+  width: 340px;
+}
+.settings label {
+  display: block;
+  font-size: 12px;
+  margin: 0 0 8px;
+}
+.ratios {
+  display: grid;
+  grid-template-columns: repeat(5, 1fr);
+  gap: 8px;
+  margin-bottom: 8px;
+}
+.ratios button {
+  height: 62px;
+  flex-direction: column;
+  gap: 8px;
+  font-size: 12px;
+  border: 1px solid #393939;
+  border-radius: 8px;
+}
+.ratios i {
+  display: block;
+  border: 1.5px solid currentColor;
+  border-radius: 1px;
+}
+.settings button[aria-pressed="true"] {
+  color: white;
+  background: #414141;
+  border-color: #999;
+}
+.duration {
+  display: flex;
+  align-items: center;
+  gap: 8px;
+  margin-bottom: 8px;
+  height: 22px;
+}
+.duration input[type="range"] {
+  min-width: 0;
+  flex: 1;
+  accent-color: white;
+  height: 4px;
+}
+.duration input[type="number"] {
+  width: 48px;
+  height: 22px;
+  appearance: textfield;
+  background: #333;
+  border: 0;
+  outline: 0;
+  border-radius: 6px;
+  color: #eee;
+  text-align: center;
+  font-size: 12px;
+}
+.duration input::-webkit-inner-spin-button {
+  appearance: none;
+}
+.duration > span {
+  font-size: 12px;
+}
+.resolutions {
+  display: flex;
+  gap: 8px;
+}
+.resolutions button {
+  flex: 1;
+  height: 34px;
+  border: 1px solid #393939;
+  border-radius: 10px;
+  font-size: 12px;
+}
+.modeMenu {
+  left: 8px;
+  width: 160px;
+  padding: 6px;
+}
+.modeMenu button {
+  width: 100%;
+  height: 36px;
+  padding: 0 8px;
+  justify-content: space-between;
+  color: #ddd;
+  font-size: 13px;
+  border-radius: 8px;
+}
+.modeMenu button:hover,
+.sourceMenu button:hover {
+  background: #383838;
+}
+.sourceMenu {
+  left: 8px;
+  bottom: calc(100% - 40px);
+  width: 260px;
+  max-height: 320px;
+  overflow: auto;
+}
+.sourceMenu > div,
+.sourceMenu > p {
+  font-size: 12px;
+  padding: 8px 4px;
+}
+.sourceMenu button {
+  width: 100%;
+  height: 36px;
+  justify-content: flex-start;
+  padding: 0 8px;
+  border-radius: 8px;
+  font-size: 13px;
+  color: #ddd;
+}
+.sourceMenu button span {
+  flex: 1;
+  text-overflow: ellipsis;
+  overflow: hidden;
+  white-space: nowrap;
+  text-align: left;
+}
+.error {
+  color: #fca5a5;
+  font-size: 12px;
+  padding: 4px;
+}
+@media (max-width: 500px) {
+  .settings {
+    left: 4px;
+    max-width: calc(100vw - 48px);
+  }
+}
+
+.tracks { max-height: 54px; overflow: hidden; }
+.tracks > div { position: relative; height: 12px; }
+.tracks i { position: absolute; top: 0; bottom: 0; border-right: 2px solid #262626; }
+
+.settings { max-height: min(600px, calc(100vh - 180px)); overflow-y: auto; }
+.settings select { width: 100%; height: 30px; padding: 0 8px; border-radius: 6px; background: #333; color: #eee; border: 0; }
+.musicUpload { display: flex; align-items: center; gap: 4px; margin-top: 8px; font-size: 12px; }
+.settings label.beatToggle { display: flex; gap: 8px; align-items: center; }

components/canvas/smart-edit/SmartEditNode.tsx

@@ -0,0 +1,476 @@
+"use client";
+import { useEffect, useRef, useState } from "react";
+import {
+  ArrowUp,
+  Check,
+  ChevronDown,
+  Maximize2,
+  Play,
+  Plus,
+  Scissors,
+  Upload,
+  X,
+} from "lucide-react";
+import {
+  editRatios,
+  editResolutions,
+  planTimeline,
+  editableCaptions,
+  SECOND,
+  type EditPlan,
+} from "@/lib/video-canvas/smart-edit/contract";
+import { resolveSmartEditIntent, assertBeatMusic } from "@/lib/video-canvas/smart-edit/intent";
+import styles from "./SmartEditNode.module.css";
+export type SmartEditSource = { id: string; name: string; url: string; duration?: number };
+const modes = [
+  { id: "auto", name: "默认模式" },
+  { id: "explain", name: "讲解视频" },
+  { id: "ads", name: "批量广告" },
+  { id: "talk", name: "口播视频" },
+  { id: "montage", name: "素材混剪" },
+];
+export function SmartEditNodeCard({
+  title,
+  previewUrl,
+  plan,
+  onOpen,
+  selected = false,
+}: {
+  title: string;
+  selected?: boolean;
+  previewUrl?: string;
+  plan?: EditPlan;
+  onOpen: () => void;
+}) {
+  let timeline: ReturnType<typeof planTimeline> | undefined;
+  try {
+    if (plan) timeline = planTimeline(plan);
+  } catch {
+    /* Invalid saved drafts open in the editor for recovery. */
+  }
+  const length = timeline?.duration || 0,
+    tick = Math.max(1, Math.ceil(length / SECOND / 5));
+  return (
+    <div className={styles.card} style={{ borderColor: selected ? "#888" : "#454545" }}>
+      <div className={styles.title}>
+        <Play size={12} fill="currentColor" />
+        {title}
+      </div>
+      <div className={styles.preview}>
+        {previewUrl && <video src={`${previewUrl}#t=0.1`} muted preload="metadata" />}
+        <div className={styles.open}>
+          <Scissors size={36} />
+          <button onPointerDown={(e) => e.stopPropagation()} onClick={onOpen}>
+            打开智能剪辑
+          </button>
+        </div>
+      </div>
+      <div className={styles.ruler} aria-label="剪辑时间刻度">
+        {timeline ? (
+          Array.from({ length: Math.floor(length / SECOND / tick) + 1 }, (_, i) => {
+            const seconds = i * tick;
+            return (
+              <span key={i}>
+                {Math.floor(seconds / 60)}:{String(seconds % 60).padStart(2, "0")}
+              </span>
+            );
+          })
+        ) : (
+          <span>尚未生成时间轴</span>
+        )}
+      </div>
+      <div className={styles.tracks} aria-label="剪辑时间轴缩略图">
+        {timeline && (
+          <>
+            <div>
+              {timeline.clips.map((clip) => (
+                <i
+                  key={clip.id}
+                  title={clip.label}
+                  style={{
+                    left: `${(clip.start / length) * 100}%`,
+                    width: `${(clip.duration / length) * 100}%`,
+                  }}
+                />
+              ))}
+            </div>
+            {!!timeline.audioClips.length && (
+              <div>
+                {timeline.audioClips.map((clip) => (
+                  <i
+                    key={clip.id}
+                    style={{
+                      left: `${(clip.start / length) * 100}%`,
+                      width: `${(clip.duration / length) * 100}%`,
+                      background: "#346752",
+                    }}
+                  />
+                ))}
+              </div>
+            )}
+            {plan?.subtitles && (
+              <div>
+                {editableCaptions(plan).map((caption) => (
+                  <i
+                    key={caption.id}
+                    title={caption.text}
+                    style={{
+                      left: `${(caption.start / length) * 100}%`,
+                      width: `${((caption.end - caption.start) / length) * 100}%`,
+                      background: "#75602e",
+                    }}
+                  />
+                ))}
+              </div>
+            )}
+          </>
+        )}
+      </div>
+    </div>
+  );
+}
+export function SmartEditNodeComposer({
+  anchor,
+  draft,
+  sources,
+  availableSources,
+  availableAudios = [],
+  onDraft,
+  onAddSource,
+  onRemoveSource,
+  onUpload,
+  onReference,
+  onSubmit,
+  busy,
+}: {
+  anchor: { centerX: number; bottomY: number; zoom: number };
+  draft: Record<string, any>;
+  sources: SmartEditSource[];
+  availableSources: SmartEditSource[];
+  availableAudios?: { name: string; url: string }[];
+  onDraft: (patch: Record<string, unknown>) => void;
+  onAddSource: (id: string) => void;
+  onRemoveSource: (id: string) => void;
+  onUpload: (files: File[]) => void;
+  onReference?: () => void;
+  onSubmit: () => Promise<unknown>;
+  busy: boolean;
+}) {
+  const [menu, setMenu] = useState<"mode" | "settings" | "sources" | null>(null),
+    [expanded, setExpanded] = useState(false),
+    [error, setError] = useState(""),
+    [submitting, setSubmitting] = useState(false);
+  useEffect(() => setError(""), [draft.music?.url, draft.beatSync]);
+  const root = useRef<HTMLDivElement>(null),
+    file = useRef<HTMLInputElement>(null),
+    musicFile = useRef<HTMLInputElement>(null);
+  useEffect(() => {
+    const close = (e: PointerEvent) => {
+      if (!root.current?.contains(e.target as Node)) setMenu(null);
+    };
+    document.addEventListener("pointerdown", close);
+    return () => document.removeEventListener("pointerdown", close);
+  }, []);
+  const ratio = draft.ratio || "16:9",
+    resolution = draft.resolution || "1080p",
+    duration = draft.targetDuration ?? 20,
+    mode = modes.find((m) => m.id === draft.mode) || modes[0];
+  const toggle = (value: typeof menu) => setMenu(menu === value ? null : value);
+  return (
+    <div
+      ref={root}
+      className={styles.composer}
+      style={{ left: anchor.centerX, top: anchor.bottomY + 16 * anchor.zoom }}
+      onPointerDown={(e) => e.stopPropagation()}
+      onWheel={(e) => e.stopPropagation()}
+      onKeyDown={(e) => e.stopPropagation()}
+    >
+      <div className={styles.referenceRow}>
+        <button
+          onClick={() => (onReference ? onReference() : toggle("sources"))}
+          onContextMenu={(e) => {
+            e.preventDefault();
+            toggle("sources");
+          }}
+        >
+          <Plus size={14} />
+          参考
+        </button>
+        <button
+          aria-label={expanded ? "收起输入框" : "展开输入框"}
+          onClick={() => setExpanded(!expanded)}
+        >
+          <Maximize2 size={16} />
+        </button>
+      </div>
+      {!!sources.length && (
+        <div className={styles.sources}>
+          {sources.map((s, i) => (
+            <div className={styles.source} key={s.id} title={s.name}>
+              <video src={`${s.url}#t=0.1`} preload="metadata" muted />
+              <b>{i + 1}</b>
+              <Play size={12} fill="white" />
+              <span>{s.duration ? `${s.duration.toFixed(1)}s` : "视频"}</span>
+              <button aria-label={`移除 ${s.name}`} onClick={() => onRemoveSource(s.id)}>
+                <X size={10} />
+              </button>
+            </div>
+          ))}
+        </div>
+      )}
+      <input
+        ref={musicFile}
+        hidden
+        type="file"
+        accept="audio/*"
+        onChange={async (e) => {
+          const file = e.target.files?.[0];
+          e.target.value = "";
+          if (!file) return;
+          if (file.size > 125 * 1024 * 1024) {
+            setError("单个素材不得超过 125 MiB");
+            return;
+          }
+          setSubmitting(true);
+          setError("");
+          try {
+            const body = new FormData();
+            body.append("file", file);
+            const response = await fetch("/api/canvas/media/upload", {
+              method: "POST",
+              body,
+              signal: AbortSignal.timeout(120000),
+            });
+            const data = await response.json(),
+              url = data.url || data.data?.url;
+            if (!response.ok || !url) throw Error(data.error || "音乐上传失败");
+            onDraft({ music: { url, gain: 0.2, start: 0 } });
+          } catch (e) {
+            setError(e instanceof Error ? e.message : "音乐上传失败");
+          } finally {
+            setSubmitting(false);
+          }
+        }}
+      />
+      <textarea
+        aria-label="智能剪辑要求"
+        placeholder="描述你想要的剪辑效果"
+        value={draft.prompt || ""}
+        onChange={(e) => onDraft({ prompt: e.target.value })}
+        style={{ height: expanded ? 192 : 96 }}
+      />
+      <div className={styles.controls}>
+        <button className={styles.mode} onClick={() => toggle("mode")}>
+          {mode.name}
+          <ChevronDown size={12} />
+        </button>
+        <button className={styles.parameters} onClick={() => toggle("settings")}>
+          {ratio} · {String(resolution).toUpperCase()} · {duration ? `${duration}s` : "自动"}
+          <ChevronDown size={12} />
+        </button>
+        <button
+          className={styles.send}
+          disabled={busy || submitting || sources.length === 0}
+          aria-label="发送"
+          onClick={async () => {
+            setError("");
+            setSubmitting(true);
+            try {
+              assertBeatMusic(resolveSmartEditIntent(draft));
+              await onSubmit();
+            } catch (e) {
+              setError(e instanceof Error ? e.message : "提交失败");
+            } finally {
+              setSubmitting(false);
+            }
+          }}
+        >
+          <ArrowUp size={20} />
+        </button>
+      </div>
+      {error && (
+        <div role="alert" className={styles.error}>
+          {error}
+        </div>
+      )}
+      {menu === "mode" && (
+        <div className={`${styles.popup} ${styles.modeMenu}`}>
+          {modes.map((m) => (
+            <button
+              key={m.id}
+              onClick={() => {
+                onDraft({ mode: m.id });
+                setMenu(null);
+              }}
+            >
+              {m.name}
+              {m.id === mode.id && <Check size={14} />}
+            </button>
+          ))}
+        </div>
+      )}
+      {menu === "settings" && (
+        <div className={`${styles.popup} ${styles.settings}`}>
+          <label>比例</label>
+          <div className={styles.ratios}>
+            {editRatios.map((r) => {
+              const [w, h] = r.split(":").map(Number);
+              return (
+                <button
+                  key={r}
+                  aria-pressed={r === ratio}
+                  onClick={() => onDraft({ ratio: r })}
+                >
+                  <i
+                    style={{
+                      width: Math.min(16, (16 * w) / h),
+                      height: Math.min(16, (16 * h) / w),
+                    }}
+                  />
+                  {r}
+                </button>
+              );
+            })}
+          </div>
+          <label htmlFor="smart-edit-duration">时长</label>
+          <div className={styles.duration}>
+            <input
+              aria-label="目标时长"
+              type="range"
+              min={0}
+              max={300}
+              step={1}
+              value={duration}
+              onChange={(e) => onDraft({ targetDuration: Number(e.target.value) })}
+            />
+            <input
+              id="smart-edit-duration"
+              type="number"
+              min={0}
+              max={300}
+              value={duration}
+              onChange={(e) =>
+                onDraft({ targetDuration: Math.max(0, Math.min(300, Number(e.target.value))) })
+              }
+            />
+            <span>s</span>
+          </div>
+          {mode.id === "ads" && (
+            <>
+              <label>广告版本</label>
+              <div className={styles.resolutions}>
+                {[1, 2, 3].map((count) => (
+                  <button
+                    key={count}
+                    aria-pressed={(draft.variants ?? 3) === count}
+                    onClick={() => onDraft({ variants: count })}
+                  >
+                    {count} 版
+                  </button>
+                ))}
+              </div>
+            </>
+          )}
+          <label htmlFor="smart-edit-music">背景音乐</label>
+          <select
+            id="smart-edit-music"
+            value={draft.music?.url || ""}
+            onChange={(e) =>
+              onDraft({
+                music: e.target.value
+                  ? { url: e.target.value, gain: 0.2, start: 0 }
+                  : undefined,
+              })
+            }
+          >
+            <option value="">未选择</option>
+            {draft.music?.url && !availableAudios.some((a) => a.url === draft.music.url) && (
+              <option value={draft.music.url}>已上传音乐</option>
+            )}
+            {availableAudios.map((audio) => (
+              <option key={audio.url} value={audio.url}>
+                {audio.name}
+              </option>
+            ))}
+          </select>
+          <button
+            className={styles.musicUpload}
+            disabled={submitting}
+            onClick={() => musicFile.current?.click()}
+          >
+            <Upload size={12} />
+            上传音乐
+          </button>
+          <label className={styles.beatToggle}>
+            <input
+              type="checkbox"
+              checked={resolveSmartEditIntent(draft).beatSync}
+              onChange={(e) => onDraft({ beatSync: e.target.checked })}
+            />
+            音乐卡点
+          </label>
+          {resolveSmartEditIntent(draft).beatSync && (
+            <div className={styles.resolutions}>
+              {["2", "4"].map((beats) => (
+                <button
+                  key={beats}
+                  aria-pressed={(draft.beatsPerCut || "4") === beats}
+                  onClick={() => onDraft({ beatsPerCut: beats })}
+                >
+                  {beats} 拍切换
+                </button>
+              ))}
+            </div>
+          )}
+          <label>清晰度</label>
+          <div className={styles.resolutions}>
+            {editResolutions.map((r) => (
+              <button
+                key={r}
+                aria-pressed={r === resolution}
+                onClick={() => onDraft({ resolution: r })}
+              >
+                {r.toUpperCase()}
+              </button>
+            ))}
+          </div>
+        </div>
+      )}
+      {menu === "sources" && (
+        <div className={`${styles.popup} ${styles.sourceMenu}`}>
+          <button onClick={() => file.current?.click()} disabled={sources.length >= 12}>
+            <Upload size={14} />
+            上传视频
+          </button>
+          <div>画布素材 · {sources.length}/12</div>
+          {availableSources.map((s) => (
+            <button
+              key={s.id}
+              disabled={sources.length >= 12 && !sources.some((a) => a.id === s.id)}
+              onClick={() =>
+                sources.some((a) => a.id === s.id) ? onRemoveSource(s.id) : onAddSource(s.id)
+              }
+            >
+              <Play size={14} />
+              <span>{s.name}</span>
+              {sources.some((a) => a.id === s.id) && <Check size={14} />}
+            </button>
+          ))}
+          {!availableSources.length && <p>上传视频或连接画布视频素材</p>}
+        </div>
+      )}
+      <input
+        ref={file}
+        type="file"
+        hidden
+        accept="video/*"
+        multiple
+        onChange={(e) => {
+          onUpload(Array.from(e.target.files || []).slice(0, 12 - sources.length));
+          e.target.value = "";
+          setMenu(null);
+        }}
+      />
+    </div>
+  );
+}

components/canvas/smart-edit/TransitionPreview.tsx

@@ -0,0 +1,75 @@
+"use client";
+import { useEffect, useMemo, useState } from "react";
+import { type EditPlan, planTimeline, SECOND } from "@/lib/video-canvas/smart-edit/contract";
+import { EditPreview } from "./EditPreview";
+
+/** Preview the selected edit boundary with the same timing/compositing used by the editor. */
+export function TransitionPreview({
+  plan,
+  index,
+  cover = false,
+}: {
+  plan: EditPlan;
+  index: number;
+  cover?: boolean;
+}) {
+  const preview = useMemo(
+    () => ({
+      ...plan,
+      subtitles: false,
+      music: undefined,
+      audioClips: [],
+      clips: plan.clips
+        .slice(index - 1, index + 1)
+        .map((c) => ({ ...c, gain: 0, audioDetached: true })),
+    }),
+    [plan, index],
+  );
+  const timeline = useMemo(() => planTimeline(preview), [preview]);
+  const boundary = timeline.clips[1];
+  const start = Math.max(0, boundary.transitionStart / SECOND - 0.25);
+  const end = Math.min(
+    timeline.duration / SECOND,
+    (boundary.transitionStart + boundary.overlap) / SECOND + 0.25,
+  );
+  const [time, setTime] = useState(start),
+    [error, setError] = useState(""),
+    [visible, setVisible] = useState(true);
+  useEffect(() => {
+    const change = () => setVisible(!document.hidden);
+    document.addEventListener("visibilitychange", change);
+    return () => document.removeEventListener("visibilitychange", change);
+  }, []);
+  useEffect(() => {
+    let request = 0,
+      epoch = 0;
+    setError("");
+    setTime(start);
+    const tick = (now: number) => {
+      if (document.hidden) epoch = 0;
+      else {
+        if (!epoch) epoch = now;
+        setTime(start + (((now - epoch) / 1000) % (end - start)));
+      }
+      request = requestAnimationFrame(tick);
+    };
+    request = requestAnimationFrame(tick);
+    return () => cancelAnimationFrame(request);
+  }, [start, end, preview]);
+  return (
+    <div className="se-transition-preview" aria-label="当前片段转场动态预览">
+      <EditPreview
+        plan={preview}
+        time={time}
+        playing={visible}
+        externalClock
+        onTime={setTime}
+        onPlaying={() => {}}
+        onError={setError}
+        fit
+        cover={cover}
+      />
+      {error && <div className="se-transition-preview-error">{error}</div>}
+    </div>
+  );
+}

components/canvas/smart-edit/usePreviewGain.ts

@@ -0,0 +1,111 @@
+"use client";
+import { useEffect, useRef, useState } from "react";
+import type { EditPlan } from "@/lib/video-canvas/smart-edit/contract";
+
+/** Only amplified cross-origin sources need a same-origin copy for Web Audio. */
+export function usePreviewGain(plan: EditPlan, onError: (message: string) => void) {
+  const remote = [
+    ...new Set(
+      [...plan.clips, ...(plan.audioClips || [])]
+        .filter((clip) => clip.gain > 1)
+        .map((clip) => plan.assets.find((asset) => asset.id === clip.assetId)!.url)
+        .filter(
+          (url) =>
+            typeof window !== "undefined" &&
+            new URL(url, window.location.href).origin !== window.location.origin,
+        ),
+    ),
+  ].sort();
+  const key = JSON.stringify(remote);
+  const [copies, setCopies] = useState<{ key: string; urls: Record<string, string> }>({
+    key: "",
+    urls: {},
+  });
+  const [attempt, setAttempt] = useState(0),
+    [failed, setFailed] = useState(false);
+  const error = useRef(onError);
+  error.current = onError;
+  const graph = useRef<{
+    context: AudioContext;
+    nodes: Map<HTMLMediaElement, { source: MediaElementAudioSourceNode; gain: GainNode }>;
+  }>();
+  useEffect(() => {
+    setFailed(false);
+    const controller = new AbortController(),
+      urls: Record<string, string> = {};
+    void Promise.all(
+      (JSON.parse(key) as string[]).map(async (url) => {
+        const response = await fetch("/api/canvas/download-image", {
+          method: "POST",
+          headers: { "Content-Type": "application/json" },
+          body: JSON.stringify({ url, name: "剪辑音量预览" }),
+          signal: controller.signal,
+        });
+        if (!response.ok) throw Error("无法读取音量预览素材，请检查链接后重试");
+        const blob = await response.blob();
+        if (!blob.size) throw Error("音量预览素材为空");
+        if (!controller.signal.aborted) urls[url] = URL.createObjectURL(blob);
+      }),
+    )
+      .then(() => {
+        if (!controller.signal.aborted) setCopies({ key, urls });
+      })
+      .catch((e) => {
+        if (!controller.signal.aborted) {
+          setFailed(true);
+          error.current(`${e.message}，可点击播放重试`);
+        }
+      });
+    return () => {
+      controller.abort();
+      Object.values(urls).forEach(URL.revokeObjectURL);
+    };
+  }, [key, attempt]);
+  useEffect(
+    () => () => {
+      void graph.current?.context.close();
+      graph.current = undefined;
+    },
+    [],
+  );
+  return {
+    retry() {
+      if (failed) setAttempt((value) => value + 1);
+    },
+    source(url: string) {
+      return copies.key === key ? copies.urls[url] || url : url;
+    },
+    ready(url: string) {
+      return !remote.includes(url) || (copies.key === key && Boolean(copies.urls[url]));
+    },
+    retain(elements: HTMLMediaElement[]) {
+      for (const [element, node] of graph.current?.nodes || [])
+        if (!elements.includes(element)) {
+          element.pause();
+          node.source.disconnect();
+          node.gain.disconnect();
+          graph.current!.nodes.delete(element);
+        }
+    },
+    volume(element: HTMLMediaElement, gain: number, playing: boolean) {
+      let node = graph.current?.nodes.get(element);
+      if (gain > 1 && !node) {
+        graph.current ||= { context: new AudioContext(), nodes: new Map() };
+        const source = graph.current.context.createMediaElementSource(element),
+          amplifier = graph.current.context.createGain();
+        source.connect(amplifier);
+        amplifier.connect(graph.current.context.destination);
+        node = { source, gain: amplifier };
+        graph.current.nodes.set(element, node);
+      }
+      if (node) {
+        element.volume = 1;
+        node.gain.gain.value = Math.max(0, gain);
+        if (playing && graph.current?.context.state === "suspended")
+          void graph.current.context
+            .resume()
+            .catch(() => error.current("请再次点击播放以启用音频预览"));
+      } else element.volume = Math.min(1, Math.max(0, gain));
+    },
+  };
+}

docs/smart-edit-completion-2026-09-18.md

@@ -0,0 +1,63 @@
+# 智能剪辑四项续修与本地验收（2026-09-18）
+
+本轮承接 `e18413a4`，范围为人声分离、自动转写、完整字体目录、主站智能剪辑链路。仅本地修改和验证，未部署、未推送。并非全部网站功能或所有视觉状态完成像素级验收。
+
+## 功能结果
+
+| 项目 | 实现与验证 |
+| --- | --- |
+| 人声分离 | 离线 HTDemucs，CPU 2线程、7秒分块、每素材180秒上限。视频和独立音轨均可选择“仅保留人声”“仅保留背景音”“分离为两个音轨”。保留裁剪、变速、增益、静音和转场音量包络，关闭已分离的视频原声，防止音量叠加；处理失败保留原方案。 |
+| 自动转写 | 本地 FunASR Paraformer/FSMN-VAD/CT-Transformer。字幕面板入口直接创建持久化任务，保留手动字幕、替换此前 ASR 字幕。修复缺失 `pred_timestamp` 导致没有逐句时间戳的问题，负起点归零，按照裁剪及变速映射到时间轴。 |
+| 完整字体库 | 参考目录共2002项，5项随应用打包、1997项新增目录全部下载、解析及SHA校验通过；198个参考源字体，其他字体来自官方 Google Fonts。原始与处理后字节均固定校验和。浏览器和合成使用相同字体文件；字体内版权信息保留。支持搜索、全部/常用/推荐/Google分类、虚拟滚动。修复小高度窗口弹层被参数栏截断。 |
+| 主站智能剪辑链路 | 实际 Next 生产构建、登录、隔离数据库、上传、任务队列、模型处理、FFmpeg合成、成片与SRT、画布回填播放、编辑方案保存和刷新恢复均验证。未使用独立演示页的样例任务回调。 |
+
+模型官方来源：[Demucs](https://github.com/facebookresearch/demucs)、[FunASR](https://github.com/modelscope/FunASR)。部署安装及缓存配置见 [服务说明](../services/smart-edit/README.md)。
+
+## 真实数据
+
+中文样例8.3935秒，视频8.9秒。实际识别三句话：
+
+- 0–1.349秒：“欢迎使用智能剪辑，”
+- 1.349–4.710秒：“我们正在测试人声分离和自动字幕，”
+- 4.710–8.205秒：“请保留原始视频的画面和声音。”
+
+实际混音分离后，人声及背景声相加与原信号的最大误差 `1.49e-8`。这验证音量与时间同步；不等于对任意歌曲的人声提取质量作保证。
+
+生产构建 API 链路：方案1.6秒、视频分离10.6秒、转写10.6秒、独立音轨分离9.1秒、合成1.6秒。两个单轨保留分支另测11.6/10.2秒，均只新增所选的一轨。此为本机短素材实测，不是线上性能承诺。成片854×480、H.264/AAC、8.900秒。
+
+新增字体1997项全部处理通过，缓存793,509,828字节（约757MiB）。实际FFmpeg验证原有5种及新增6种（大宝桃桃体、ABeeZee、Abril Fatface、CJGDGMH、Molle、SCJSSHT），检查实际字形像素。主站实际选择大宝桃桃体，应用全部字幕并确认数据库保存为所选字体。两款参考字体OS/2版本字段与长度不匹配，仅修正字段版本，未改变字形轮廓。
+
+## 发现并修复的主站问题
+
+- 新画布依赖数据库 `now()` 写入无时区时间时，与应用生成的任务时间产生8小时时差，导致任务回填被判为旧结果。新建画布的两个接口改为显式应用时间；隔离测试库为 Asia/Shanghai，新增项目时间断言通过。未迁移或修改生产历史数据。
+- 视频卡片原先忽略 `videoMeta.durationSeconds`，显示5秒而实际8.9秒。优先读取成片元数据及浏览器解码时长，现显示9秒。
+- 音频处理结果不得把编辑器节点变成视频节点；更新当前方案时保留其他备选方案，反复回填保持幂等。
+- 开发热更新期间曾遇到编辑器样式失效；最终生产构建重新加载、播放和操作正常。未将开发热更新现象误判为已定位的线上缺陷。
+
+## 验证与复现
+
+- TypeScript及Next生产构建通过。
+- Helper回归498通过、1跳过。
+- 真实FFmpeg：6比例、3转场、居中转场、裁剪/变速/音量/静音/字幕位置与字体；PySceneDetect 3镜头、RMS 2有效区间、librosa 22拍。
+- 独立合成服务：鉴权、真实成片、进程中断后恢复、幂等、冲突拒绝、完成结果重启保留。
+- 主站浏览器：实际自动转写、字体搜索/选择/全部应用、两个分离子菜单、仅人声提交、刷新后回填、成片播放和导出到画布。
+
+可复现脚本：
+
+```sh
+$SMART_EDIT_PYTHON scripts/qa/smart-edit-models.py input.wav .qa/model-check
+SMART_EDIT_TEST_CATALOG_FONTS=1 node scripts/qa/video-canvas-smart-edit.cjs
+SMART_EDIT_QA_DIR=.qa/isolated-main-chain node scripts/qa/smart-edit-main-chain.cjs
+node services/video-render/build.cjs
+node scripts/qa/video-render-service.cjs
+```
+
+主站脚本仅允许localhost/127.0.0.1，需独立测试库、实际登录账号 `account.json`（id/email/password）和含“智能剪辑”的中文 `speech.mp4`。不得使用生产账号或生产库。数据库、模型权重、字体缓存、测试凭据和成片不提交Git。
+
+本地证据目录 `.qa/smart-edit/remaining-0918/`：`model-regression/report.json`、`production-chain.log`、`production-chain/main-result.mp4`、`branches-report.json`、`catalog-render.log`、`browser-verification.json`、`main-editor-final.jpg`、`browser-export.mp4`、`build-release.log`、`helpers-final.log`、`render-service.log`。
+
+## 验收边界
+
+本地主站使用隔离 PostgreSQL 数据库及本地S3兼容服务。测试进程注入仅限本地下载传输的映射，使真实下载器可访问该存储；没有放宽生产SSRF规则。尚未验证线上对象存储或部署。
+
+目录及字体文件完整，不代表2002字体在所有DPR、中文混排、字形抗锯齿方面都完成逐像素对拍。后台休眠/断网/所有窄屏矩阵、运行中远程任务主动取消、五种模式算法与参考逐案一致性仍单列待验，不混入本轮完成项。

docs/smart-edit-fixes-2026-09-18.json

@@ -0,0 +1,777 @@
+{
+  "scope": "智能剪辑节点、编辑器及直接依赖的音频/导出链路；不是全站所有模块",
+  "baseline": "68f62dae",
+  "local_only": true,
+  "counts": {
+    "通过": 39,
+    "待验": 6,
+    "代码核对": 17,
+    "已修复": 28,
+    "部分修复": 4,
+    "既有证据": 1
+  },
+  "items": [
+    {
+      "id": "SE-001",
+      "group": "节点与输入",
+      "item": "打开剪辑",
+      "status": "通过",
+      "detail": "参考和本地均有入口；本地进入真实编辑组件",
+      "evidence": "本轮UI"
+    },
+    {
+      "id": "SE-002",
+      "group": "节点与输入",
+      "item": "五种模式名称和顺序",
+      "status": "通过",
+      "detail": "默认、讲解、批量广告、口播、素材混剪相同",
+      "evidence": "参考node-modes截图/SmartEditNode"
+    },
+    {
+      "id": "SE-003",
+      "group": "节点与输入",
+      "item": "五种模式算法效果",
+      "status": "待验",
+      "detail": "本地有模式分支；没有重新提交参考站付费生成，不能认定效果等同",
+      "evidence": "execute/contract"
+    },
+    {
+      "id": "SE-004",
+      "group": "节点与输入",
+      "item": "六种比例",
+      "status": "通过",
+      "detail": "21:9、16:9、9:16、1:1、4:3、3:4；实际合成覆盖",
+      "evidence": "参考node-settings/本轮FFmpeg"
+    },
+    {
+      "id": "SE-005",
+      "group": "节点与输入",
+      "item": "三种清晰度",
+      "status": "通过",
+      "detail": "480P、720P、1080P；控件一致",
+      "evidence": "参考node-settings/SmartEditNode"
+    },
+    {
+      "id": "SE-006",
+      "group": "节点与输入",
+      "item": "时长范围",
+      "status": "通过",
+      "detail": "0–300秒、步长1秒；本地0为自动",
+      "evidence": "参考可见DOM/contract"
+    },
+    {
+      "id": "SE-007",
+      "group": "节点与输入",
+      "item": "提示词编辑与展开",
+      "status": "代码核对",
+      "detail": "本地受控输入、展开及提交；参考既有提示词未改动",
+      "evidence": "SmartEditNode"
+    },
+    {
+      "id": "SE-008",
+      "group": "节点与输入",
+      "item": "引用画布素材",
+      "status": "代码核对",
+      "detail": "参考进入引用选择；生产组件onReference连接连线状态，本地独立测试页只是候选列表",
+      "evidence": "参考引用模式/主画布调用点"
+    },
+    {
+      "id": "SE-009",
+      "group": "节点与输入",
+      "item": "素材序号、时长、移除",
+      "status": "代码核对",
+      "detail": "本地有缩略图、序号、时长及移除；未删除参考原素材",
+      "evidence": "SmartEditNode"
+    },
+    {
+      "id": "SE-010",
+      "group": "节点与输入",
+      "item": "本地上传与数量限制",
+      "status": "代码核对",
+      "detail": "真实上传入口，上限12个素材、单个125MiB；参考上限尚未验证",
+      "evidence": "SmartEditNode/SmartEditDialog"
+    },
+    {
+      "id": "SE-011",
+      "group": "节点与输入",
+      "item": "节点缩略时间轴",
+      "status": "已修复",
+      "detail": "节点刻度、视频、字幕、独立音轨从实际计划生成；重新生成后按任务版本选择新计划，保留所选版本。",
+      "evidence": "本轮节点UI/SmartEditNodeCard"
+    },
+    {
+      "id": "SE-012",
+      "group": "节点与输入",
+      "item": "批量广告多版本",
+      "status": "部分修复",
+      "detail": "节点新增1–3版选择，未指定时广告默认3版；实际仍是候选顺序变体，未认定为参考站的多创意算法。",
+      "evidence": "本轮UI/intent helpers"
+    },
+    {
+      "id": "SE-013",
+      "group": "节点与输入",
+      "item": "提示词要求卡点",
+      "status": "已修复",
+      "detail": "节点提示词的常见“卡点/beat sync”表达关联开关，显式关闭优先；支持画布音乐选择/上传及2或4拍切换，缺音乐阻止提交。并非通用自然语言意图解析。",
+      "evidence": "本轮UI/intent helpers/真实librosa 22拍"
+    },
+    {
+      "id": "SE-014",
+      "group": "布局与操作",
+      "item": "顶栏高度和边距",
+      "status": "通过",
+      "detail": "48px顶栏、8px页面边距；本轮图像对照",
+      "evidence": "参考DOM/本地CSS"
+    },
+    {
+      "id": "SE-015",
+      "group": "布局与操作",
+      "item": "左右参数栏宽度",
+      "status": "通过",
+      "detail": "转场库372px、参数栏300px，均可拖动调整",
+      "evidence": "参考DOM/本地CSS"
+    },
+    {
+      "id": "SE-016",
+      "group": "布局与操作",
+      "item": "面板背景与关闭按钮",
+      "status": "已修复",
+      "detail": "背景统一#202020，边框白色10%，关闭按钮24px，标题500/14px",
+      "evidence": "参考transition.json/本地CSS"
+    },
+    {
+      "id": "SE-017",
+      "group": "布局与操作",
+      "item": "默认时间轴高度",
+      "status": "已修复",
+      "detail": "保留打开时的像素高度，窗口缩小时仅为可用空间收缩，不再每次重算35vh。",
+      "evidence": "resize代码"
+    },
+    {
+      "id": "SE-018",
+      "group": "布局与操作",
+      "item": "折叠预览",
+      "status": "已修复",
+      "detail": "收起后背景透明、可操作下方画布，仅底部保留时间轴；预览组件保留，实测能播放到片尾。独立页面下层为测试按钮，主画布接入仍需整体验收。",
+      "evidence": "本轮本地UI"
+    },
+    {
+      "id": "SE-019",
+      "group": "布局与操作",
+      "item": "全屏预览",
+      "status": "代码核对",
+      "detail": "本地调用Fullscreen API并显示拒绝错误；本轮未重测不同浏览器权限状态",
+      "evidence": "SmartEditEditor"
+    },
+    {
+      "id": "SE-020",
+      "group": "布局与操作",
+      "item": "撤销重做",
+      "status": "通过",
+      "detail": "本轮分离音轨后撤销成功；参考宽屏有按钮、窄屏隐藏，先前未见不代表没有",
+      "evidence": "本轮UI/参考1512宽度"
+    },
+    {
+      "id": "SE-021",
+      "group": "布局与操作",
+      "item": "播放、暂停、逐帧",
+      "status": "代码核对",
+      "detail": "Space与左右键、30fps步进、片尾停止；前批浏览器已验证，本轮保留代码核对口径",
+      "evidence": "SmartEditEditor/EditPreview"
+    },
+    {
+      "id": "SE-022",
+      "group": "布局与操作",
+      "item": "吸附默认值",
+      "status": "通过",
+      "detail": "参考关闭，本地SmartEditEditor也是false；纠正旧文档“本地开启”误记",
+      "evidence": "参考AX/SmartEditEditor:state"
+    },
+    {
+      "id": "SE-023",
+      "group": "布局与操作",
+      "item": "退出与Esc",
+      "status": "已修复",
+      "detail": "Esc退出整个编辑器，输入框编辑时不误退出；收起时保留关闭按钮。",
+      "evidence": "本轮本地UI"
+    },
+    {
+      "id": "SE-024",
+      "group": "布局与操作",
+      "item": "窄屏可用性",
+      "status": "待验",
+      "detail": "参考1024宽度测得左右仍372/300；本地三栏与工具条更窄窗口尚无完整可用性验收",
+      "evidence": "reference-transition-1024"
+    },
+    {
+      "id": "SE-025",
+      "group": "视频片段",
+      "item": "四角缩放、移动与重置",
+      "status": "通过",
+      "detail": "本轮真实合成像素验证；本地有选框和重置，参考四角手柄可见",
+      "evidence": "本轮FFmpeg/参考video"
+    },
+    {
+      "id": "SE-026",
+      "group": "视频片段",
+      "item": "变速上下调步长",
+      "status": "已修复",
+      "detail": "参考1.00→1.01，本地0.05改为0.01；浏览器已复测",
+      "evidence": "本轮UI"
+    },
+    {
+      "id": "SE-027",
+      "group": "视频片段",
+      "item": "变速支持范围",
+      "status": "待验",
+      "detail": "本地0.5–2倍；参考未穷尽上下限，不能因步长一致声称范围一致",
+      "evidence": "contract/参考video"
+    },
+    {
+      "id": "SE-028",
+      "group": "视频片段",
+      "item": "时长输入",
+      "status": "代码核对",
+      "detail": "按原素材剩余长度和变速约束，不能超过素材",
+      "evidence": "SmartEditEditor/contract"
+    },
+    {
+      "id": "SE-029",
+      "group": "视频片段",
+      "item": "数值越界后的显示",
+      "status": "已修复",
+      "detail": "到上限2后再输入99，显示回到2.00，不残留非法数字",
+      "evidence": "本轮UI"
+    },
+    {
+      "id": "SE-030",
+      "group": "视频片段",
+      "item": "增益步长",
+      "status": "已修复",
+      "detail": "参考0.5dB，本地由1改0.5；浏览器0→-0.5确认",
+      "evidence": "本轮UI"
+    },
+    {
+      "id": "SE-031",
+      "group": "视频片段",
+      "item": "正增益",
+      "status": "已修复",
+      "detail": "本地支持最高+12dB，预览使用GainNode，导出实际音量放大；参考上限尚未确认。跨域正增益素材经登录下载接口取得同源副本，需另测真实主站链路。",
+      "evidence": "本轮浏览器+0.5dB/FFmpeg约2倍振幅"
+    },
+    {
+      "id": "SE-032",
+      "group": "视频片段",
+      "item": "边缘裁剪和排序",
+      "status": "代码核对",
+      "detail": "已有按source时间/速度换算和撤销；本轮未重新拖遍所有方向",
+      "evidence": "SmartEditEditor"
+    },
+    {
+      "id": "SE-033",
+      "group": "视频片段",
+      "item": "分割和左右裁剪",
+      "status": "通过",
+      "detail": "边界禁用、源范围及片长合同回归通过；本轮未对拍参考破坏性编辑",
+      "evidence": "本轮helpers/SmartEditEditor"
+    },
+    {
+      "id": "SE-034",
+      "group": "视频片段",
+      "item": "至少保留一个片段",
+      "status": "待验",
+      "detail": "本地阻止删最后一段；参考空时间轴行为未删除原素材验证",
+      "evidence": "SmartEditEditor"
+    },
+    {
+      "id": "SE-035",
+      "group": "视频片段",
+      "item": "右键菜单",
+      "status": "部分修复",
+      "detail": "已增加人声分离子菜单及仅保留人声、仅保留背景音、两轨输出；菜单其他文案与参考仍有差异。",
+      "evidence": "本轮主站UI/真实HTDemucs"
+    },
+    {
+      "id": "SE-036",
+      "group": "视频片段",
+      "item": "转场覆盖区右键",
+      "status": "已修复",
+      "detail": "转场覆盖区域按指针时间选择底层片段并打开视频右键菜单。",
+      "evidence": "本轮浏览器：覆盖框右键选中有声片段并成功分离"
+    },
+    {
+      "id": "SE-037",
+      "group": "视频片段",
+      "item": "静音后恢复原增益",
+      "status": "已修复",
+      "detail": "新增可选muted独立于gain保存，原声和独立音轨恢复不会覆盖原增益；分离时继承静音状态。",
+      "evidence": "浏览器原声/独立轨+0.5dB静音恢复；真实FFmpeg两类静音"
+    },
+    {
+      "id": "SE-038",
+      "group": "字幕",
+      "item": "空列表、新建、搜索",
+      "status": "通过",
+      "detail": "参考空列表和新建已操作；本地新建成功、搜索有真实过滤",
+      "evidence": "本轮UI/源码"
+    },
+    {
+      "id": "SE-039",
+      "group": "字幕",
+      "item": "列表逐条新增与删除",
+      "status": "已修复",
+      "detail": "每行新增在后面插入/删除，插入按时间排序、遇下条字幕或片尾缩短，并定位新字幕；片尾不能插入时给提示。",
+      "evidence": "本轮UI/代码"
+    },
+    {
+      "id": "SE-040",
+      "group": "字幕",
+      "item": "列表时间格式",
+      "status": "已修复",
+      "detail": "时间框采用MM:SS.xx，可输入小数秒；失焦校验范围，非法值还原，Esc取消编辑。参考只展示整秒的像素细节尚未作为黄金图验收。",
+      "evidence": "本轮UI"
+    },
+    {
+      "id": "SE-041",
+      "group": "字幕",
+      "item": "选中文本、编辑与位置",
+      "status": "通过",
+      "detail": "本轮本地样式编辑和实际字幕位置像素测试通过",
+      "evidence": "本轮UI/FFmpeg"
+    },
+    {
+      "id": "SE-042",
+      "group": "字幕",
+      "item": "显示和隐藏",
+      "status": "代码核对",
+      "detail": "subtitles控制预览和导出；关闭面板仅隐藏编辑轨道",
+      "evidence": "SmartEditEditor/compiledSubtitles"
+    },
+    {
+      "id": "SE-043",
+      "group": "字幕",
+      "item": "字体搜索和长列表",
+      "status": "通过",
+      "detail": "参考目录2002项完整收录，1997新增文件全部下载解析与双SHA校验；搜索、四分类、虚拟滚动、按需加载可用。",
+      "evidence": "remaining-0918/reference-fonts.json、prepare-catalog.log；主站字体切换"
+    },
+    {
+      "id": "SE-044",
+      "group": "字幕",
+      "item": "字体分类",
+      "status": "已修复",
+      "detail": "新增全部/常用/推荐/Google四类筛选；范围限已打包的实际字体。",
+      "evidence": "本轮UI/CaptionFonts"
+    },
+    {
+      "id": "SE-045",
+      "group": "字幕",
+      "item": "预览与导出的字形",
+      "status": "部分修复",
+      "detail": "2002字体目录已齐；预览与FFmpeg使用同一固定文件，5原有+6新增字体实际合成通过；未对2002种字体逐字形、行高、DPR作黄金图对拍。",
+      "evidence": "catalog-render.log/主站大宝桃桃体切换"
+    },
+    {
+      "id": "SE-046",
+      "group": "字幕",
+      "item": "字号范围和显示单位",
+      "status": "已修复",
+      "detail": "从参考可见iframe确认1920×1080、56px；新字幕明确1080基准，默认56/12–72px；旧720基准按比例显示并保留历史超范围值。",
+      "evidence": "本轮参考DOM/legacy兼容helpers/本地UI"
+    },
+    {
+      "id": "SE-047",
+      "group": "字幕",
+      "item": "字间距范围和控件",
+      "status": "已修复",
+      "detail": "字间距改X数值框0–100；旧工程负值保留，预览与ASS依各自基准缩放。",
+      "evidence": "StyleNumber/contract"
+    },
+    {
+      "id": "SE-048",
+      "group": "字幕",
+      "item": "字体颜色和十六进制",
+      "status": "代码核对",
+      "detail": "两处均有取色和hex；本地失焦校验非法值；样式烧录已回归",
+      "evidence": "StyleColor/FFmpeg"
+    },
+    {
+      "id": "SE-049",
+      "group": "字幕",
+      "item": "六种预设入口",
+      "status": "通过",
+      "detail": "默认、白黑、黑白、黄黑、白粉、粉黑均有",
+      "evidence": "本轮参考及本地UI"
+    },
+    {
+      "id": "SE-050",
+      "group": "字幕",
+      "item": "预设按钮外观",
+      "status": "已修复",
+      "detail": "默认预设显示“默认”，其他为T；选中描边按实际样式匹配。",
+      "evidence": "本轮字幕UI"
+    },
+    {
+      "id": "SE-051",
+      "group": "字幕",
+      "item": "描边颜色和粗细",
+      "status": "已修复",
+      "detail": "新字幕默认白色描边，粗细0–20；旧字幕不变。预览和ASS同步缩放。",
+      "evidence": "本轮UI/contract/FFmpeg"
+    },
+    {
+      "id": "SE-052",
+      "group": "字幕",
+      "item": "描边不透明度语义",
+      "status": "已修复",
+      "detail": "独立outlineOpacity；0时保留文字，只去描边；旧整行opacity保持兼容",
+      "evidence": "UI截图/ASS合同/实际像素测试"
+    },
+    {
+      "id": "SE-053",
+      "group": "字幕",
+      "item": "字号等数值单位装饰",
+      "status": "已修复",
+      "detail": "增加px/%单位和上下步进按钮；数值失焦校验，不提交编辑中的空值。",
+      "evidence": "StyleNumber/本地构建"
+    },
+    {
+      "id": "SE-054",
+      "group": "字幕",
+      "item": "应用全部",
+      "status": "代码核对",
+      "detail": "复制完整style含新outlineOpacity；本轮未覆盖所有字幕数量组合",
+      "evidence": "SmartEditEditor"
+    },
+    {
+      "id": "SE-055",
+      "group": "字幕",
+      "item": "自动字幕与长转场",
+      "status": "已修复",
+      "detail": "居中转场不再把后片段字幕向后延迟整个转场时长，避免片头字幕丢失",
+      "evidence": "新增合同回归"
+    },
+    {
+      "id": "SE-056",
+      "group": "字幕",
+      "item": "手工字幕与SRT",
+      "status": "通过",
+      "detail": "手工字幕保留位置和样式，SRT仅保留文本时间；ASS字符转义回归通过",
+      "evidence": "本轮helpers/FFmpeg"
+    },
+    {
+      "id": "SE-057",
+      "group": "字幕",
+      "item": "语音自动转写",
+      "status": "通过",
+      "detail": "真实FunASR转写成功；修复缺少pred_timestamp造成整段无标点时间戳的问题，限制负起点并映射剪辑速度/裁剪。字幕按钮提交主站任务，完成后刷新恢复。",
+      "evidence": "model-regression/report.json；production-chain.log；主站浏览器"
+    },
+    {
+      "id": "SE-058",
+      "group": "转场",
+      "item": "三种基础转场入口",
+      "status": "通过",
+      "detail": "淡入淡出、黑场过渡、白场过渡；这不是全站视频生成特效库",
+      "evidence": "本轮参考UI"
+    },
+    {
+      "id": "SE-059",
+      "group": "转场",
+      "item": "原始动态素材",
+      "status": "通过",
+      "detail": "直接使用参考原始WebP，每张75帧/3秒；本地动画可见",
+      "evidence": "public/transitions/本轮UI"
+    },
+    {
+      "id": "SE-060",
+      "group": "转场",
+      "item": "分类栏与卡片位置",
+      "status": "已修复",
+      "detail": "分类48px高、13px字体，卡片108×130，图框100×100，列距7px",
+      "evidence": "参考DOM/本地CSS"
+    },
+    {
+      "id": "SE-061",
+      "group": "转场",
+      "item": "选中边框和颜色",
+      "status": "已修复",
+      "detail": "2px改1px，避免图片缩小；边框/文字改#09caf5，圆角6px",
+      "evidence": "参考DOM/本地截图"
+    },
+    {
+      "id": "SE-062",
+      "group": "转场",
+      "item": "时长与名称行间距",
+      "status": "已修复",
+      "detail": "改16px行高，时长框与90px小预览纵向位置对齐",
+      "evidence": "参考DOM/本地CSS"
+    },
+    {
+      "id": "SE-063",
+      "group": "转场",
+      "item": "时长上下限与步长",
+      "status": "通过",
+      "detail": "0.2–4秒，0.1步进，非法值失焦归一化",
+      "evidence": "前批UI/本轮合同"
+    },
+    {
+      "id": "SE-064",
+      "group": "转场",
+      "item": "相邻素材小预览",
+      "status": "通过",
+      "detail": "本地使用相邻真实视频，静音循环并应用效果",
+      "evidence": "本轮UI"
+    },
+    {
+      "id": "SE-065",
+      "group": "转场",
+      "item": "主画面预览时机",
+      "status": "已修复",
+      "detail": "主画面保持播放头，右侧小预览独立循环；取消选择转场时自动替换整个主画面的分支。",
+      "evidence": "本轮浏览器主画面停0秒、小预览独立循环/既有参考观察"
+    },
+    {
+      "id": "SE-066",
+      "group": "转场",
+      "item": "居中时序和总片长",
+      "status": "通过",
+      "detail": "三种2.2/4秒真实成片均保留4秒总长",
+      "evidence": "本轮FFmpeg"
+    },
+    {
+      "id": "SE-067",
+      "group": "转场",
+      "item": "黑场与白场中点",
+      "status": "通过",
+      "detail": "真实像素验证纯黑/纯白中点",
+      "evidence": "本轮FFmpeg"
+    },
+    {
+      "id": "SE-068",
+      "group": "转场",
+      "item": "极短片段和连续接点",
+      "status": "通过",
+      "detail": "短片段约束和避免三片同时重叠的合同通过",
+      "evidence": "本轮helpers/前批压力合成"
+    },
+    {
+      "id": "SE-069",
+      "group": "转场",
+      "item": "删除与硬切",
+      "status": "通过",
+      "detail": "已有删除恢复硬切；本轮逐片段与硬切组合实际合成通过",
+      "evidence": "FFmpeg/源码"
+    },
+    {
+      "id": "SE-070",
+      "group": "转场",
+      "item": "时间轴覆盖框",
+      "status": "已修复",
+      "detail": "覆盖框top4、高60，最小点击宽44；补选中填充、两侧标记和16px参考图标。缩放滑块扩0–100，以参考19.6787=100px/s、50=251px/s、100=500px/s测量映射。",
+      "evidence": "本轮参考可见DOM/本地CSS"
+    },
+    {
+      "id": "SE-071",
+      "group": "转场",
+      "item": "旧方案兼容",
+      "status": "通过",
+      "detail": "未设置alignment的旧重叠方案保留原片长语义",
+      "evidence": "本轮helpers/FFmpeg"
+    },
+    {
+      "id": "SE-072",
+      "group": "音频",
+      "item": "有声视频分离",
+      "status": "通过",
+      "detail": "本轮真实点按生成独立音轨，撤销后消失",
+      "evidence": "本轮本地UI"
+    },
+    {
+      "id": "SE-073",
+      "group": "音频",
+      "item": "无音轨时分离禁用",
+      "status": "通过",
+      "detail": "参考当前无音轨素材禁用；本地按hasAudio禁用",
+      "evidence": "本轮参考UI/源码"
+    },
+    {
+      "id": "SE-074",
+      "group": "音频",
+      "item": "分离后声音不叠加",
+      "status": "通过",
+      "detail": "实际音频RMS与转场包络回归通过",
+      "evidence": "本轮FFmpeg"
+    },
+    {
+      "id": "SE-075",
+      "group": "音频",
+      "item": "独立音轨变速、淡入淡出",
+      "status": "通过",
+      "detail": "实际2倍速、增益、延迟、淡入淡出及长音轨黑尾回归通过",
+      "evidence": "本轮FFmpeg"
+    },
+    {
+      "id": "SE-076",
+      "group": "音频",
+      "item": "独立音轨裁剪、分割、移动",
+      "status": "代码核对",
+      "detail": "有完整source范围与时间轴处理；前批UI通过，本轮未全部重拖",
+      "evidence": "SmartEditEditor/contract"
+    },
+    {
+      "id": "SE-077",
+      "group": "音频",
+      "item": "人声分离两个分支",
+      "status": "通过",
+      "detail": "HTDemucs真实离线分离，视频和独立音轨可用；支持仅人声、仅背景及两轨输出，保留速度/增益/静音/包络。混合信号重构最大误差1.49e-8。",
+      "evidence": "smart-edit-models.py；主站API与UI"
+    },
+    {
+      "id": "SE-078",
+      "group": "音频",
+      "item": "音频参数栏调整宽度",
+      "status": "已修复",
+      "detail": "音频参数栏新增同一宽度拖拽柄；参考有声参数更多行为仍待验。",
+      "evidence": "SmartEditEditor"
+    },
+    {
+      "id": "SE-079",
+      "group": "音频",
+      "item": "真实波形",
+      "status": "待验",
+      "detail": "本地音频条是纯色色块，没有采样波形；本轮未取得参考有声轨波形证据",
+      "evidence": "se-wave/se-audio-clip"
+    },
+    {
+      "id": "SE-080",
+      "group": "音频",
+      "item": "音乐与卡点",
+      "status": "通过",
+      "detail": "节拍与候选剪辑合同回归通过，音乐混音真实输出通过；不是参考算法效果对拍",
+      "evidence": "本轮helpers/FFmpeg"
+    },
+    {
+      "id": "SE-081",
+      "group": "导出与恢复",
+      "item": "导出菜单位置",
+      "status": "通过",
+      "detail": "本地/画布两个目的地和参考相同",
+      "evidence": "参考export-menu/源码"
+    },
+    {
+      "id": "SE-082",
+      "group": "导出与恢复",
+      "item": "导出菜单视觉",
+      "status": "代码核对",
+      "detail": "本地有200px菜单、40px条目；边框/图标需独立像素精调",
+      "evidence": "参考截图/CSS"
+    },
+    {
+      "id": "SE-083",
+      "group": "导出与恢复",
+      "item": "真实本地合成",
+      "status": "通过",
+      "detail": "FFmpeg实际生成结果，不用模拟成功替代",
+      "evidence": "本轮ffmpeg.log"
+    },
+    {
+      "id": "SE-084",
+      "group": "导出与恢复",
+      "item": "提交任务ID和幂等",
+      "status": "通过",
+      "detail": "请求ID回归与远程确定性任务身份测试通过",
+      "evidence": "本轮helpers"
+    },
+    {
+      "id": "SE-085",
+      "group": "导出与恢复",
+      "item": "刷新恢复导出任务",
+      "status": "通过",
+      "detail": "任务检查点/reconcile合同通过；本轮未执行真实Next页面断网恢复",
+      "evidence": "本轮helpers"
+    },
+    {
+      "id": "SE-086",
+      "group": "导出与恢复",
+      "item": "浏览器中返回画布结果",
+      "status": "通过",
+      "detail": "主站Next生产构建：真实登录/隔离数据库/上传/排队/分离/转写/合成/MP4和SRT下载/浏览器回填播放及刷新恢复通过。存储用本地S3兼容服务，未测线上存储或部署。修复新画布时区错位漏回填与成片时长显示。",
+      "evidence": "production-chain.log；main-result.mp4；主站浏览器"
+    },
+    {
+      "id": "SE-087",
+      "group": "导出与恢复",
+      "item": "任务错误和服务不可用",
+      "status": "代码核对",
+      "detail": "导出按钮按能力禁用，失败状态显示错误；本轮未切断实际服务故障注入",
+      "evidence": "SmartEditEditor/SmartEditDialog"
+    },
+    {
+      "id": "SE-088",
+      "group": "导出与恢复",
+      "item": "取消远程导出",
+      "status": "部分修复",
+      "detail": "已接排队导出的取消按钮到现有鉴权取消API，取消状态不再等待。运行中远程FFmpeg的主动终止仍未实现，主站取消按钮全链路未验。",
+      "evidence": "onCancel/现有task-store"
+    },
+    {
+      "id": "SE-089",
+      "group": "导出与恢复",
+      "item": "多版本全导出",
+      "status": "代码核对",
+      "detail": "高级设置有导出全部版本；主节点默认只一版，参考完整多版结果未生成",
+      "evidence": "SmartEditDialog"
+    },
+    {
+      "id": "SE-090",
+      "group": "导出与恢复",
+      "item": "最高时长与比例",
+      "status": "通过",
+      "detail": "300秒约束、宽屏远程结果下载合同、6比例真实本地输出通过",
+      "evidence": "本轮helpers/FFmpeg"
+    },
+    {
+      "id": "SE-091",
+      "group": "边界与性能",
+      "item": "最大片段数",
+      "status": "代码核对",
+      "detail": "最多12视频片段、24独立音轨、500字幕；不是参考站已确认相同限制",
+      "evidence": "contract"
+    },
+    {
+      "id": "SE-092",
+      "group": "边界与性能",
+      "item": "后台标签恢复",
+      "status": "待验",
+      "detail": "循环预览监听visibilitychange；本轮未覆盖系统休眠/后台节流完整矩阵",
+      "evidence": "TransitionPreview"
+    },
+    {
+      "id": "SE-093",
+      "group": "边界与性能",
+      "item": "原生媒体加载失败",
+      "status": "代码核对",
+      "detail": "play的AbortError忽略，其余报错可见；本轮未测试全部外部域名/CORS",
+      "evidence": "EditPreview"
+    },
+    {
+      "id": "SE-094",
+      "group": "边界与性能",
+      "item": "高清与12片段性能",
+      "status": "既有证据",
+      "detail": "前批1080P4秒合成12.81秒、12片段7.604秒；本轮未重复此压力测试",
+      "evidence": "transition-stress-native.log"
+    },
+    {
+      "id": "SE-095",
+      "group": "边界与性能",
+      "item": "服务部署状态",
+      "status": "通过",
+      "detail": "本轮所有修改仅本地，未更新48或算力服务器",
+      "evidence": "git/操作记录"
+    }
+  ],
+  "date": "2026-09-18",
+  "note": "09-18续修：人声分离、自动转写、2002字体目录和本地主站完整智能剪辑链路已接通；仍非全站所有功能验收，未部署或推送；逐字体像素/DPR与休眠断网矩阵未完成。"
+}

docs/smart-edit-fixes-2026-09-18.md

@@ -0,0 +1,177 @@
+# 智能剪辑修复进度（2026-09-18）
+
+基线 `68f62dae`，承接09-17的95项清单；**仅本地修改和验证，未部署、未推送**。累计状态不是全项本轮复测，保留每行的证据边界。尚不能认定为全部功能齐全或像素级完全一致。
+
+## 本轮结果
+
+已处理字幕行操作、字体搜索/分类、五种实际字体与成片一致性、新字幕1080坐标基准、字间距/描边/单位、正音量增益和静音恢复、收起预览与Esc、真实节点时间轴、音乐卡点与广告版本设置、音频面板拖宽、转场覆盖框与主画面预览时机。
+
+增加排队导出的取消入口。独立合成服务构建携带字体包；真实测试发现临时字体目录清理会误报合成失败，已修正为递归清理，非关键清理失败不删除已验证成片。
+
+## 验证
+
+- TypeScript：通过。
+- 回归：496通过、1跳过；最后定向剪辑回归21通过。
+- 真实FFmpeg：六种比例、三种转场、居中/旧重叠时序、字幕位置/透明度；正增益振幅、原声及独立音轨静音；五种字体实际字形输出。
+- 真实分析：PySceneDetect 3镜头、RMS 2有效区段、librosa 22个节拍。
+- 真实本地HTTP合成服务：鉴权、真实字体成片、进程强制中断后恢复、幂等、冲突拒绝、完成记录重启保留。
+- 浏览器独立测试：字幕新增和字体切换、新字幕56px、+0.5dB静音后恢复、折叠播放到片尾、Esc退出、卡点缺音乐拦截、选择音乐、广告默认3版。
+
+本地独立页面仍用样例任务回调，不能冒充主站Next API端到端成功；上面的真实HTTP测试针对独立合成服务。测试目录 `.qa/smart-edit/fixes-0918/` 的日志与截图不提交到Git，复现测试脚本提交。
+
+## 09-18续修：本轮指定的四项
+
+人声分离、自动转写、2002项字体目录、主站智能剪辑本地全链路已接通并使用真实模型/FFmpeg验证。详细环境、数据及边界见 [续修验收记录](smart-edit-completion-2026-09-18.md)。以上早先的“独立测试页”结论保留为历史；本轮新增真实Next生产构建验收。
+
+## 仍未完成的其他范围
+
+- 全部2002字体逐字形/行高/DPR黄金图对拍；字体目录和实际字体文件已齐全。
+- 运行中远程合成主动取消；后台休眠/断网/各窄窗口完整矩阵。
+- 五种模式与参考的算法效果逐案对比、广告多创意策略、真实波形及未穷尽边界。
+- 线上部署与真实线上对象存储验收。本轮只用隔离本地数据库及S3兼容测试服务。
+
+参考字号证据：预览iframe可见SPAN为56px，父画布1920×1080。缩放证据：slider实际状态19.67871485943775时2.2秒转场宽220px，50时552.2px，100时1100px；采用2+4.98×slider的映射。最小转场覆盖框44px单独约束。
+
+## 累计95项
+
+- 通过：39
+- 待验：6
+- 代码核对：17
+- 已修复：28
+- 部分修复：4
+- 既有证据：1
+
+### 节点与输入
+
+| ID | 检查项 | 状态 | 当前结论 | 证据 |
+| --- | --- | --- | --- | --- |
+| SE-001 | 打开剪辑 | 通过 | 参考和本地均有入口；本地进入真实编辑组件 | 本轮UI |
+| SE-002 | 五种模式名称和顺序 | 通过 | 默认、讲解、批量广告、口播、素材混剪相同 | 参考node-modes截图/SmartEditNode |
+| SE-003 | 五种模式算法效果 | 待验 | 本地有模式分支；没有重新提交参考站付费生成，不能认定效果等同 | execute/contract |
+| SE-004 | 六种比例 | 通过 | 21:9、16:9、9:16、1:1、4:3、3:4；实际合成覆盖 | 参考node-settings/本轮FFmpeg |
+| SE-005 | 三种清晰度 | 通过 | 480P、720P、1080P；控件一致 | 参考node-settings/SmartEditNode |
+| SE-006 | 时长范围 | 通过 | 0–300秒、步长1秒；本地0为自动 | 参考可见DOM/contract |
+| SE-007 | 提示词编辑与展开 | 代码核对 | 本地受控输入、展开及提交；参考既有提示词未改动 | SmartEditNode |
+| SE-008 | 引用画布素材 | 代码核对 | 参考进入引用选择；生产组件onReference连接连线状态，本地独立测试页只是候选列表 | 参考引用模式/主画布调用点 |
+| SE-009 | 素材序号、时长、移除 | 代码核对 | 本地有缩略图、序号、时长及移除；未删除参考原素材 | SmartEditNode |
+| SE-010 | 本地上传与数量限制 | 代码核对 | 真实上传入口，上限12个素材、单个125MiB；参考上限尚未验证 | SmartEditNode/SmartEditDialog |
+| SE-011 | 节点缩略时间轴 | 已修复 | 节点刻度、视频、字幕、独立音轨从实际计划生成；重新生成后按任务版本选择新计划，保留所选版本。 | 本轮节点UI/SmartEditNodeCard |
+| SE-012 | 批量广告多版本 | 部分修复 | 节点新增1–3版选择，未指定时广告默认3版；实际仍是候选顺序变体，未认定为参考站的多创意算法。 | 本轮UI/intent helpers |
+| SE-013 | 提示词要求卡点 | 已修复 | 节点提示词的常见“卡点/beat sync”表达关联开关，显式关闭优先；支持画布音乐选择/上传及2或4拍切换，缺音乐阻止提交。并非通用自然语言意图解析。 | 本轮UI/intent helpers/真实librosa 22拍 |
+
+### 布局与操作
+
+| ID | 检查项 | 状态 | 当前结论 | 证据 |
+| --- | --- | --- | --- | --- |
+| SE-014 | 顶栏高度和边距 | 通过 | 48px顶栏、8px页面边距；本轮图像对照 | 参考DOM/本地CSS |
+| SE-015 | 左右参数栏宽度 | 通过 | 转场库372px、参数栏300px，均可拖动调整 | 参考DOM/本地CSS |
+| SE-016 | 面板背景与关闭按钮 | 已修复 | 背景统一#202020，边框白色10%，关闭按钮24px，标题500/14px | 参考transition.json/本地CSS |
+| SE-017 | 默认时间轴高度 | 已修复 | 保留打开时的像素高度，窗口缩小时仅为可用空间收缩，不再每次重算35vh。 | resize代码 |
+| SE-018 | 折叠预览 | 已修复 | 收起后背景透明、可操作下方画布，仅底部保留时间轴；预览组件保留，实测能播放到片尾。独立页面下层为测试按钮，主画布接入仍需整体验收。 | 本轮本地UI |
+| SE-019 | 全屏预览 | 代码核对 | 本地调用Fullscreen API并显示拒绝错误；本轮未重测不同浏览器权限状态 | SmartEditEditor |
+| SE-020 | 撤销重做 | 通过 | 本轮分离音轨后撤销成功；参考宽屏有按钮、窄屏隐藏，先前未见不代表没有 | 本轮UI/参考1512宽度 |
+| SE-021 | 播放、暂停、逐帧 | 代码核对 | Space与左右键、30fps步进、片尾停止；前批浏览器已验证，本轮保留代码核对口径 | SmartEditEditor/EditPreview |
+| SE-022 | 吸附默认值 | 通过 | 参考关闭，本地SmartEditEditor也是false；纠正旧文档“本地开启”误记 | 参考AX/SmartEditEditor:state |
+| SE-023 | 退出与Esc | 已修复 | Esc退出整个编辑器，输入框编辑时不误退出；收起时保留关闭按钮。 | 本轮本地UI |
+| SE-024 | 窄屏可用性 | 待验 | 参考1024宽度测得左右仍372/300；本地三栏与工具条更窄窗口尚无完整可用性验收 | reference-transition-1024 |
+
+### 视频片段
+
+| ID | 检查项 | 状态 | 当前结论 | 证据 |
+| --- | --- | --- | --- | --- |
+| SE-025 | 四角缩放、移动与重置 | 通过 | 本轮真实合成像素验证；本地有选框和重置，参考四角手柄可见 | 本轮FFmpeg/参考video |
+| SE-026 | 变速上下调步长 | 已修复 | 参考1.00→1.01，本地0.05改为0.01；浏览器已复测 | 本轮UI |
+| SE-027 | 变速支持范围 | 待验 | 本地0.5–2倍；参考未穷尽上下限，不能因步长一致声称范围一致 | contract/参考video |
+| SE-028 | 时长输入 | 代码核对 | 按原素材剩余长度和变速约束，不能超过素材 | SmartEditEditor/contract |
+| SE-029 | 数值越界后的显示 | 已修复 | 到上限2后再输入99，显示回到2.00，不残留非法数字 | 本轮UI |
+| SE-030 | 增益步长 | 已修复 | 参考0.5dB，本地由1改0.5；浏览器0→-0.5确认 | 本轮UI |
+| SE-031 | 正增益 | 已修复 | 本地支持最高+12dB，预览使用GainNode，导出实际音量放大；参考上限尚未确认。跨域正增益素材经登录下载接口取得同源副本，需另测真实主站链路。 | 本轮浏览器+0.5dB/FFmpeg约2倍振幅 |
+| SE-032 | 边缘裁剪和排序 | 代码核对 | 已有按source时间/速度换算和撤销；本轮未重新拖遍所有方向 | SmartEditEditor |
+| SE-033 | 分割和左右裁剪 | 通过 | 边界禁用、源范围及片长合同回归通过；本轮未对拍参考破坏性编辑 | 本轮helpers/SmartEditEditor |
+| SE-034 | 至少保留一个片段 | 待验 | 本地阻止删最后一段；参考空时间轴行为未删除原素材验证 | SmartEditEditor |
+| SE-035 | 右键菜单 | 部分修复 | 已增加人声分离子菜单及仅保留人声、仅保留背景音、两轨输出；菜单其他文案与参考仍有差异。 | 本轮主站UI/真实HTDemucs |
+| SE-036 | 转场覆盖区右键 | 已修复 | 转场覆盖区域按指针时间选择底层片段并打开视频右键菜单。 | 本轮浏览器：覆盖框右键选中有声片段并成功分离 |
+| SE-037 | 静音后恢复原增益 | 已修复 | 新增可选muted独立于gain保存，原声和独立音轨恢复不会覆盖原增益；分离时继承静音状态。 | 浏览器原声及独立轨+0.5dB静音恢复/真实FFmpeg两类静音 |
+
+### 字幕
+
+| ID | 检查项 | 状态 | 当前结论 | 证据 |
+| --- | --- | --- | --- | --- |
+| SE-038 | 空列表、新建、搜索 | 通过 | 参考空列表和新建已操作；本地新建成功、搜索有真实过滤 | 本轮UI/源码 |
+| SE-039 | 列表逐条新增与删除 | 已修复 | 每行新增在后面插入/删除，插入按时间排序、遇下条字幕或片尾缩短，并定位新字幕；片尾不能插入时给提示。 | 本轮UI/代码 |
+| SE-040 | 列表时间格式 | 已修复 | 时间框采用MM:SS.xx，可输入小数秒；失焦校验范围，非法值还原，Esc取消编辑。参考只展示整秒的像素细节尚未作为黄金图验收。 | 本轮UI |
+| SE-041 | 选中文本、编辑与位置 | 通过 | 本轮本地样式编辑和实际字幕位置像素测试通过 | 本轮UI/FFmpeg |
+| SE-042 | 显示和隐藏 | 代码核对 | subtitles控制预览和导出；关闭面板仅隐藏编辑轨道 | SmartEditEditor/compiledSubtitles |
+| SE-043 | 字体搜索和长列表 | 通过 | 参考目录2002项完整收录，1997新增文件全部下载解析与双SHA校验；搜索、四分类、虚拟滚动、按需加载可用。 | remaining-0918/reference-fonts.json、prepare-catalog.log；主站字体切换 |
+| SE-044 | 字体分类 | 已修复 | 新增全部/常用/推荐/Google四类筛选；范围限已打包的实际字体。 | 本轮UI/CaptionFonts |
+| SE-045 | 预览与导出的字形 | 部分修复 | 2002字体目录已齐；预览与FFmpeg使用同一固定文件，5原有+6新增字体实际合成通过；未对2002种字体逐字形、行高、DPR作黄金图对拍。 | catalog-render.log/主站大宝桃桃体切换 |
+| SE-046 | 字号范围和显示单位 | 已修复 | 从参考可见iframe确认1920×1080、56px；新字幕明确1080基准，默认56/12–72px；旧720基准按比例显示并保留历史超范围值。 | 本轮参考DOM/legacy兼容helpers/本地UI |
+| SE-047 | 字间距范围和控件 | 已修复 | 字间距改X数值框0–100；旧工程负值保留，预览与ASS依各自基准缩放。 | StyleNumber/contract |
+| SE-048 | 字体颜色和十六进制 | 代码核对 | 两处均有取色和hex；本地失焦校验非法值；样式烧录已回归 | StyleColor/FFmpeg |
+| SE-049 | 六种预设入口 | 通过 | 默认、白黑、黑白、黄黑、白粉、粉黑均有 | 本轮参考及本地UI |
+| SE-050 | 预设按钮外观 | 已修复 | 默认预设显示“默认”，其他为T；选中描边按实际样式匹配。 | 本轮字幕UI |
+| SE-051 | 描边颜色和粗细 | 已修复 | 新字幕默认白色描边，粗细0–20；旧字幕不变。预览和ASS同步缩放。 | 本轮UI/contract/FFmpeg |
+| SE-052 | 描边不透明度语义 | 已修复 | 独立outlineOpacity；0时保留文字，只去描边；旧整行opacity保持兼容 | UI截图/ASS合同/实际像素测试 |
+| SE-053 | 字号等数值单位装饰 | 已修复 | 增加px/%单位和上下步进按钮；数值失焦校验，不提交编辑中的空值。 | StyleNumber/本地构建 |
+| SE-054 | 应用全部 | 代码核对 | 复制完整style含新outlineOpacity；本轮未覆盖所有字幕数量组合 | SmartEditEditor |
+| SE-055 | 自动字幕与长转场 | 已修复 | 居中转场不再把后片段字幕向后延迟整个转场时长，避免片头字幕丢失 | 新增合同回归 |
+| SE-056 | 手工字幕与SRT | 通过 | 手工字幕保留位置和样式，SRT仅保留文本时间；ASS字符转义回归通过 | 本轮helpers/FFmpeg |
+| SE-057 | 语音自动转写 | 通过 | 真实FunASR转写成功；修复缺少pred_timestamp造成整段无标点时间戳的问题，限制负起点并映射剪辑速度/裁剪。字幕按钮提交主站任务，完成后刷新恢复。 | model-regression/report.json；production-chain.log；主站浏览器 |
+
+### 转场
+
+| ID | 检查项 | 状态 | 当前结论 | 证据 |
+| --- | --- | --- | --- | --- |
+| SE-058 | 三种基础转场入口 | 通过 | 淡入淡出、黑场过渡、白场过渡；这不是全站视频生成特效库 | 本轮参考UI |
+| SE-059 | 原始动态素材 | 通过 | 直接使用参考原始WebP，每张75帧/3秒；本地动画可见 | public/transitions/本轮UI |
+| SE-060 | 分类栏与卡片位置 | 已修复 | 分类48px高、13px字体，卡片108×130，图框100×100，列距7px | 参考DOM/本地CSS |
+| SE-061 | 选中边框和颜色 | 已修复 | 2px改1px，避免图片缩小；边框/文字改#09caf5，圆角6px | 参考DOM/本地截图 |
+| SE-062 | 时长与名称行间距 | 已修复 | 改16px行高，时长框与90px小预览纵向位置对齐 | 参考DOM/本地CSS |
+| SE-063 | 时长上下限与步长 | 通过 | 0.2–4秒，0.1步进，非法值失焦归一化 | 前批UI/本轮合同 |
+| SE-064 | 相邻素材小预览 | 通过 | 本地使用相邻真实视频，静音循环并应用效果 | 本轮UI |
+| SE-065 | 主画面预览时机 | 已修复 | 主画面保持播放头，右侧小预览独立循环；取消选择转场时自动替换整个主画面的分支。 | 本轮浏览器主画面停0秒、小预览独立循环/既有参考观察 |
+| SE-066 | 居中时序和总片长 | 通过 | 三种2.2/4秒真实成片均保留4秒总长 | 本轮FFmpeg |
+| SE-067 | 黑场与白场中点 | 通过 | 真实像素验证纯黑/纯白中点 | 本轮FFmpeg |
+| SE-068 | 极短片段和连续接点 | 通过 | 短片段约束和避免三片同时重叠的合同通过 | 本轮helpers/前批压力合成 |
+| SE-069 | 删除与硬切 | 通过 | 已有删除恢复硬切；本轮逐片段与硬切组合实际合成通过 | FFmpeg/源码 |
+| SE-070 | 时间轴覆盖框 | 已修复 | 覆盖框top4、高60，最小点击宽44；补选中填充、两侧标记和16px参考图标。缩放滑块扩0–100，以参考19.6787=100px/s、50=251px/s、100=500px/s测量映射。 | 本轮参考可见DOM/本地CSS |
+| SE-071 | 旧方案兼容 | 通过 | 未设置alignment的旧重叠方案保留原片长语义 | 本轮helpers/FFmpeg |
+
+### 音频
+
+| ID | 检查项 | 状态 | 当前结论 | 证据 |
+| --- | --- | --- | --- | --- |
+| SE-072 | 有声视频分离 | 通过 | 本轮真实点按生成独立音轨，撤销后消失 | 本轮本地UI |
+| SE-073 | 无音轨时分离禁用 | 通过 | 参考当前无音轨素材禁用；本地按hasAudio禁用 | 本轮参考UI/源码 |
+| SE-074 | 分离后声音不叠加 | 通过 | 实际音频RMS与转场包络回归通过 | 本轮FFmpeg |
+| SE-075 | 独立音轨变速、淡入淡出 | 通过 | 实际2倍速、增益、延迟、淡入淡出及长音轨黑尾回归通过 | 本轮FFmpeg |
+| SE-076 | 独立音轨裁剪、分割、移动 | 代码核对 | 有完整source范围与时间轴处理；前批UI通过，本轮未全部重拖 | SmartEditEditor/contract |
+| SE-077 | 人声分离两个分支 | 通过 | HTDemucs真实离线分离，视频和独立音轨可用；支持仅人声、仅背景及两轨输出，保留速度/增益/静音/包络。混合信号重构最大误差1.49e-8。 | smart-edit-models.py；主站API与UI |
+| SE-078 | 音频参数栏调整宽度 | 已修复 | 音频参数栏新增同一宽度拖拽柄；参考有声参数更多行为仍待验。 | SmartEditEditor |
+| SE-079 | 真实波形 | 待验 | 本地音频条是纯色色块，没有采样波形；本轮未取得参考有声轨波形证据 | se-wave/se-audio-clip |
+| SE-080 | 音乐与卡点 | 通过 | 节拍与候选剪辑合同回归通过，音乐混音真实输出通过；不是参考算法效果对拍 | 本轮helpers/FFmpeg |
+
+### 导出与恢复
+
+| ID | 检查项 | 状态 | 当前结论 | 证据 |
+| --- | --- | --- | --- | --- |
+| SE-081 | 导出菜单位置 | 通过 | 本地/画布两个目的地和参考相同 | 参考export-menu/源码 |
+| SE-082 | 导出菜单视觉 | 代码核对 | 本地有200px菜单、40px条目；边框/图标需独立像素精调 | 参考截图/CSS |
+| SE-083 | 真实本地合成 | 通过 | FFmpeg实际生成结果，不用模拟成功替代 | 本轮ffmpeg.log |
+| SE-084 | 提交任务ID和幂等 | 通过 | 请求ID回归与远程确定性任务身份测试通过 | 本轮helpers |
+| SE-085 | 刷新恢复导出任务 | 通过 | 任务检查点/reconcile合同通过；本轮未执行真实Next页面断网恢复 | 本轮helpers |
+| SE-086 | 浏览器中返回画布结果 | 通过 | 主站Next生产构建：真实登录/隔离数据库/上传/排队/分离/转写/合成/MP4和SRT下载/浏览器回填播放及刷新恢复通过。存储用本地S3兼容服务，未测线上存储或部署。修复新画布时区错位漏回填与成片时长显示。 | production-chain.log；main-result.mp4；主站浏览器 |
+| SE-087 | 任务错误和服务不可用 | 代码核对 | 导出按钮按能力禁用，失败状态显示错误；本轮未切断实际服务故障注入 | SmartEditEditor/SmartEditDialog |
+| SE-088 | 取消远程导出 | 部分修复 | 已接排队导出的取消按钮到现有鉴权取消API，取消状态不再等待。运行中远程FFmpeg的主动终止仍未实现，主站取消按钮全链路未验。 | onCancel/现有task-store |
+| SE-089 | 多版本全导出 | 代码核对 | 高级设置有导出全部版本；主节点默认只一版，参考完整多版结果未生成 | SmartEditDialog |
+| SE-090 | 最高时长与比例 | 通过 | 300秒约束、宽屏远程结果下载合同、6比例真实本地输出通过 | 本轮helpers/FFmpeg |
+
+### 边界与性能
+
+| ID | 检查项 | 状态 | 当前结论 | 证据 |
+| --- | --- | --- | --- | --- |
+| SE-091 | 最大片段数 | 代码核对 | 最多12视频片段、24独立音轨、500字幕；不是参考站已确认相同限制 | contract |
+| SE-092 | 后台标签恢复 | 待验 | 循环预览监听visibilitychange；本轮未覆盖系统休眠/后台节流完整矩阵 | TransitionPreview |
+| SE-093 | 原生媒体加载失败 | 代码核对 | play的AbortError忽略，其余报错可见；本轮未测试全部外部域名/CORS | EditPreview |
+| SE-094 | 高清与12片段性能 | 既有证据 | 前批1080P4秒合成12.81秒、12片段7.604秒；本轮未重复此压力测试 | transition-stress-native.log |
+| SE-095 | 服务部署状态 | 通过 | 本轮所有修改仅本地，未更新48或算力服务器 | git/操作记录 |

docs/smart-edit-full-recheck-2026-09-17.json

@@ -0,0 +1,775 @@
+{
+  "scope": "智能剪辑节点、编辑器及直接依赖的音频/导出链路；不是全站所有模块",
+  "baseline": "431ca03f",
+  "local_only": true,
+  "counts": {
+    "通过": 35,
+    "待验": 8,
+    "代码核对": 17,
+    "差异": 25,
+    "已修复": 9,
+    "既有证据": 1
+  },
+  "items": [
+    {
+      "id": "SE-001",
+      "group": "节点与输入",
+      "item": "打开剪辑",
+      "status": "通过",
+      "detail": "参考和本地均有入口；本地进入真实编辑组件",
+      "evidence": "本轮UI"
+    },
+    {
+      "id": "SE-002",
+      "group": "节点与输入",
+      "item": "五种模式名称和顺序",
+      "status": "通过",
+      "detail": "默认、讲解、批量广告、口播、素材混剪相同",
+      "evidence": "参考node-modes截图/SmartEditNode"
+    },
+    {
+      "id": "SE-003",
+      "group": "节点与输入",
+      "item": "五种模式算法效果",
+      "status": "待验",
+      "detail": "本地有模式分支；没有重新提交参考站付费生成，不能认定效果等同",
+      "evidence": "execute/contract"
+    },
+    {
+      "id": "SE-004",
+      "group": "节点与输入",
+      "item": "六种比例",
+      "status": "通过",
+      "detail": "21:9、16:9、9:16、1:1、4:3、3:4；实际合成覆盖",
+      "evidence": "参考node-settings/本轮FFmpeg"
+    },
+    {
+      "id": "SE-005",
+      "group": "节点与输入",
+      "item": "三种清晰度",
+      "status": "通过",
+      "detail": "480P、720P、1080P；控件一致",
+      "evidence": "参考node-settings/SmartEditNode"
+    },
+    {
+      "id": "SE-006",
+      "group": "节点与输入",
+      "item": "时长范围",
+      "status": "通过",
+      "detail": "0–300秒、步长1秒；本地0为自动",
+      "evidence": "参考可见DOM/contract"
+    },
+    {
+      "id": "SE-007",
+      "group": "节点与输入",
+      "item": "提示词编辑与展开",
+      "status": "代码核对",
+      "detail": "本地受控输入、展开及提交；参考既有提示词未改动",
+      "evidence": "SmartEditNode"
+    },
+    {
+      "id": "SE-008",
+      "group": "节点与输入",
+      "item": "引用画布素材",
+      "status": "代码核对",
+      "detail": "参考进入引用选择；生产组件onReference连接连线状态，本地独立测试页只是候选列表",
+      "evidence": "参考引用模式/主画布调用点"
+    },
+    {
+      "id": "SE-009",
+      "group": "节点与输入",
+      "item": "素材序号、时长、移除",
+      "status": "代码核对",
+      "detail": "本地有缩略图、序号、时长及移除；未删除参考原素材",
+      "evidence": "SmartEditNode"
+    },
+    {
+      "id": "SE-010",
+      "group": "节点与输入",
+      "item": "本地上传与数量限制",
+      "status": "代码核对",
+      "detail": "真实上传入口，上限12个素材、单个125MiB；参考上限尚未验证",
+      "evidence": "SmartEditNode/SmartEditDialog"
+    },
+    {
+      "id": "SE-011",
+      "group": "节点与输入",
+      "item": "节点缩略时间轴",
+      "status": "差异",
+      "detail": "本地节点卡片仍固定生成12个刻度、两行8个色块，未绑定实际剪辑轨道",
+      "evidence": "SmartEditNodeCard"
+    },
+    {
+      "id": "SE-012",
+      "group": "节点与输入",
+      "item": "批量广告多版本",
+      "status": "差异",
+      "detail": "本地高级设置支持1–3版，节点默认variants为1；候选顺序变体不等于多套广告创意",
+      "evidence": "contract/buildEditPlans"
+    },
+    {
+      "id": "SE-013",
+      "group": "节点与输入",
+      "item": "提示词要求卡点",
+      "status": "差异",
+      "detail": "真正卡点要求音乐和beatSync；仅在提示词写“卡点”不会自动打开节拍流程",
+      "evidence": "execute/主画布onSubmit"
+    },
+    {
+      "id": "SE-014",
+      "group": "布局与操作",
+      "item": "顶栏高度和边距",
+      "status": "通过",
+      "detail": "48px顶栏、8px页面边距；本轮图像对照",
+      "evidence": "参考DOM/本地CSS"
+    },
+    {
+      "id": "SE-015",
+      "group": "布局与操作",
+      "item": "左右参数栏宽度",
+      "status": "通过",
+      "detail": "转场库372px、参数栏300px，均可拖动调整",
+      "evidence": "参考DOM/本地CSS"
+    },
+    {
+      "id": "SE-016",
+      "group": "布局与操作",
+      "item": "面板背景与关闭按钮",
+      "status": "已修复",
+      "detail": "背景统一#202020，边框白色10%，关闭按钮24px，标题500/14px",
+      "evidence": "参考transition.json/本地CSS"
+    },
+    {
+      "id": "SE-017",
+      "group": "布局与操作",
+      "item": "默认时间轴高度",
+      "status": "差异",
+      "detail": "打开时均约35vh；参考改变视口保留原像素高度，本地未手动拖动时随窗口重算",
+      "evidence": "本轮宽屏切换/resize处理"
+    },
+    {
+      "id": "SE-018",
+      "group": "布局与操作",
+      "item": "折叠预览",
+      "status": "差异",
+      "detail": "参考保留底部时间轴并露出画布；本地把时间轴铺满编辑器且保留顶栏",
+      "evidence": "参考collapsed截图/SmartEditEditor"
+    },
+    {
+      "id": "SE-019",
+      "group": "布局与操作",
+      "item": "全屏预览",
+      "status": "代码核对",
+      "detail": "本地调用Fullscreen API并显示拒绝错误；本轮未重测不同浏览器权限状态",
+      "evidence": "SmartEditEditor"
+    },
+    {
+      "id": "SE-020",
+      "group": "布局与操作",
+      "item": "撤销重做",
+      "status": "通过",
+      "detail": "本轮分离音轨后撤销成功；参考宽屏有按钮、窄屏隐藏，先前未见不代表没有",
+      "evidence": "本轮UI/参考1512宽度"
+    },
+    {
+      "id": "SE-021",
+      "group": "布局与操作",
+      "item": "播放、暂停、逐帧",
+      "status": "代码核对",
+      "detail": "Space与左右键、30fps步进、片尾停止；前批浏览器已验证，本轮保留代码核对口径",
+      "evidence": "SmartEditEditor/EditPreview"
+    },
+    {
+      "id": "SE-022",
+      "group": "布局与操作",
+      "item": "吸附默认值",
+      "status": "通过",
+      "detail": "参考关闭，本地SmartEditEditor也是false；纠正旧文档“本地开启”误记",
+      "evidence": "参考AX/SmartEditEditor:state"
+    },
+    {
+      "id": "SE-023",
+      "group": "布局与操作",
+      "item": "退出与Esc",
+      "status": "差异",
+      "detail": "参考Esc退出整个编辑器；本地Esc只关闭菜单/参数和取消选中",
+      "evidence": "前批参考UI/本轮源码"
+    },
+    {
+      "id": "SE-024",
+      "group": "布局与操作",
+      "item": "窄屏可用性",
+      "status": "待验",
+      "detail": "参考1024宽度测得左右仍372/300；本地三栏与工具条更窄窗口尚无完整可用性验收",
+      "evidence": "reference-transition-1024"
+    },
+    {
+      "id": "SE-025",
+      "group": "视频片段",
+      "item": "四角缩放、移动与重置",
+      "status": "通过",
+      "detail": "本轮真实合成像素验证；本地有选框和重置，参考四角手柄可见",
+      "evidence": "本轮FFmpeg/参考video"
+    },
+    {
+      "id": "SE-026",
+      "group": "视频片段",
+      "item": "变速上下调步长",
+      "status": "已修复",
+      "detail": "参考1.00→1.01，本地0.05改为0.01；浏览器已复测",
+      "evidence": "本轮UI"
+    },
+    {
+      "id": "SE-027",
+      "group": "视频片段",
+      "item": "变速支持范围",
+      "status": "待验",
+      "detail": "本地0.5–2倍；参考未穷尽上下限，不能因步长一致声称范围一致",
+      "evidence": "contract/参考video"
+    },
+    {
+      "id": "SE-028",
+      "group": "视频片段",
+      "item": "时长输入",
+      "status": "代码核对",
+      "detail": "按原素材剩余长度和变速约束，不能超过素材",
+      "evidence": "SmartEditEditor/contract"
+    },
+    {
+      "id": "SE-029",
+      "group": "视频片段",
+      "item": "数值越界后的显示",
+      "status": "已修复",
+      "detail": "到上限2后再输入99，显示回到2.00，不残留非法数字",
+      "evidence": "本轮UI"
+    },
+    {
+      "id": "SE-030",
+      "group": "视频片段",
+      "item": "增益步长",
+      "status": "已修复",
+      "detail": "参考0.5dB，本地由1改0.5；浏览器0→-0.5确认",
+      "evidence": "本轮UI"
+    },
+    {
+      "id": "SE-031",
+      "group": "视频片段",
+      "item": "正增益",
+      "status": "差异",
+      "detail": "参考能到+0.5dB；本地最大0dB，预览及合同最大gain=1",
+      "evidence": "参考UI/contract/EditPreview"
+    },
+    {
+      "id": "SE-032",
+      "group": "视频片段",
+      "item": "边缘裁剪和排序",
+      "status": "代码核对",
+      "detail": "已有按source时间/速度换算和撤销；本轮未重新拖遍所有方向",
+      "evidence": "SmartEditEditor"
+    },
+    {
+      "id": "SE-033",
+      "group": "视频片段",
+      "item": "分割和左右裁剪",
+      "status": "通过",
+      "detail": "边界禁用、源范围及片长合同回归通过；本轮未对拍参考破坏性编辑",
+      "evidence": "本轮helpers/SmartEditEditor"
+    },
+    {
+      "id": "SE-034",
+      "group": "视频片段",
+      "item": "至少保留一个片段",
+      "status": "待验",
+      "detail": "本地阻止删最后一段；参考空时间轴行为未删除原素材验证",
+      "evidence": "SmartEditEditor"
+    },
+    {
+      "id": "SE-035",
+      "group": "视频片段",
+      "item": "右键菜单",
+      "status": "差异",
+      "detail": "本地缺人声分离子菜单，增加撤销/重做/设置；“删除片段”与参考“删除”不同",
+      "evidence": "本轮参考及本地UI"
+    },
+    {
+      "id": "SE-036",
+      "group": "视频片段",
+      "item": "转场覆盖区右键",
+      "status": "差异",
+      "detail": "本地右键落在覆盖区时显示时间轴菜单，不能直接选中底下视频；从未覆盖位置可正常分离",
+      "evidence": "本轮本地UI"
+    },
+    {
+      "id": "SE-037",
+      "group": "视频片段",
+      "item": "静音后恢复原增益",
+      "status": "差异",
+      "detail": "本地静音后恢复统一gain=1，会丢失原先降低的增益；独立音轨也同样处理",
+      "evidence": "SmartEditEditor静音处理"
+    },
+    {
+      "id": "SE-038",
+      "group": "字幕",
+      "item": "空列表、新建、搜索",
+      "status": "通过",
+      "detail": "参考空列表和新建已操作；本地新建成功、搜索有真实过滤",
+      "evidence": "本轮UI/源码"
+    },
+    {
+      "id": "SE-039",
+      "group": "字幕",
+      "item": "列表逐条新增与删除",
+      "status": "差异",
+      "detail": "参考行内有“在后面新增字幕/删除字幕”；本地只有顶端新建、右键删除",
+      "evidence": "reference-subtitle-list/源码"
+    },
+    {
+      "id": "SE-040",
+      "group": "字幕",
+      "item": "列表时间格式",
+      "status": "差异",
+      "detail": "参考00:00–00:02；本地数字秒输入，布局与交互不同",
+      "evidence": "参考截图/本轮本地UI"
+    },
+    {
+      "id": "SE-041",
+      "group": "字幕",
+      "item": "选中文本、编辑与位置",
+      "status": "通过",
+      "detail": "本轮本地样式编辑和实际字幕位置像素测试通过",
+      "evidence": "本轮UI/FFmpeg"
+    },
+    {
+      "id": "SE-042",
+      "group": "字幕",
+      "item": "显示和隐藏",
+      "status": "代码核对",
+      "detail": "subtitles控制预览和导出；关闭面板仅隐藏编辑轨道",
+      "evidence": "SmartEditEditor/compiledSubtitles"
+    },
+    {
+      "id": "SE-043",
+      "group": "字幕",
+      "item": "字体搜索和长列表",
+      "status": "差异",
+      "detail": "参考可搜索Source Han、Noto、Roboto、Inter等；本地单一select且onChange为空",
+      "evidence": "reference-font-menu/源码"
+    },
+    {
+      "id": "SE-044",
+      "group": "字幕",
+      "item": "字体分类",
+      "status": "差异",
+      "detail": "参考全部/常用/推荐/Google字体；本地无分类",
+      "evidence": "reference-font-categories"
+    },
+    {
+      "id": "SE-045",
+      "group": "字幕",
+      "item": "预览与导出的字形",
+      "status": "差异",
+      "detail": "本地浏览器依赖系统字体回退，导出固定Noto CJK；不能声明与参考字体逐字形一致",
+      "evidence": "EditPreview/subtitlesToAss"
+    },
+    {
+      "id": "SE-046",
+      "group": "字幕",
+      "item": "字号范围和显示单位",
+      "status": "差异",
+      "detail": "参考默认56、12–72；本地默认28、8–160且以720为基准，需统一单位再调整",
+      "evidence": "参考DOM/contract"
+    },
+    {
+      "id": "SE-047",
+      "group": "字幕",
+      "item": "字间距范围和控件",
+      "status": "差异",
+      "detail": "参考X数值框0–100；本地滑条与数值框-5–30",
+      "evidence": "参考DOM/本地UI"
+    },
+    {
+      "id": "SE-048",
+      "group": "字幕",
+      "item": "字体颜色和十六进制",
+      "status": "代码核对",
+      "detail": "两处均有取色和hex；本地失焦校验非法值；样式烧录已回归",
+      "evidence": "StyleColor/FFmpeg"
+    },
+    {
+      "id": "SE-049",
+      "group": "字幕",
+      "item": "六种预设入口",
+      "status": "通过",
+      "detail": "默认、白黑、黑白、黄黑、白粉、粉黑均有",
+      "evidence": "本轮参考及本地UI"
+    },
+    {
+      "id": "SE-050",
+      "group": "字幕",
+      "item": "预设按钮外观",
+      "status": "差异",
+      "detail": "本地六个按钮都显示T，参考默认项显示“默认”，间距和选中态仍不同",
+      "evidence": "参考subtitle-style/源码"
+    },
+    {
+      "id": "SE-051",
+      "group": "字幕",
+      "item": "描边颜色和粗细",
+      "status": "差异",
+      "detail": "参考默认白色、粗细0–20；本地默认黑色、0–10；颜色选择可用",
+      "evidence": "参考DOM/contract"
+    },
+    {
+      "id": "SE-052",
+      "group": "字幕",
+      "item": "描边不透明度语义",
+      "status": "已修复",
+      "detail": "独立outlineOpacity；0时保留文字，只去描边；旧整行opacity保持兼容",
+      "evidence": "UI截图/ASS合同/实际像素测试"
+    },
+    {
+      "id": "SE-053",
+      "group": "字幕",
+      "item": "字号等数值单位装饰",
+      "status": "差异",
+      "detail": "参考px/%以及上下箭头；本地多为原生数字框，尺寸及结构未完全一致",
+      "evidence": "参考subtitle-style/StyleNumber"
+    },
+    {
+      "id": "SE-054",
+      "group": "字幕",
+      "item": "应用全部",
+      "status": "代码核对",
+      "detail": "复制完整style含新outlineOpacity；本轮未覆盖所有字幕数量组合",
+      "evidence": "SmartEditEditor"
+    },
+    {
+      "id": "SE-055",
+      "group": "字幕",
+      "item": "自动字幕与长转场",
+      "status": "已修复",
+      "detail": "居中转场不再把后片段字幕向后延迟整个转场时长，避免片头字幕丢失",
+      "evidence": "新增合同回归"
+    },
+    {
+      "id": "SE-056",
+      "group": "字幕",
+      "item": "手工字幕与SRT",
+      "status": "通过",
+      "detail": "手工字幕保留位置和样式，SRT仅保留文本时间；ASS字符转义回归通过",
+      "evidence": "本轮helpers/FFmpeg"
+    },
+    {
+      "id": "SE-057",
+      "group": "字幕",
+      "item": "语音自动转写",
+      "status": "差异",
+      "detail": "已有能力探测，但当前本地测试服务asr=false；完整转写尚无可用验收结果",
+      "evidence": "本地capabilities/已有部署记录"
+    },
+    {
+      "id": "SE-058",
+      "group": "转场",
+      "item": "三种基础转场入口",
+      "status": "通过",
+      "detail": "淡入淡出、黑场过渡、白场过渡；这不是全站视频生成特效库",
+      "evidence": "本轮参考UI"
+    },
+    {
+      "id": "SE-059",
+      "group": "转场",
+      "item": "原始动态素材",
+      "status": "通过",
+      "detail": "直接使用参考原始WebP，每张75帧/3秒；本地动画可见",
+      "evidence": "public/transitions/本轮UI"
+    },
+    {
+      "id": "SE-060",
+      "group": "转场",
+      "item": "分类栏与卡片位置",
+      "status": "已修复",
+      "detail": "分类48px高、13px字体，卡片108×130，图框100×100，列距7px",
+      "evidence": "参考DOM/本地CSS"
+    },
+    {
+      "id": "SE-061",
+      "group": "转场",
+      "item": "选中边框和颜色",
+      "status": "已修复",
+      "detail": "2px改1px，避免图片缩小；边框/文字改#09caf5，圆角6px",
+      "evidence": "参考DOM/本地截图"
+    },
+    {
+      "id": "SE-062",
+      "group": "转场",
+      "item": "时长与名称行间距",
+      "status": "已修复",
+      "detail": "改16px行高，时长框与90px小预览纵向位置对齐",
+      "evidence": "参考DOM/本地CSS"
+    },
+    {
+      "id": "SE-063",
+      "group": "转场",
+      "item": "时长上下限与步长",
+      "status": "通过",
+      "detail": "0.2–4秒，0.1步进，非法值失焦归一化",
+      "evidence": "前批UI/本轮合同"
+    },
+    {
+      "id": "SE-064",
+      "group": "转场",
+      "item": "相邻素材小预览",
+      "status": "通过",
+      "detail": "本地使用相邻真实视频，静音循环并应用效果",
+      "evidence": "本轮UI"
+    },
+    {
+      "id": "SE-065",
+      "group": "转场",
+      "item": "主画面预览时机",
+      "status": "待验",
+      "detail": "本地选中转场会循环预览；参考本轮主画面多保持原播放头，小预览循环，触发规则需进一步对拍",
+      "evidence": "reference-transition/TransitionPreview"
+    },
+    {
+      "id": "SE-066",
+      "group": "转场",
+      "item": "居中时序和总片长",
+      "status": "通过",
+      "detail": "三种2.2/4秒真实成片均保留4秒总长",
+      "evidence": "本轮FFmpeg"
+    },
+    {
+      "id": "SE-067",
+      "group": "转场",
+      "item": "黑场与白场中点",
+      "status": "通过",
+      "detail": "真实像素验证纯黑/纯白中点",
+      "evidence": "本轮FFmpeg"
+    },
+    {
+      "id": "SE-068",
+      "group": "转场",
+      "item": "极短片段和连续接点",
+      "status": "通过",
+      "detail": "短片段约束和避免三片同时重叠的合同通过",
+      "evidence": "本轮helpers/前批压力合成"
+    },
+    {
+      "id": "SE-069",
+      "group": "转场",
+      "item": "删除与硬切",
+      "status": "通过",
+      "detail": "已有删除恢复硬切；本轮逐片段与硬切组合实际合成通过",
+      "evidence": "FFmpeg/源码"
+    },
+    {
+      "id": "SE-070",
+      "group": "转场",
+      "item": "时间轴覆盖框",
+      "status": "差异",
+      "detail": "参考top=4px、高60px；本地top=0、高64px，选中填充/图标仍需精调",
+      "evidence": "参考DOM/本地CSS"
+    },
+    {
+      "id": "SE-071",
+      "group": "转场",
+      "item": "旧方案兼容",
+      "status": "通过",
+      "detail": "未设置alignment的旧重叠方案保留原片长语义",
+      "evidence": "本轮helpers/FFmpeg"
+    },
+    {
+      "id": "SE-072",
+      "group": "音频",
+      "item": "有声视频分离",
+      "status": "通过",
+      "detail": "本轮真实点按生成独立音轨，撤销后消失",
+      "evidence": "本轮本地UI"
+    },
+    {
+      "id": "SE-073",
+      "group": "音频",
+      "item": "无音轨时分离禁用",
+      "status": "通过",
+      "detail": "参考当前无音轨素材禁用；本地按hasAudio禁用",
+      "evidence": "本轮参考UI/源码"
+    },
+    {
+      "id": "SE-074",
+      "group": "音频",
+      "item": "分离后声音不叠加",
+      "status": "通过",
+      "detail": "实际音频RMS与转场包络回归通过",
+      "evidence": "本轮FFmpeg"
+    },
+    {
+      "id": "SE-075",
+      "group": "音频",
+      "item": "独立音轨变速、淡入淡出",
+      "status": "通过",
+      "detail": "实际2倍速、增益、延迟、淡入淡出及长音轨黑尾回归通过",
+      "evidence": "本轮FFmpeg"
+    },
+    {
+      "id": "SE-076",
+      "group": "音频",
+      "item": "独立音轨裁剪、分割、移动",
+      "status": "代码核对",
+      "detail": "有完整source范围与时间轴处理；前批UI通过，本轮未全部重拖",
+      "evidence": "SmartEditEditor/contract"
+    },
+    {
+      "id": "SE-077",
+      "group": "音频",
+      "item": "人声分离两个分支",
+      "status": "差异",
+      "detail": "参考“仅保留人声/仅保留背景音”子菜单；本地未接专用分离服务",
+      "evidence": "reference-video-menu"
+    },
+    {
+      "id": "SE-078",
+      "group": "音频",
+      "item": "音频参数栏调整宽度",
+      "status": "差异",
+      "detail": "本地视频/字幕/转场栏有拖拽柄，音频栏缺同类柄；参考有声参数栏未验",
+      "evidence": "SmartEditEditor"
+    },
+    {
+      "id": "SE-079",
+      "group": "音频",
+      "item": "真实波形",
+      "status": "待验",
+      "detail": "本地音频条是纯色色块，没有采样波形；本轮未取得参考有声轨波形证据",
+      "evidence": "se-wave/se-audio-clip"
+    },
+    {
+      "id": "SE-080",
+      "group": "音频",
+      "item": "音乐与卡点",
+      "status": "通过",
+      "detail": "节拍与候选剪辑合同回归通过，音乐混音真实输出通过；不是参考算法效果对拍",
+      "evidence": "本轮helpers/FFmpeg"
+    },
+    {
+      "id": "SE-081",
+      "group": "导出与恢复",
+      "item": "导出菜单位置",
+      "status": "通过",
+      "detail": "本地/画布两个目的地和参考相同",
+      "evidence": "参考export-menu/源码"
+    },
+    {
+      "id": "SE-082",
+      "group": "导出与恢复",
+      "item": "导出菜单视觉",
+      "status": "代码核对",
+      "detail": "本地有200px菜单、40px条目；边框/图标需独立像素精调",
+      "evidence": "参考截图/CSS"
+    },
+    {
+      "id": "SE-083",
+      "group": "导出与恢复",
+      "item": "真实本地合成",
+      "status": "通过",
+      "detail": "FFmpeg实际生成结果，不用模拟成功替代",
+      "evidence": "本轮ffmpeg.log"
+    },
+    {
+      "id": "SE-084",
+      "group": "导出与恢复",
+      "item": "提交任务ID和幂等",
+      "status": "通过",
+      "detail": "请求ID回归与远程确定性任务身份测试通过",
+      "evidence": "本轮helpers"
+    },
+    {
+      "id": "SE-085",
+      "group": "导出与恢复",
+      "item": "刷新恢复导出任务",
+      "status": "通过",
+      "detail": "任务检查点/reconcile合同通过；本轮未执行真实Next页面断网恢复",
+      "evidence": "本轮helpers"
+    },
+    {
+      "id": "SE-086",
+      "group": "导出与恢复",
+      "item": "浏览器中返回画布结果",
+      "status": "待验",
+      "detail": "独立测试页任务回调仍是样例；本轮未跑主站Next API全链路",
+      "evidence": "ui-entry/onRender"
+    },
+    {
+      "id": "SE-087",
+      "group": "导出与恢复",
+      "item": "任务错误和服务不可用",
+      "status": "代码核对",
+      "detail": "导出按钮按能力禁用，失败状态显示错误；本轮未切断实际服务故障注入",
+      "evidence": "SmartEditEditor/SmartEditDialog"
+    },
+    {
+      "id": "SE-088",
+      "group": "导出与恢复",
+      "item": "取消远程导出",
+      "status": "差异",
+      "detail": "没有主动取消运行中合成任务的完整UI/API链路；本轮参考菜单也未验证取消",
+      "evidence": "源码"
+    },
+    {
+      "id": "SE-089",
+      "group": "导出与恢复",
+      "item": "多版本全导出",
+      "status": "代码核对",
+      "detail": "高级设置有导出全部版本；主节点默认只一版，参考完整多版结果未生成",
+      "evidence": "SmartEditDialog"
+    },
+    {
+      "id": "SE-090",
+      "group": "导出与恢复",
+      "item": "最高时长与比例",
+      "status": "通过",
+      "detail": "300秒约束、宽屏远程结果下载合同、6比例真实本地输出通过",
+      "evidence": "本轮helpers/FFmpeg"
+    },
+    {
+      "id": "SE-091",
+      "group": "边界与性能",
+      "item": "最大片段数",
+      "status": "代码核对",
+      "detail": "最多12视频片段、24独立音轨、500字幕；不是参考站已确认相同限制",
+      "evidence": "contract"
+    },
+    {
+      "id": "SE-092",
+      "group": "边界与性能",
+      "item": "后台标签恢复",
+      "status": "待验",
+      "detail": "循环预览监听visibilitychange；本轮未覆盖系统休眠/后台节流完整矩阵",
+      "evidence": "TransitionPreview"
+    },
+    {
+      "id": "SE-093",
+      "group": "边界与性能",
+      "item": "原生媒体加载失败",
+      "status": "代码核对",
+      "detail": "play的AbortError忽略，其余报错可见；本轮未测试全部外部域名/CORS",
+      "evidence": "EditPreview"
+    },
+    {
+      "id": "SE-094",
+      "group": "边界与性能",
+      "item": "高清与12片段性能",
+      "status": "既有证据",
+      "detail": "前批1080P4秒合成12.81秒、12片段7.604秒；本轮未重复此压力测试",
+      "evidence": "transition-stress-native.log"
+    },
+    {
+      "id": "SE-095",
+      "group": "边界与性能",
+      "item": "服务部署状态",
+      "status": "通过",
+      "detail": "本轮所有修改仅本地，未更新48或算力服务器",
+      "evidence": "git/操作记录"
+    }
+  ]
+}

docs/smart-edit-full-recheck-2026-09-17.md

@@ -0,0 +1,208 @@
+# 智能剪辑全面复核与像素差异（2026-09-17）
+
+> 后续修复已更新至 [2026-09-18记录](smart-edit-fixes-2026-09-18.md)。本文保留09-17历史审计状态。
+
+本轮在基线`431ca03f`之上复核 **95个检查点**，覆盖智能剪辑节点输入、编辑器布局、视频、字幕、转场、音频、导出恢复和边界。结论：**仍有遗漏，不能验收为全部功能齐全或100%像素一致**。这轮继续只在本地修复和验证，没有更新服务器。
+
+该范围承接本次“转场预览图”对话，不扩大为整个视频画布、3D导演台或参考站全站重新验收；它们的旧清单不能被本轮结果覆盖。各检查点区别本轮UI、真实合成、代码核对、既有证据和待验，避免把功能存在误写为全链路成功。
+
+## 结果分布
+
+| 状态 | 数量 |
+| --- | ---: |
+| 通过 | 35 |
+| 待验 | 8 |
+| 代码核对 | 17 |
+| 差异 | 25 |
+| 已修复 | 9 |
+| 既有证据 | 1 |
+
+“通过”仅对该行明确的测试边界成立，不能换算为全站完成率。“差异”包含缺失能力、交互不同和视觉偏差，不代表同等严重程度。
+
+## 本轮修复
+
+- 自动字幕时序：居中转场不能按旧重叠转场推迟后片段字幕，否则片头字幕会被丢弃。
+- 独立描边透明度：新增可选`outlineOpacity`，预览和ASS输出保持一致；旧`opacity`仍保持历史整行透明度语义，避免改坏已保存工程。
+- 变速按钮步长改0.01、音量步长改0.5dB；输入超上限后立即显示约束后的数值，包括已经到上限再输入的情况。
+- 转场分类栏、卡片内边距、圆角、选中蓝色、边框粗细、名称/时长行高及参数面板背景、关闭按钮尺寸按参考可见DOM修正。
+
+## 像素测量
+
+在参考站1331×681、1512×752和1024×768 CSS视口测量可见DOM。独立本地页也设置过1331×681设备视口检查三栏；最终宽屏截图来自本机原生浏览器。原生截图包含浏览器外壳且经过显示缩放，汇总图裁去顶部106px并将参考图归一到1331×662显示尺寸。为排除面板状态差异，宽屏对照临时把参考时间轴增高25px后截图，随后恢复。
+
+这些图用于布局叠位和控件对照，不计算混入不同视频、不同动画帧、系统字体抗锯齿的整屏“相似度百分比”。严格逐像素验收仍需要固定同一视频帧、字体文件、DPR和全部交互状态；本轮没有宣称完成这种全矩阵黄金截图验收。
+
+| 控件 | 参考测量 | 本轮本地修正后 | 结论 |
+| --- | --- | --- | --- |
+| 顶栏 | 48px | 48px | 一致 |
+| 页面边距/栏间距 | 8px/8px | 8px/8px | 一致 |
+| 转场库/参数栏 | 372px/300px | 372px/300px | 一致 |
+| 参数面板 | #202020、白10%边框、12px圆角 | 相同CSS值 | 修正 |
+| 面板关闭按钮 | 24×24px | 24×24px | 修正 |
+| 分类栏 | 48px高、13px字 | 相同 | 修正 |
+| 卡片 | 默认栏宽下108×130px，列距7px | 相同盒模型 | 修正 |
+| 动态图框 | 100×100px，1px边框 | 选中也保留1px，内图不缩小 | 修正 |
+| 选中边框/文字 | #09caf5 | #09caf5 | 修正 |
+| 卡片圆角/文字内边距 | 6px / 左右4px | 相同 | 修正 |
+| 名称/时长文字 | 12px/16px行高 | 相同 | 修正 |
+| 右侧实际预览 | 90px高 | 90px | 一致 |
+| 转场轨道覆盖框 | top4、高60px | top0、高64px | 仍有差异 |
+| 时间轴缩放数值范围 | 0–100、默认20 | 0–40、默认20 | 范围不同；尚未对齐最大倍率与刻度密度 |
+| 字幕数值控件 | px/%装饰、自绘步进、不同字号范围 | 原生数值输入、旧720基准 | 仍有差异 |
+
+## 尚未完成的优先事项
+
+1. 字体库、搜索/分类、真实字体文件与预览/导出映射；字号、字间距、描边粗细的单位及范围。
+2. 正音量增益、静音恢复原增益、人声/背景音分离。
+3. 字幕列表逐条新增/删除、时间格式与样式控件；折叠预览露出画布。
+4. 节点缩略时间轴绑定实际计划；节点卡点需求到音乐/节拍设置的联动；批量广告版本设置。
+5. 转场覆盖框和时间轴缩放细节，长转场覆盖区域的右键交互，以及主画面转场演示触发规则对拍。
+6. 主站端到端导出、断网/后台恢复、有声参考轨更多功能、窄窗口完整验收；目前不能用本地样例任务回调代替这些结果。
+
+## 逐项记录
+
+
+### 节点与输入
+
+| ID | 检查项 | 状态 | 结论/范围 | 证据 |
+| --- | --- | --- | --- | --- |
+| SE-001 | 打开剪辑 | 通过 | 参考和本地均有入口；本地进入真实编辑组件 | 本轮UI |
+| SE-002 | 五种模式名称和顺序 | 通过 | 默认、讲解、批量广告、口播、素材混剪相同 | 参考node-modes截图/SmartEditNode |
+| SE-003 | 五种模式算法效果 | 待验 | 本地有模式分支；没有重新提交参考站付费生成，不能认定效果等同 | execute/contract |
+| SE-004 | 六种比例 | 通过 | 21:9、16:9、9:16、1:1、4:3、3:4；实际合成覆盖 | 参考node-settings/本轮FFmpeg |
+| SE-005 | 三种清晰度 | 通过 | 480P、720P、1080P；控件一致 | 参考node-settings/SmartEditNode |
+| SE-006 | 时长范围 | 通过 | 0–300秒、步长1秒；本地0为自动 | 参考可见DOM/contract |
+| SE-007 | 提示词编辑与展开 | 代码核对 | 本地受控输入、展开及提交；参考既有提示词未改动 | SmartEditNode |
+| SE-008 | 引用画布素材 | 代码核对 | 参考进入引用选择；生产组件onReference连接连线状态，本地独立测试页只是候选列表 | 参考引用模式/主画布调用点 |
+| SE-009 | 素材序号、时长、移除 | 代码核对 | 本地有缩略图、序号、时长及移除；未删除参考原素材 | SmartEditNode |
+| SE-010 | 本地上传与数量限制 | 代码核对 | 真实上传入口，上限12个素材、单个125MiB；参考上限尚未验证 | SmartEditNode/SmartEditDialog |
+| SE-011 | 节点缩略时间轴 | 差异 | 本地节点卡片仍固定生成12个刻度、两行8个色块，未绑定实际剪辑轨道 | SmartEditNodeCard |
+| SE-012 | 批量广告多版本 | 差异 | 本地高级设置支持1–3版，节点默认variants为1；候选顺序变体不等于多套广告创意 | contract/buildEditPlans |
+| SE-013 | 提示词要求卡点 | 差异 | 真正卡点要求音乐和beatSync；仅在提示词写“卡点”不会自动打开节拍流程 | execute/主画布onSubmit |
+
+### 布局与操作
+
+| ID | 检查项 | 状态 | 结论/范围 | 证据 |
+| --- | --- | --- | --- | --- |
+| SE-014 | 顶栏高度和边距 | 通过 | 48px顶栏、8px页面边距；本轮图像对照 | 参考DOM/本地CSS |
+| SE-015 | 左右参数栏宽度 | 通过 | 转场库372px、参数栏300px，均可拖动调整 | 参考DOM/本地CSS |
+| SE-016 | 面板背景与关闭按钮 | 已修复 | 背景统一#202020，边框白色10%，关闭按钮24px，标题500/14px | 参考transition.json/本地CSS |
+| SE-017 | 默认时间轴高度 | 差异 | 打开时均约35vh；参考改变视口保留原像素高度，本地未手动拖动时随窗口重算 | 本轮宽屏切换/resize处理 |
+| SE-018 | 折叠预览 | 差异 | 参考保留底部时间轴并露出画布；本地把时间轴铺满编辑器且保留顶栏 | 参考collapsed截图/SmartEditEditor |
+| SE-019 | 全屏预览 | 代码核对 | 本地调用Fullscreen API并显示拒绝错误；本轮未重测不同浏览器权限状态 | SmartEditEditor |
+| SE-020 | 撤销重做 | 通过 | 本轮分离音轨后撤销成功；参考宽屏有按钮、窄屏隐藏，先前未见不代表没有 | 本轮UI/参考1512宽度 |
+| SE-021 | 播放、暂停、逐帧 | 代码核对 | Space与左右键、30fps步进、片尾停止；前批浏览器已验证，本轮保留代码核对口径 | SmartEditEditor/EditPreview |
+| SE-022 | 吸附默认值 | 通过 | 参考关闭，本地SmartEditEditor也是false；纠正旧文档“本地开启”误记 | 参考AX/SmartEditEditor:state |
+| SE-023 | 退出与Esc | 差异 | 参考Esc退出整个编辑器；本地Esc只关闭菜单/参数和取消选中 | 前批参考UI/本轮源码 |
+| SE-024 | 窄屏可用性 | 待验 | 参考1024宽度测得左右仍372/300；本地三栏与工具条更窄窗口尚无完整可用性验收 | reference-transition-1024 |
+
+### 视频片段
+
+| ID | 检查项 | 状态 | 结论/范围 | 证据 |
+| --- | --- | --- | --- | --- |
+| SE-025 | 四角缩放、移动与重置 | 通过 | 本轮真实合成像素验证；本地有选框和重置，参考四角手柄可见 | 本轮FFmpeg/参考video |
+| SE-026 | 变速上下调步长 | 已修复 | 参考1.00→1.01，本地0.05改为0.01；浏览器已复测 | 本轮UI |
+| SE-027 | 变速支持范围 | 待验 | 本地0.5–2倍；参考未穷尽上下限，不能因步长一致声称范围一致 | contract/参考video |
+| SE-028 | 时长输入 | 代码核对 | 按原素材剩余长度和变速约束，不能超过素材 | SmartEditEditor/contract |
+| SE-029 | 数值越界后的显示 | 已修复 | 到上限2后再输入99，显示回到2.00，不残留非法数字 | 本轮UI |
+| SE-030 | 增益步长 | 已修复 | 参考0.5dB，本地由1改0.5；浏览器0→-0.5确认 | 本轮UI |
+| SE-031 | 正增益 | 差异 | 参考能到+0.5dB；本地最大0dB，预览及合同最大gain=1 | 参考UI/contract/EditPreview |
+| SE-032 | 边缘裁剪和排序 | 代码核对 | 已有按source时间/速度换算和撤销；本轮未重新拖遍所有方向 | SmartEditEditor |
+| SE-033 | 分割和左右裁剪 | 通过 | 边界禁用、源范围及片长合同回归通过；本轮未对拍参考破坏性编辑 | 本轮helpers/SmartEditEditor |
+| SE-034 | 至少保留一个片段 | 待验 | 本地阻止删最后一段；参考空时间轴行为未删除原素材验证 | SmartEditEditor |
+| SE-035 | 右键菜单 | 差异 | 本地缺人声分离子菜单，增加撤销/重做/设置；“删除片段”与参考“删除”不同 | 本轮参考及本地UI |
+| SE-036 | 转场覆盖区右键 | 差异 | 本地右键落在覆盖区时显示时间轴菜单，不能直接选中底下视频；从未覆盖位置可正常分离 | 本轮本地UI |
+| SE-037 | 静音后恢复原增益 | 差异 | 本地静音后恢复统一gain=1，会丢失原先降低的增益；独立音轨也同样处理 | SmartEditEditor静音处理 |
+
+### 字幕
+
+| ID | 检查项 | 状态 | 结论/范围 | 证据 |
+| --- | --- | --- | --- | --- |
+| SE-038 | 空列表、新建、搜索 | 通过 | 参考空列表和新建已操作；本地新建成功、搜索有真实过滤 | 本轮UI/源码 |
+| SE-039 | 列表逐条新增与删除 | 差异 | 参考行内有“在后面新增字幕/删除字幕”；本地只有顶端新建、右键删除 | reference-subtitle-list/源码 |
+| SE-040 | 列表时间格式 | 差异 | 参考00:00–00:02；本地数字秒输入，布局与交互不同 | 参考截图/本轮本地UI |
+| SE-041 | 选中文本、编辑与位置 | 通过 | 本轮本地样式编辑和实际字幕位置像素测试通过 | 本轮UI/FFmpeg |
+| SE-042 | 显示和隐藏 | 代码核对 | subtitles控制预览和导出；关闭面板仅隐藏编辑轨道 | SmartEditEditor/compiledSubtitles |
+| SE-043 | 字体搜索和长列表 | 差异 | 参考可搜索Source Han、Noto、Roboto、Inter等；本地单一select且onChange为空 | reference-font-menu/源码 |
+| SE-044 | 字体分类 | 差异 | 参考全部/常用/推荐/Google字体；本地无分类 | reference-font-categories |
+| SE-045 | 预览与导出的字形 | 差异 | 本地浏览器依赖系统字体回退，导出固定Noto CJK；不能声明与参考字体逐字形一致 | EditPreview/subtitlesToAss |
+| SE-046 | 字号范围和显示单位 | 差异 | 参考默认56、12–72；本地默认28、8–160且以720为基准，需统一单位再调整 | 参考DOM/contract |
+| SE-047 | 字间距范围和控件 | 差异 | 参考X数值框0–100；本地滑条与数值框-5–30 | 参考DOM/本地UI |
+| SE-048 | 字体颜色和十六进制 | 代码核对 | 两处均有取色和hex；本地失焦校验非法值；样式烧录已回归 | StyleColor/FFmpeg |
+| SE-049 | 六种预设入口 | 通过 | 默认、白黑、黑白、黄黑、白粉、粉黑均有 | 本轮参考及本地UI |
+| SE-050 | 预设按钮外观 | 差异 | 本地六个按钮都显示T，参考默认项显示“默认”，间距和选中态仍不同 | 参考subtitle-style/源码 |
+| SE-051 | 描边颜色和粗细 | 差异 | 参考默认白色、粗细0–20；本地默认黑色、0–10；颜色选择可用 | 参考DOM/contract |
+| SE-052 | 描边不透明度语义 | 已修复 | 独立outlineOpacity；0时保留文字，只去描边；旧整行opacity保持兼容 | UI截图/ASS合同/实际像素测试 |
+| SE-053 | 字号等数值单位装饰 | 差异 | 参考px/%以及上下箭头；本地多为原生数字框，尺寸及结构未完全一致 | 参考subtitle-style/StyleNumber |
+| SE-054 | 应用全部 | 代码核对 | 复制完整style含新outlineOpacity；本轮未覆盖所有字幕数量组合 | SmartEditEditor |
+| SE-055 | 自动字幕与长转场 | 已修复 | 居中转场不再把后片段字幕向后延迟整个转场时长，避免片头字幕丢失 | 新增合同回归 |
+| SE-056 | 手工字幕与SRT | 通过 | 手工字幕保留位置和样式，SRT仅保留文本时间；ASS字符转义回归通过 | 本轮helpers/FFmpeg |
+| SE-057 | 语音自动转写 | 差异 | 已有能力探测，但当前本地测试服务asr=false；完整转写尚无可用验收结果 | 本地capabilities/已有部署记录 |
+
+### 转场
+
+| ID | 检查项 | 状态 | 结论/范围 | 证据 |
+| --- | --- | --- | --- | --- |
+| SE-058 | 三种基础转场入口 | 通过 | 淡入淡出、黑场过渡、白场过渡；这不是全站视频生成特效库 | 本轮参考UI |
+| SE-059 | 原始动态素材 | 通过 | 直接使用参考原始WebP，每张75帧/3秒；本地动画可见 | public/transitions/本轮UI |
+| SE-060 | 分类栏与卡片位置 | 已修复 | 分类48px高、13px字体，卡片108×130，图框100×100，列距7px | 参考DOM/本地CSS |
+| SE-061 | 选中边框和颜色 | 已修复 | 2px改1px，避免图片缩小；边框/文字改#09caf5，圆角6px | 参考DOM/本地截图 |
+| SE-062 | 时长与名称行间距 | 已修复 | 改16px行高，时长框与90px小预览纵向位置对齐 | 参考DOM/本地CSS |
+| SE-063 | 时长上下限与步长 | 通过 | 0.2–4秒，0.1步进，非法值失焦归一化 | 前批UI/本轮合同 |
+| SE-064 | 相邻素材小预览 | 通过 | 本地使用相邻真实视频，静音循环并应用效果 | 本轮UI |
+| SE-065 | 主画面预览时机 | 待验 | 本地选中转场会循环预览；参考本轮主画面多保持原播放头，小预览循环，触发规则需进一步对拍 | reference-transition/TransitionPreview |
+| SE-066 | 居中时序和总片长 | 通过 | 三种2.2/4秒真实成片均保留4秒总长 | 本轮FFmpeg |
+| SE-067 | 黑场与白场中点 | 通过 | 真实像素验证纯黑/纯白中点 | 本轮FFmpeg |
+| SE-068 | 极短片段和连续接点 | 通过 | 短片段约束和避免三片同时重叠的合同通过 | 本轮helpers/前批压力合成 |
+| SE-069 | 删除与硬切 | 通过 | 已有删除恢复硬切；本轮逐片段与硬切组合实际合成通过 | FFmpeg/源码 |
+| SE-070 | 时间轴覆盖框 | 差异 | 参考top=4px、高60px；本地top=0、高64px，选中填充/图标仍需精调 | 参考DOM/本地CSS |
+| SE-071 | 旧方案兼容 | 通过 | 未设置alignment的旧重叠方案保留原片长语义 | 本轮helpers/FFmpeg |
+
+### 音频
+
+| ID | 检查项 | 状态 | 结论/范围 | 证据 |
+| --- | --- | --- | --- | --- |
+| SE-072 | 有声视频分离 | 通过 | 本轮真实点按生成独立音轨，撤销后消失 | 本轮本地UI |
+| SE-073 | 无音轨时分离禁用 | 通过 | 参考当前无音轨素材禁用；本地按hasAudio禁用 | 本轮参考UI/源码 |
+| SE-074 | 分离后声音不叠加 | 通过 | 实际音频RMS与转场包络回归通过 | 本轮FFmpeg |
+| SE-075 | 独立音轨变速、淡入淡出 | 通过 | 实际2倍速、增益、延迟、淡入淡出及长音轨黑尾回归通过 | 本轮FFmpeg |
+| SE-076 | 独立音轨裁剪、分割、移动 | 代码核对 | 有完整source范围与时间轴处理；前批UI通过，本轮未全部重拖 | SmartEditEditor/contract |
+| SE-077 | 人声分离两个分支 | 差异 | 参考“仅保留人声/仅保留背景音”子菜单；本地未接专用分离服务 | reference-video-menu |
+| SE-078 | 音频参数栏调整宽度 | 差异 | 本地视频/字幕/转场栏有拖拽柄，音频栏缺同类柄；参考有声参数栏未验 | SmartEditEditor |
+| SE-079 | 真实波形 | 待验 | 本地音频条是纯色色块，没有采样波形；本轮未取得参考有声轨波形证据 | se-wave/se-audio-clip |
+| SE-080 | 音乐与卡点 | 通过 | 节拍与候选剪辑合同回归通过，音乐混音真实输出通过；不是参考算法效果对拍 | 本轮helpers/FFmpeg |
+
+### 导出与恢复
+
+| ID | 检查项 | 状态 | 结论/范围 | 证据 |
+| --- | --- | --- | --- | --- |
+| SE-081 | 导出菜单位置 | 通过 | 本地/画布两个目的地和参考相同 | 参考export-menu/源码 |
+| SE-082 | 导出菜单视觉 | 代码核对 | 本地有200px菜单、40px条目；边框/图标需独立像素精调 | 参考截图/CSS |
+| SE-083 | 真实本地合成 | 通过 | FFmpeg实际生成结果，不用模拟成功替代 | 本轮ffmpeg.log |
+| SE-084 | 提交任务ID和幂等 | 通过 | 请求ID回归与远程确定性任务身份测试通过 | 本轮helpers |
+| SE-085 | 刷新恢复导出任务 | 通过 | 任务检查点/reconcile合同通过；本轮未执行真实Next页面断网恢复 | 本轮helpers |
+| SE-086 | 浏览器中返回画布结果 | 待验 | 独立测试页任务回调仍是样例；本轮未跑主站Next API全链路 | ui-entry/onRender |
+| SE-087 | 任务错误和服务不可用 | 代码核对 | 导出按钮按能力禁用，失败状态显示错误；本轮未切断实际服务故障注入 | SmartEditEditor/SmartEditDialog |
+| SE-088 | 取消远程导出 | 差异 | 没有主动取消运行中合成任务的完整UI/API链路；本轮参考菜单也未验证取消 | 源码 |
+| SE-089 | 多版本全导出 | 代码核对 | 高级设置有导出全部版本；主节点默认只一版，参考完整多版结果未生成 | SmartEditDialog |
+| SE-090 | 最高时长与比例 | 通过 | 300秒约束、宽屏远程结果下载合同、6比例真实本地输出通过 | 本轮helpers/FFmpeg |
+
+### 边界与性能
+
+| ID | 检查项 | 状态 | 结论/范围 | 证据 |
+| --- | --- | --- | --- | --- |
+| SE-091 | 最大片段数 | 代码核对 | 最多12视频片段、24独立音轨、500字幕；不是参考站已确认相同限制 | contract |
+| SE-092 | 后台标签恢复 | 待验 | 循环预览监听visibilitychange；本轮未覆盖系统休眠/后台节流完整矩阵 | TransitionPreview |
+| SE-093 | 原生媒体加载失败 | 代码核对 | play的AbortError忽略，其余报错可见；本轮未测试全部外部域名/CORS | EditPreview |
+| SE-094 | 高清与12片段性能 | 既有证据 | 前批1080P4秒合成12.81秒、12片段7.604秒；本轮未重复此压力测试 | transition-stress-native.log |
+| SE-095 | 服务部署状态 | 通过 | 本轮所有修改仅本地，未更新48或算力服务器 | git/操作记录 |
+
+## 验证与原始证据
+
+- TypeScript通过；智能剪辑18项回归通过；完整helpers493项通过、1项跳过。
+- 本轮真实FFmpeg回归通过：旧重叠转场、三种居中2.2/4秒转场、六比例、字幕烧录、画面变换、音轨变速/延迟/包络、描边透明度。
+- 新的描边像素检查：白字像素仍存在，原黑描边像素消失，避免只测表单值。
+- 本轮未调用参考站付费生成或重新部署服务；参考原有2.2秒白场、1倍速、0dB保持；临时字幕已从列表删除，引用模式已退出。
+- 原始截图、DOM测量、测试日志位于`.qa/smart-edit/recheck/`；`transition-side-by-side.png`为并排对照，`local-outline-opacity-zero.jpg`为描边语义验证。
+- 参考站一度在字幕列表为空后仍显示旧字幕预览，重开编辑器后消失；这不是本地持久化成功或失败的证据。
+- 机器可读逐项清单：[JSON](smart-edit-full-recheck-2026-09-17.json)。旧报告如有冲突，以本轮明确复核的条目为准。

docs/smart-edit-subfeature-audit-2026-09-17.md

@@ -0,0 +1,59 @@
+# 智能剪辑子功能再核对
+
+> 后续全面复核见[95项检查及像素差异](smart-edit-full-recheck-2026-09-17.md)。本文保留此前快照：描边透明度、步长和转场样式已在后续修复；本地吸附默认关闭，本文旧“默认开启”结论不准确。
+
+范围：2026-09-17 在用户指定的 LibTV 项目逐层展开编辑器，结合此前节点输入卡片核对记录。只报告可见界面与实际操作，不推断参考站私有实现。用户随后明确要求直接使用参考网页动态预览图。所有新增代码只在本地；继续遵守“本地完整验收后再更新服务器”。
+
+## 本轮转场差异及修复
+
+| 子功能 | 参考实测 | 本地处理 |
+| --- | --- | --- |
+| 库分类与类型 | 基础转场：淡入淡出、黑场过渡、白场过渡 | 三项保留；当前展开未见其他分类，不等同于完整特效库 |
+| 卡片素材 | 三张动态WebP，人物/场景演示；75帧、3秒 | 原始文件保存到public/video-canvas/transitions，替换原素材缩略图+渐变遮罩 |
+| 选中样式 | 灰色底、蓝色边框和文字 | 本地CSS已对应调整 |
+| 动态效果预览 | 右侧90px高预览当前相邻视频；主画面展示选中转场 | 新增TransitionPreview，复用编辑器合成时序；小预览静音循环、关闭卸载；普通播放/拖动播放头可退出效果演示 |
+| 时长输入 | 最小0.2、最大4、步长0.1秒，左右减/加按钮 | 修正旧0–0.5秒上限；输入失焦提交，避免输入小数时截断 |
+| 总片长 | 把已有白场2.2改成0.2、再恢复2.2，总片长仍14秒 | 新设置使用接点居中的转场，保留剪辑总长；旧保存方案保留原重叠语义，避免自动改旧工程 |
+| 时间轴转场区域 | 覆盖接点两侧的矩形区域 | 按实际生效时长展示整个轨道高度的区域 |
+| 极短相邻片段 | 本轮未穷尽所有参考边界 | 本地约束防止相邻转场争用同一片段，面板提示实际生效时长；不谎报完整4秒 |
+| 删除 | 右下删除，恢复片段接点 | 已保留；新转场删除不改变总片长 |
+| 实际黑白场 | 黑/白覆盖从透明达到全色，再恢复下一视频 | 修正FFmpeg内置曲线与浏览器线性遮罩差异；以两个原生fade经过纯色中间帧合成，避免慢速逐像素表达式 |
+| 分离音轨联动 | 参考源无音轨，未验证其音频细节 | 本地补齐居中转场的半程音量包络，分离后保持声音增益 |
+
+导出使用的原生滤镜依据[FFmpeg官方xfade说明](https://ffmpeg.org/ffmpeg-filters.html#xfade)，效果以本地像素采样验证，未声称取得参考站私有算法。
+
+## 其他入口逐项状态
+
+| 入口/子项 | 核对结果 | 状态 |
+| --- | --- | --- |
+| 视频预览四角缩放、移动 | 参考可见四角手柄；本地已有画面变换与实际导出验证 | 已实现，见前批验收 |
+| 视频参数：速度、时长 | 参考两项均可输入、上下调；本地已提供 | 入口已对齐，参考完整速度范围仍待边界核对 |
+| 视频音量增益 | 参考0dB可点到+0.5dB；本地最高0dB且步进1dB | **待补齐正增益及步长**，预览放大需要音频增益节点，不能只放宽表单 |
+| 视频右键：音视频分离 | 参考当前无音轨源显示禁用；本地有声源已可分离 | 已实现基础分离；参考有声源分支仍待核对 |
+| 视频右键：人声分离 | 参考可见带子菜单的入口；当前源无音轨，未执行 | **待接入/验证真实人声分离**，不放置假按钮冒充成功 |
+| 视频右键：删除 | 可见删除入口 | 本地已有，保留至少一个视频片段 |
+| 时间轴：分割、向左裁剪、向右裁剪 | 与选中片段、播放头位置联动禁用 | 本地已有，前批真实操作验证 |
+| 时间轴：缩放、吸附、折叠预览、拖动高度 | 参考工具栏可见 | 本地已有，前批浏览器验证；参考吸附默认关闭，本地默认开启，默认状态仍有差异 |
+| 字幕：显示/隐藏、列表、搜索、新建 | 参考均可展开；本轮临时新建字幕检查后已删除恢复 | 本地已有 |
+| 字幕：编辑文字、移动、时间范围 | 参考提供文字面板和起止手柄 | 本地已有，前批浏览器及实际烧录验证 |
+| 字幕：字体 | 参考有搜索框、“全部字体”、Source Han/Noto/Roboto/Inter等长列表；本地只有一个字体选项 | **待补齐可搜索字体库和导出字体映射** |
+| 字幕：字号 | 本次参考默认56，范围12–72；本地默认28、范围8–160（720基准） | **待统一显示单位和字号范围**，不能仅照抄数值导致导出大小翻倍 |
+| 字幕：字间距 | 参考0–100；本地-5–30 | **待对齐范围/单位** |
+| 字幕：颜色、六种预设 | 参考默认/白黑/黑白/黄黑/白粉/粉黑 | 本地六种已有，不能据此推导所有字体效果一致 |
+| 字幕：描边粗细 | 参考0–20；本地0–10 | **待对齐范围/单位** |
+| 字幕：描边不透明度 | 参考专门作用于描边；本地opacity作用于整行文字 | **待拆分描边透明度**，此前“透明度已支持”的记录不代表语义一致 |
+| 字幕：应用到全部 | 参考有底部应用按钮 | 本地已有，字体库扩展后需覆盖字体字段 |
+| 导出：本地/画布 | 参考菜单只有这两个目的地 | 本地已有；完整线上页面联动仍未发布验收 |
+| 独立音轨更深功能 | 当前参考素材无音轨，不能以灰色入口断言无其他功能 | **待有声样本核对**；本地独立音轨已有裁剪、移动、分割、增益和淡入淡出 |
+| 画布节点模式/参数/连接 | 参考原项目七素材、素材混剪、16:9/1080P/20s；详情见前批文档 | 已有此前核对记录，不重复宣称本轮重测所有模式 |
+
+## 证据与发布边界
+
+- 动态图片逐文件SHA-256、帧数及周期：素材目录README和`.qa/smart-edit/transition-assets.json`。
+- TypeScript及16项智能剪辑合同回归：`transition-parity-tsc.log`、`transition-parity-helpers.log`。
+- 完整helpers回归：491项通过、1项跳过，见`transition-parity-all-helpers.log`。
+- 本地浏览器效果截图：`.qa/smart-edit/transition-reference-assets-local.jpg`；已检查动态卡片、循环预览、整段播放及4秒转场。
+- 真实合成回归：`transition-parity-ffmpeg.log`；三种效果均验证2.2秒/4秒居中转场、保持4秒总长和中点像素。
+- 高清与连续短片段：`transition-stress-native.log`。使用独立本地组件与本地FFmpeg，未把模拟的导出任务回调当作真实Next API端到端成功。
+- 参考项目的2.2秒白场和0dB参数已恢复；临时新建字幕已清理；没有扣费生成或发布。
+- 字体、人声分离、字幕细节和正增益这些差异仍然存在，**不能标记“全部功能已齐/全站像素级一致”**。服务器更新继续冻结。

docs/smart-edit-ui-parity-2026-09-17.md

@@ -0,0 +1,56 @@
+# 智能剪辑界面对照与验收（2026-09-17）
+
+> 最新状态见[全面复核与像素差异](smart-edit-full-recheck-2026-09-17.md)：95个检查点，本轮修复9项，仍有25项明确差异及8项待验。下方各批验收结果是历史记录，不能解释为所有功能或全部像素已一致。
+
+参考：LibTV 已登录画布中的智能剪辑节点及编辑器。通过可见 DOM、截图和实际点击测量，不读取参考站隐藏状态；没有提交参考站生成任务。
+
+## 已逐项实现
+
+| 功能 | 对照与实现 | 验证 |
+| --- | --- | --- |
+| 画布入口 | 566×350 节点、标题外置、暗化素材预览、剪刀图标、打开按钮、简化时间尺和轨道 | 本地实际组件截图检查；旧节点恢复时按新尺寸归一化 |
+| 节点输入卡片 | 660 宽、16 圆角、48×48 素材缩略图、序号/时长/移除、提示词及发送 | 浏览器打开、参数修改；参考按钮连接画布视频，上限12份 |
+| 五种模式 | 默认/讲解/批量广告/口播/素材混剪 | 映射已有方案服务；默认模式使用通用选片，广告多版仍是候选顺序变体 |
+| 输出参数 | 六比例、480P/720P/1080P、0–300秒（0自动） | 浏览器选择21:9/480P/300s；六比例真实FFmpeg输出通过；远程21:9/1080P 2520×1080通过 |
+| 编辑器默认视图 | 48px顶栏、导出/关闭、默认大预览、底部35vh时间轴 | 同窗口截图比较；左侧配置改为按需打开 |
+| 视频参数 | 300px右栏、变速、时长、音量、暗色数值步进器 | 输入小数、选片及预览实际操作 |
+| 时间轴 | 刻度、视频缩略图、音轨条、播放头、缩放、吸附、裁剪和分割 | 鼠标排序/边缘裁剪通过；修复拖动结束遗漏历史记录；一次撤销只撤回一次操作 |
+| 字幕 | 列表/搜索/新建/时间范围/位置拖动、字体尺寸、间距、颜色、描边、透明度及样式预设 | 手动中文预览、实际ASS烧录；SRT保留文字，不把样式伪装成SRT能力 |
+| 转场 | 左侧转场库、右侧时长/删除；淡入淡出/黑场/白场 | 真实成片含硬切混用；像素采样确认黑场变暗、白场变亮 |
+| 导出 | “导出到本地/画布”菜单、提交阶段、关闭后重新打开恢复状态 | UI任务生命周期样例通过；真实远程接口回传与对象存储归档另行通过 |
+| 打开已有连接 | 自动建立顺序时间轴，不调用视觉模型/Python分析 | 七素材顺序方案测试通过；不会覆盖用户选择的智能剪辑参数 |
+| 任务标识 | 使用服务端实际接受的requestId；明确4xx拒绝不显示为持续导出 | 类型/回归检查，避免幂等返回旧任务后UI追踪新ID |
+
+## 验证结果和边界
+
+- 全量 helpers：488通过、1跳过（此批包含新增合同与远程宽屏结果测试）。
+- 真实FFmpeg：两种全局转场、三种逐片段转场、中文字幕样式、六种480P比例全部通过。
+- 真实300秒21:9/480P成片通过。AAC封装使实际容器时长300.022秒，输出探测允许0.2秒封装尾差，并仍要求成片与方案差异≤0.15秒；方案上限严格300秒。
+- 真实远程2520×1080合成与回传通过；普通合成没有调用付费模型或本地千问。
+- 浏览器验收使用独立本地入口加载真实React组件；任务回调样例与真正的Next API/服务器合成验证分开记录，不把样例回调当作生产端到端成功。
+- 当前记录不等于“全站所有分辨率100%像素一致”：参考站素材、浏览器缩放、窗口尺寸会影响结果；主项目集成页及窄屏还需发布候选验证。字幕内核使用现有Noto CJK字体，不能承诺与参考站未授权字体逐字形相同。
+- FunASR权重仍未安装，自动转写保持不可用；可导入SRT、手工编辑字幕。完整复杂对象多轨、音频等其他模块的状态不由这份记录扩大承诺。
+
+部署状态以 `video-render-service-2026-09-17.md` 最后更新为准。
+
+## 本地补齐与发布冻结
+
+用户要求“在本地测试确保功能都全了再更新到服务器”后，停止服务器更新。以下为此后仅在本地实现、验证的增量，不能视为已经上线。
+
+| 子功能 | 当前实现 | 本地验证 |
+| --- | --- | --- |
+| 预览画面变换 | 拖动画面、四角等比例缩放、重置、撤销；导出保留位置/缩放 | 浏览器实操；真实成片采样确认移动后的颜色区域和黑色留白 |
+| 字幕画面位置 | 在预览中拖动字幕；受控坐标生成ASS位置，用户文字继续转义 | 浏览器拖动；真实字幕输出像素区域检查；ASS注入回归通过 |
+| 音视频分离 | 有音轨视频右键分离；原视频静音，音频独立保存，不随后续视频排序移动 | 浏览器分离、拖动、分割、撤销；真实音频RMS确认音量未翻倍 |
+| 独立音轨编辑 | 起止裁剪、变速、音量、淡入淡出、静音、删除、最多24片段 | 合同边界和真实2倍速/延迟/增益/淡入淡出验证；超过视频尾部输出黑场 |
+| 转场库联动 | 三种基础转场，接点选择、时长和删除；与字幕/画面变换/独立音轨共用时间轴 | 三种转场再次通过FFmpeg像素检验；浏览器黑场转场操作通过 |
+
+本轮类型检查通过，智能剪辑合同测试15项通过；真实FFmpeg覆盖既有两种全局转场、三种逐接点转场、六种比例及新增画面/音频/字幕编辑。浏览器测试仍使用真实组件的独立本地入口，不把模拟任务回调当作线上端到端验证。转场库与完整特效库是不同范围：参考站目前展开可见的基础转场为三项，未确认的其他特效不能登记为已完成。
+
+证据保存在本地 `.qa/smart-edit/local-extension-helpers.log`、`local-extension-tsc.log`、`local-extension-ffmpeg.log` 和 `transition-library-local.png`。这些测试产物不随源码提交。
+
+## 动态转场与逐项再核对
+
+后续按用户要求直接采用参考站原始动态WebP（三张、每张75帧/3秒），并补齐当前相邻素材的右侧/主画面转场预览。新增0.2–4秒、0.1秒步进及不缩短总片长的居中转场；修正黑白场中点和独立音轨音量包络。三种转场2.2秒/4秒、1080P及12片段混用均已本地真实合成验证；整段播放可退出转场循环，完成后自动停止。
+
+逐层核对仍发现字体库、字幕数值/描边透明度、正值音量增益、人声分离等差异。详见 [子功能逐项清单](smart-edit-subfeature-audit-2026-09-17.md)，不要把此前概括性“已实现”解释为参考站所有子功能均已一致。该批继续保持本地，服务器未更新。

docs/video-canvas-base-audit-2026-09-18.md

@@ -0,0 +1,59 @@
+# 视频画布基础功能核对与 P40 资源调整（2026-09-18）
+
+## 本轮范围与结论
+
+直接核对 LibTV 现有画布的九宫格菜单、手写剧本工作流、资产准备和编组工具栏，并检查本地实现；没有触发参考站付费生成。应用改动仅在本地验收，未更新主站。本次服务器操作仅降低千问实例数量，未部署人声分离或视频超分模型。
+
+| 功能 | 参考站 | 本地核对结果 |
+| --- | --- | --- |
+| 九宫格菜单 | 多机位九宫格、剧情推演四宫格、角色脸部三视图、角色设定图、场景设定图、产品设定图、25宫格连贯分镜、电影级光影校正、角色三视图、3秒后/5秒前推演，共11项 | 均有配置与生成入口；参考图模型路由及宫格拆分已实现。本轮未重复实际付费生图 |
+| 宫格默认提示词 | 应随输入素材保持主体与场景 | 发现4项写死绿色服装或森林；已改成素材通用约束，保留2x2/3x3/5x5布局。仅迁移已知旧默认值，保留管理员自定义提示词 |
+| 自己编写分镜脚本 | 直接打开确认镜头，初始1个空镜头 | 原入口仅选中节点，未打开编辑器；已修复。编辑、新增、自动保存及刷新持久化在本地浏览器通过 |
+| 分镜编辑 | 确认镜头、准备资产、合成提示词；单元格弹层、行拖动 | 本地已有三阶段、逐格编辑、增删/上移、逐行/全部合成、导出；弹层样式和拖动排序未完全对齐 |
+| 资产准备 | 角色/场景/道具；AI生成、当前画布、本地上传、个人资产库 | 已有前3种来源，个人资产库入口及逐资产模型/比例参数仍缺；不能标记完全复刻 |
+| 批量生成条件 | 三阶段完成状态及视频生成门槛 | 本地已统计资产/提示词就绪数，但批量入口尚未完整落实三阶段门槛，空镜头也应补校验 |
+| 编组 | 组颜色、布局、整组执行、添加到工具箱、转分镜组、解组、批量下载 | 现有 TOOLBOX_TEMPLATE_GROUP 为预设背景框，没有完整成员关系和通用编组操作，尚未完整实现 |
+| 视频高清 | 视频超分功能 | 已有1080/2K/4K节点参数，但执行函数明确报未接通；没有实际超分服务 |
+| 人声分离 | 人声/伴奏处理 | 本地已实现HTDemucs实际分离，但当前Python脚本使用主应用所在机器CPU，尚未迁移算力服务器 |
+
+## 千问与 GPU 调整：已执行
+
+- 服务器4张Tesla P40，每张驱动报告23040MiB。
+- 调整前 GPU0、1、3 各运行一个 `qwen38-27b-32k` 实例，单实例 `--parallel 1`；GPU2 被 masked。
+- GPU2 的既有 Nginx 配置明确记录历史 `NVIDIA Xid 79 / fallen off bus`。虽然当前可枚举、显存5MiB，不能因此宣称可稳定计算。本轮没有解除该卡禁用。
+- 将 `127.0.0.1:18233` 从 Nginx upstream 移除，`nginx -t` 通过并 reload。
+- 连续3次确认18233的槽位空闲后，disable并stop `qwen38-27b-32k@3.service`。
+- GPU3显存从19055MiB降到5MiB，无计算进程，复查利用率0%。GPU0、1仍各19055MiB。
+- 并发由3路降为2路；模型、量化和单实例32K上下文保持不变。超过两路的请求等待或上游限流风险相应增加，未扩大请求队列。
+- 18230、18231直接健康检查、18080/18188代理健康检查通过；18084鉴权模型列表返回200。
+- 两个保留实例同时执行短文本推理，均200且返回2个输出token。
+- 手动运行CUDA驻留守护检查后，GPU3仍inactive。守护脚本会跳过disabled/masked实例。
+- 原视频服务所用18231保留；video-render与FRP服务均active。
+- 配置备份位于服务器 `/root/qwen-gpu-release-20260918-104112/qwen38-27b-local.conf`，备份目录仅root可访问。
+
+### 回滚
+
+先启用并启动 `qwen38-27b-32k@3.service`，等待18233健康且确认GPU3模型已驻留，再从备份恢复原Nginx配置，检查语法后reload。不能先恢复路由到尚未就绪的实例。部署媒体模型后回滚应先排空媒体任务并释放GPU3。
+
+## 算力服务方案与开源检索
+
+优先把新释放的GPU3作为媒体任务卡，人声分离和视频超分共享有界队列，初期每卡仅一个重任务。GPU2故障卡另行稳定性测试后才考虑加入。主应用仅处理鉴权、任务状态和结果引用；媒体下载、推理、封装及上传在算力服务执行，禁止失败后静默回落主站CPU。
+
+| 项目 | 依据与用途 | 本机边界 |
+| --- | --- | --- |
+| [Demucs](https://github.com/facebookresearch/demucs) | MIT，官方支持GPU；可沿用当前HTDemucs模型与人声/伴奏输出 | 将脚本硬编码CPU改为显式GPU设备，独立依赖环境、异步任务和取消/超时；目前未部署GPU实例 |
+| [Real-ESRGAN](https://github.com/xinntao/Real-ESRGAN) | BSD-3-Clause，官方包含视频推理、分块和FP32选项；优先作为可验证的超分基线 | P40先验证兼容环境和分块推理；逐帧方法可能闪烁，实际速度和质量未测，不能承诺实时4K |
+| [RealBasicVSR](https://github.com/ckkelvinchan/RealBasicVSR) | Apache-2.0，面向真实视频恢复，具有时序处理 | 旧MMCV/MMEditing依赖需隔离。原示例整段读帧、默认25fps，不可原样上线；应分段并保持源帧率及音轨 |
+| [FlashVSR](https://github.com/OpenImagingLab/FlashVSR) | Apache-2.0，官方性能展示基于A100等，依赖稀疏注意力实现 | 官方性能不能套用P40，不作为首选部署目标 |
+| [SeedVR2](https://github.com/ByteDance-Seed/SeedVR) | Apache-2.0，官方高规格GPU视频恢复方案 | 官方H100示例不能用于推断P40吞吐，暂不作为首选 |
+
+Real-ESRGAN的视频示例存在shell路径调用、FPS表达式求值和帧失败继续执行等行为，不能直接封装为公网API。实现时使用安全argv调用FFmpeg、验证媒体参数、限制时长/像素/磁盘量；输入帧失败应明确失败，成片校验帧率、帧数、时长及音轨，不把普通缩放标成AI超分。
+
+## 本轮验证
+
+- 本地浏览器3027：手写入口打开含1个空镜头的工作流；填写柜门画面描述，新增第2镜头；保存后刷新，2个镜头和描述均保留。
+- 四个相关测试文件：66通过，1项FFmpeg集成测试因当前命令PATH无FFmpeg跳过。本轮没有修改媒体编解码实现。
+- TypeScript 全量 `tsc --noEmit --incremental false` 通过。
+- 新增回归覆盖：旧示例提示词迁移、布局约束保留，以及管理员自定义绿色服装/森林提示词不被覆盖。
+- 参考站临时测试节点已删除，原11个节点保留，恢复到素材混剪分组视图。
+- 本轮不宣称编组、超分服务、像素级分镜编辑器或GPU人声分离已完成。

docs/video-canvas-connections-2026-09-18.md

@@ -0,0 +1,54 @@
+# 画布连线语义与请求验收（2026-09-18）
+
+## 核对范围
+
+参考 LibTV 同一项目的素材混剪：7条视频输入边进入智能剪辑节点，进入真实剪辑器可见14秒时间轴。只查看原有节点、连线及编辑器，未调用参考站生成、修改素材或导出。
+
+本地读取前端节点、连线、任务提交、轮询恢复及后端请求契约，并使用真实Next页面、隔离账号和隔离画布进行浏览器验收。拦截付费生成请求，检查浏览器实际序列化的请求体；不以“画出一条线”作为功能完成依据。
+
+## 连线应该表示什么
+
+- `source → target`：target在执行时读取source当前已完成的输出。节点移动不改变依赖，输入顺序由保存的连线顺序决定。
+- 图片/视频/音频输出与输入预览是两回事。`sourceUrl`、poster不能证明上游已生成成功；失败、排队和未完成的上游应阻止下游提交。
+- 普通多图参考可以按URL去重；首尾帧是两个位置，即使两张图URL相同也必须保留两个槽位。
+- 文字输出作为下游提示词或配音内容；视频素材作为剪辑输入；角色图片和角色/特效描述应进入相应视频请求。
+- 当前连线是下一次执行的输入关系，已提交任务中的参数是不可变的历史快照。改线不会自动付费重生成，也不篡改已经生成的成片。
+- 完整“整组执行”还需要独立调度器：拓扑排序、等待上游成功、失败阻断、重试和幂等提交。当前没有这套完整实现，不能把连线动画或逐个点生成当作整组调度。
+
+## 本轮实际修复
+
+| 问题 | 修复 |
+| --- | --- |
+| 视频/图片重跑会混入params.referenceImages中的旧URL | 标准重跑路径从当前连线重新解析输入，断线后不使用旧URL或节点自身预览补位 |
+| 首尾帧相同URL被Set去重 | 保留连线顺序及两个槽位，后端引用校验也保留相同URL |
+| 文字连线没有参与视频/图片/配音请求 | 接入文字输出；再次执行替换此前连接的文本区段。音频保存原始表单，提交时组合当前文本，避免重复朗读；编辑表单移除解析器标记，任务参数仍保留已解析文本 |
+| 上游失败但参考预览仍可被使用 | 提交前检查上游状态与实际输出，缺失节点、空文本、未完成媒体阻止下游 |
+| 可以自连、循环连接或把分组背景当输入 | 手动连线、从端口创建节点和选参考图统一检查；旧画布循环在执行前也拦截 |
+| 不匹配素材可以挂一条无作用的线 | 常用图片、视频、文字、音频、剪辑节点检查输入类型；多余图片超出视频模式容量明确提示，不静默截取 |
+| 任务响应/轮询晚到重新补回删除的线 | 记录本页断开的source/target关系，恢复任务边时不覆盖这次断线；明确重连会解除该限制 |
+| 角色素材只有描述没有参考图 | 视频请求同时包含角色图片和角色描述 |
+| 宫格裁切失败会退回整张图或少数切片 | 所选切片必须全部准备成功，失败则停止提交并提示重试 |
+| 空闲连线一直流动 | 仅下游任务处于pending/running时显示流动，空闲关系保留静态连线 |
+
+有成品的图片继续派生新图时，该成品本身作为新输入，不自动把它所有历史图片依赖再追加一遍。尚无成品的图像节点则读取其当前直接输入。显式选取的额外参考仍会建立对应输入边。
+
+## 验收证据
+
+浏览器脚本位于本地忽略目录 `.qa/connection-audit-0918/browser.cjs`，请求证据位于同目录`request-evidence.json`。使用隔离测试账号，不提交供应商，不导出账号凭据。
+
+- 真实视频页面提交：相同URL的首尾帧保留2项，连接文字进入prompt，历史stale.png没有进入请求。
+- 改接素材且更新源URL后，下一次请求只包含新地址，断开的文字不再进入prompt。
+- 删除所有图片连线后，首帧模式提示缺少参考图，无新增请求。
+- 上游failed时提示先完成上游，无新增请求。
+- 暂缓任务POST响应，在页面删除连线后再释放响应，连线不会重新出现。
+- 文字连入音频节点，真实生成音频提交中的text包含该文字；通过右键再次执行，文字没有重复追加。
+- TypeScript全量检查 `tsc --noEmit --incremental false` 通过。
+- 回归测试：62通过、1跳过（命令PATH未提供FFmpeg；本轮未修改编解码实现）。覆盖循环、端点缺失、类型、模式容量、宫格部分失败、文本更新及首尾帧位置语义。
+
+## 尚未完成与边界
+
+- 整组依赖执行器、完整编组成员关系尚未实现。
+- 剪辑器内部独立上传的素材、编辑好的时间轴与画布输入边是不同层数据，现有成片/时间轴不会被改线自动改写；编辑器内素材增删和画布编组之间还需要完整双向同步验收。
+- 部分专用节点有自己的提交路径，本轮不宣称所有导演台、标记对象和第三方工具组合已全链路验收。
+- 模型结果质量、耗时、真实供应商生成没有在本轮重测。本轮验证输入关系和请求边界。
+- 代码仅本地修复，未推送、未更新主站服务器。

docs/video-canvas-groups-upscale-2026-09-18.md

@@ -0,0 +1,39 @@
+# 编组和独立视频高清验收 — 2026-09-18
+
+工作分支 `codex/video-canvas-parity-0916`。本次没有更新主站前端/主项目版本。
+
+## 编组
+
+- Shift 点击/框选、Ctrl/⌘+G、新建组、嵌套成员、整体移动、四角调整边框、颜色/名称、网格/横排/竖排、移动到组/移出、解组保留节点和连线。旧预设背景框按几何包含关系迁移一次，后续按明确 parentGroupId 工作。
+- 全部就绪图片/视频可转分镜组，预览、前后镜头、调整顺序；混合工具或未就绪素材时禁用转换。顺序保存在画布中，导出重建 ID。
+- 我的工具箱保存到所属画布，按用户权限跨画布查询；模板 JSON 导入/导出。内部节点和连线重建 ID，运行编号清除；外部媒体依赖转换为必填输入节点，缺素材时整批拒绝。角色/特效外部依赖要求先纳入组内。
+- 整组执行一次性原子入库，客户端持久 outbox 使用同一 runId/requestId 重试；关闭/刷新浏览器后服务端继续执行。按真实连线获取本次上游结果，拒绝循环、错误类型、缺输入、跨批次依赖。上游失败/取消/无可恢复编号的中断会阻断下游，不取旧输出。
+- 支持现有图片/视频/文字分析/音频/图片处理/视频高清任务。智能剪辑遇到本次新视频时拆成隐藏选片阶段 + 成片阶段，依赖新生成素材；保留已确认且输入未变的现有剪辑方案。默认新组合使用顺序剪辑，已有智能参数按保存配置。取消只取消尚未提交的任务。
+- 选中宫格切片需先落成独立图片节点再整组执行，有明确拦截，未用整张原图冒充切片。
+
+## 视频高清
+
+Real-ESRGAN 独立部署到算力服务器物理 GPU 3（P40）。保留千问 GPU 0/1，GPU 2 不使用。服务与 48→算力私有 API 隧道均启用；48 的 `/etc/video-upscale-client.env` 已准备，尚未注入/发布主项目版本。部署和约束见 `services/video-upscale/README.md`。
+
+真实样例均由色彩图案/正弦音合成，无付费模型：
+
+| 输入 | 输出 | 验证 |
+| --- | --- | --- |
+| 960×540、2fps、1秒、音轨 | 3840×2160 | 2帧，2fps，音轨保留，1.002秒 |
+| 270×480、2fps、1秒、无音轨 | 1080×1920 | 2帧，2fps，1秒 |
+| 960×540 + 90度旋转信息、含音轨 | 1080×1920 | 归一化方向，音轨/帧率/帧数保留 |
+| 主项目编组提交 960×540 | 2560×1440 | 真实 GPU → 持久任务成功 → 本地对象存储归档 → 页面恢复 src |
+
+第一轮横屏/竖屏短片含传输与轮询分别约 8.4/6.9 秒。仅是 1 秒 2 帧链路测试，不能推算常见 24/30fps 长视频吞吐。
+
+重启测试发现并修复两处问题：128核机器 FFmpeg 默认滤镜线程超过服务限额；SIGTERM 杀子进程时曾将正在处理的任务记成 failed。修复后同一任务重启续跑，attempts=2，返回真实 4K 结果。未鉴权 401、重复请求复用记录、参数冲突 409 均通过。
+
+## 验证证据与边界
+
+- 75 项定向回归通过，包含连线、编组、任务恢复、智能剪辑；新增组内自动选片/成片依赖测试。
+- 浏览器实测：多选编组、布局、整体拖动、调整边框、模板保存导入/必填输入拦截、分镜转换/预览/排序、真实提交参数依赖、刷新恢复排队、解组保留成员。
+- 队列读取仅查询当前待处理任务及直接依赖，避免每次 claim 读取整个项目历史。
+- 真实 PostgreSQL：原子批量提交、幂等、冲突拒绝、项目所有权；上游失败时下游 attempts=0。
+- 真实主项目 API：2K 成片归档成功，FFprobe 验证 H.264/AAC、2560×1440、帧数/时长；刷新后保留结果链接。
+- 当前验收 Chrome 对 `127.0.0.1:19440` 本地对象存储的媒体加载返回 `Media load rejected by URL safety check`。没有绕过浏览器策略；网页实际播放不能记为通过。界面增加明确“结果已保存、预览加载失败”的提示与打开链接，避免黑屏被理解为仍在处理。尚需在获准访问的正式/测试对象存储域名上完成播放验收后发布主项目。
+- 本地证据保存在忽略目录 `.qa/group-upscale-0918/`，可复用 GPU 验证脚本 `scripts/qa/video-upscale-service.cjs`。正式运行代码/文档不包含密钥或 QA 账号。

docs/video-render-service-2026-09-17.md

@@ -26,3 +26,15 @@
 FunASR适配与安装脚本已实现，未装权重、未做真实ASR验收，界面禁用自动转写并支持导入SRT。当前本地千问接口不能替代视觉选片，视觉分析继续使用主项目已有模型配置。
 
 部署边界：独立合成服务已启动且两机链路实测通过；主项目正式应用的切换状态将在发布验证后追加。此前音频/导演台等开发分支改动不能凭本记录视为已经上线。
+
+## 2026-09-17 智能剪辑界面与输出范围补齐
+
+用户再次要求按参考站像素级对齐。新增独立编辑器和画布节点输入卡片，逐项记录见 `smart-edit-ui-parity-2026-09-17.md`。新增手动字幕样式、逐片段三种转场、六种比例、480P和300秒上限，主项目与算力服务共用合同；修正远程返回结果旧90秒/1920限制。真实300秒输出发现AAC封装尾差，输出探测及校验保留0.2/0.15秒容差。
+
+48候选应用仍运行独立目录 `/var/www/ai-chatbot-render-0917` 和回环3135，公网应用3125尚未切换。候选旧版已完成真实Next API鉴权、排队、选片、节拍、远程渲染和归档测试；新UI必须完成候选重建与验证后再更新此状态。不要将算力服务已部署理解成主站UI已上线。
+
+## 最新状态：本地验收优先，停止服务器更新
+
+上述候选随后已在 `6869bd8c` 完成构建及启动；七个素材、手工字幕、21:9/480P、重复提交幂等、真实Next API到远程渲染和对象存储回传通过。此时公网仍为3125旧应用，未切换。算力服务当时的包为 `0f7d1c0eeb42971c`。
+
+用户进一步要求先在本地补齐和测试特效等缺项，再更新服务器。因此此后的画面变换、字幕坐标和独立音轨等增量只在本地开发验证，尚未同步至候选或算力服务；需要将主项目与算力服务的共享合同同步发布后才能在线使用。完成已确认功能不代表全部参考站子功能已核对完毕，继续对照的记录见智能剪辑界面验收文档。

hooks/use-canvas-group-runs.ts

@@ -0,0 +1,90 @@
+"use client";
+import { useCallback, useEffect, useRef, useState } from "react";
+import {
+  canvasTaskRequestSchema,
+  type CanvasTaskRequest,
+} from "@/lib/video-canvas/task-contract";
+import { toast } from "sonner";
+export function useCanvasGroupRuns(projectId: string, refresh: () => void) {
+  const [submitting, setSubmitting] = useState(false),
+    active = useRef(projectId),
+    refreshRef = useRef(refresh),
+    sending = useRef(false);
+  active.current = projectId;
+  refreshRef.current = refresh;
+  const key = `video-canvas-group-outbox:${projectId}`;
+  const read = useCallback(() => {
+    try {
+      const parsed = canvasTaskRequestSchema
+        .array()
+        .min(1)
+        .max(40)
+        .safeParse(JSON.parse(localStorage.getItem(key) || "null"));
+      return parsed.success &&
+        parsed.data.every((r) => r.projectId === projectId && r.workflow)
+        ? parsed.data
+        : null;
+    } catch {
+      return null;
+    }
+  }, [key, projectId]);
+  const send = useCallback(
+    async (requests: CanvasTaskRequest[]) => {
+      if (sending.current) return;
+      sending.current = true;
+      setSubmitting(true);
+      try {
+        const response = await fetch("/api/canvas/group-runs", {
+          method: "POST",
+          headers: { "Content-Type": "application/json" },
+          body: JSON.stringify(requests),
+          signal: AbortSignal.timeout(20000),
+        });
+        if (!response.ok) {
+          if (response.status < 500 && response.status !== 429) {
+            localStorage.removeItem(key);
+            const body = await response.json();
+            throw Error(body.error || "编组请求被拒绝");
+          }
+          throw Error("编组提交尚未确认，已保存，将查询同一批任务");
+        }
+        localStorage.removeItem(key);
+        if (active.current === projectId) refreshRef.current();
+      } finally {
+        sending.current = false;
+        if (active.current === projectId) setSubmitting(false);
+      }
+    },
+    [key, projectId]
+  );
+  const submit = useCallback(
+    async (requests: CanvasTaskRequest[]) => {
+      const pending = read();
+      if (pending && pending[0].workflow?.runId !== requests[0].workflow?.runId)
+        throw Error("上一批编组提交仍在确认，请稍后重试");
+      localStorage.setItem(key, JSON.stringify(requests));
+      await send(requests);
+    },
+    [key, read, send]
+  );
+  useEffect(() => {
+    setSubmitting(false);
+    const retry = () => {
+      const requests = read();
+      if (requests && !sending.current)
+        void send(requests).catch((error) => {
+          if (!read()) toast.error(error.message);
+        });
+    };
+    retry();
+    const timer = setInterval(retry, 10000);
+    window.addEventListener("online", retry);
+    window.addEventListener("focus", retry);
+    return () => {
+      clearInterval(timer);
+      window.removeEventListener("online", retry);
+      window.removeEventListener("focus", retry);
+    };
+  }, [read, send]);
+  return { submit, submitting };
+}

hooks/use-video-canvas-tasks.ts

@@ -1,4 +1,5 @@
 "use client";
+import { withConnectedText } from "@/lib/video-canvas/connections";
 import { useCallback, useEffect, useRef, useState } from "react";
 import { toast } from "sonner";
 import {
@@ -7,6 +8,8 @@ import {
   type CanvasTaskRequest,
 } from "@/lib/video-canvas/task-contract";
 
+class CanvasTaskRejectedError extends Error {}
+
 type Node = CanvasTaskRequest["node"];
 type Edge = CanvasTaskRequest["edges"][number];
 const storageKey = (projectId: string) => `video-canvas-outbox:${projectId}`;
@@ -38,10 +41,18 @@ export function useVideoCanvasTasks<T extends Node>(
   loaded: { id: string; revision: string } | null,
   setNodes: React.Dispatch<React.SetStateAction<T[]>>,
   setEdges: React.Dispatch<React.SetStateAction<Edge[]>>,
+  canvasEdges: Edge[] = [],
 ) {
   const [tasks, setTasks] = useState<CanvasGenerationTask[]>([]);
   const hiddenNodes = useRef(new Set<string>());
   const recoveredEdges = useRef(new Set<string>());
+  const edgeEdits = useRef({ projectId, current: new Set<string>(), removed: new Set<string>() });
+  if (edgeEdits.current.projectId !== projectId)
+    edgeEdits.current = { projectId, current: new Set(), removed: new Set() };
+  const currentEdgeKeys = new Set(canvasEdges.map(edge => `${edge.source}>${edge.target}`));
+  for (const key of edgeEdits.current.current) if (!currentEdgeKeys.has(key)) edgeEdits.current.removed.add(key);
+  for (const key of currentEdgeKeys) edgeEdits.current.removed.delete(key);
+  edgeEdits.current.current = currentEdgeKeys;
   const refreshRef = useRef<() => void>(() => {});
   const activeProject = useRef(projectId);
   activeProject.current = projectId;
@@ -78,7 +89,9 @@ export function useVideoCanvasTasks<T extends Node>(
           )
           .filter((edge) => {
             const key = `${edge.source}>${edge.target}`;
-            if (existing.has(key)) return false;
+            // A request is a historical snapshot, not permission to undo a
+            // later disconnect while its response/poll is in flight.
+            if (existing.has(key) || edgeEdits.current.removed.has(key)) return false;
             existing.add(key);
             return true;
           });
@@ -115,6 +128,8 @@ export function useVideoCanvasTasks<T extends Node>(
               ),
             );
         }
+        if (response.status >= 400 && response.status < 500)
+          throw new CanvasTaskRejectedError(body.error || "任务参数未通过校验");
         throw Error(body.error || "提交暂时失败，恢复连接后自动核对");
       }
       forget(projectId, request.requestId);
@@ -149,10 +164,19 @@ export function useVideoCanvasTasks<T extends Node>(
     ) => {
       if (!loaded || loaded.id !== projectId) throw Error("画布尚未加载完成，请稍后提交");
       const requestId = crypto.randomUUID();
+      // The task params retain resolved text; the editable form retains only
+      // the user's prompt, so linked content is neither duplicated nor exposed
+      // as resolver markup when the node is reopened.
+      const editableDraft = { ...node.draftValues };
+      if (typeof editableDraft.prompt === "string") editableDraft.prompt = withConnectedText(editableDraft.prompt, []);
+      if (editableDraft.params && typeof editableDraft.params === "object" && !Array.isArray(editableDraft.params)) {
+        const params = editableDraft.params as Record<string, unknown>;
+        if (typeof params.prompt === "string") editableDraft.params = { ...params, prompt: withConnectedText(params.prompt, []) };
+      }
       const snapshot = {
         ...node,
         draftValues: {
-          ...node.draftValues,
+          ...editableDraft,
           taskRequestId: requestId,
           generationTaskId: undefined,
           appliedTaskStatus: undefined,
@@ -175,13 +199,17 @@ export function useVideoCanvasTasks<T extends Node>(
       remember(request);
       setNodes((nodes) => nodes.map((item) => (item.id === node.id ? snapshot : item)));
       try {
-        await send(request);
+        const accepted = await send(request);
+        refreshRef.current();
+        return accepted.requestId;
       } catch (error) {
+        if (kind === "smart-edit" && error instanceof CanvasTaskRejectedError) throw error;
         toast.error(
           error instanceof Error ? error.message : "连接中断，恢复后将核对任务，请勿重复提交",
         );
       }
       refreshRef.current();
+      return requestId;
     },
     [loaded, projectId, send, setNodes],
   );

lib/video-canvas/config.ts

@@ -125,7 +125,7 @@ const defaultFunctions: VideoCanvasFunctionDefinition[] = [
     route: "canvas-generate-image",
     enabled: true,
     prompt:
-      "以选中图片作为角色身份、服装、发型、场景光线、镜头距离和画面质感的唯一参考，生成一张 2x2 剧情推演四宫格合成图。画面中只能有 4 个画面区域，只允许一条竖向分割线和一条横向分割线；严禁出现 3 行 3 列、9 个画面区域或九宫格布局。四格都必须保持参考图的近景/半身近中景人物构图，主体占画面主要面积，不要远景、全景、森林空镜或主体很小的画面。四格是同一角色在同一自然光场景中的连续轻微剧情变化：1 正面安静凝视，2 保留参考图抬手/袖摆动作的延续，3 视线或表情轻微变化，4 情绪收束的近中景反应。保持五官、发型、绿色服装、柔和森林逆光、色调和景深一致；每格角度和动作有轻微差异，但不能像 9 个不同机位分镜。输出单张 16:9 图片，2x2 等分排版，分割线清晰干净，不能出现文字、水印、编号、边框标签或 3x3/5x5 宫格。",
+      "以选中图片作为主体身份、外观、场景光线、镜头距离和画面质感的唯一参考，生成一张 2x2 剧情推演四宫格合成图。只能有 4 个等大画面区域，一条横向和一条竖向分割线，严禁生成 3x3 或 5x5 布局。四格依次呈现初始状态、动作或事件延续、细微变化、结果或反应，形成连贯的短叙事。根据参考图实际主体安排变化，保持原有身份、数量、服装或产品结构、配色、场景、光线和景别一致，不引入无依据的人物、服装颜色或环境。输出单张 16:9 图片，2x2 等分排版，分割线清晰干净，不添加文字、水印、编号或边框标签。",
     description: "官方剧情推演四宫格：基于参考图生成 2x2 连续剧情分镜合成图。",
   },
   {
@@ -135,7 +135,7 @@ const defaultFunctions: VideoCanvasFunctionDefinition[] = [
     route: "canvas-generate-image",
     enabled: true,
     prompt:
-      "以选中图片作为唯一角色参考，生成一张角色三视图设定图。只输出正面、侧面、背面 3 个完整站姿人物，三人横向并排在同一张 16:9 画面中，白色或浅灰干净背景。必须保持同一角色身份、五官、发型、绿色服装、材质、体型比例和配饰一致；正面、侧面、背面视角要清晰可辨，人物全身完整入画，站姿自然。禁止生成 3x3 九宫格、多机位分镜、黑色分割线、重复头像、文字、水印、编号、环境场景或额外人物。",
+      "以选中图片作为唯一角色参考，生成一张角色三视图设定图。只输出同一角色的正面、侧面、背面 3 个完整站姿视图，横向并排在同一张 16:9 画面中，白色或浅灰干净背景。保持参考图中的身份、五官、发型、服装颜色、材质、体型比例和配饰一致，不擅自添加或改变服装颜色。三个视角清晰可辨，全身完整入画，站姿自然。禁止生成 3x3 九宫格、多机位分镜、黑色分割线、重复头像、文字、水印、编号、环境场景或额外人物。",
     description: "官方角色三视图预设，使用参考图生图/编辑模型。",
   },
   {
@@ -145,7 +145,7 @@ const defaultFunctions: VideoCanvasFunctionDefinition[] = [
     route: "canvas-generate-image",
     enabled: true,
     prompt:
-      "以选中图片作为角色、服装、发型、场景光线的唯一参考，生成一张 3x3 多机位分镜合成图。必须是 9 个明显不同的镜头画面，不是同一张图裁切或重复人脸。九格分别包含：1 森林环境远景/主体很小，2 全身远景，3 中景正面，4 侧面半身，5 低机位或俯拍，6 面部近景，7 手部/衣袖/动作细节，8 背面或侧背，9 仰视或情绪特写。保持同一角色身份、绿色服装、自然森林光线一致，每格构图、景别、机位和主体大小必须不同。输出为单张 16:9 图片，九格之间使用清晰黑色分割线，不能出现文字、水印、编号，不能白底棚拍。",
+      "以选中图片作为主体、外观、场景和光线的唯一参考，生成一张 3x3 多机位分镜合成图。必须是 9 个明显不同的镜头，不是同一张图裁切或重复。九格分别呈现：1 环境建立远景，2 主体完整全景，3 正面中景，4 侧面中景，5 俯拍，6 主体近景，7 关键局部细节，8 背面或侧背视角，9 低机位或特写。根据实际主体选择合理构图，保持身份、数量、服装或产品结构、配色、材质、场景和光线一致，不凭空加入森林、指定服装颜色或人物。输出单张 16:9 图片，严格 3x3 等分排版，九格之间使用清晰黑色分割线，不添加文字、水印或编号。",
     description: "官方九宫格工具：先生成带黑色分割线的 3x3 分镜图，再由宫格切分拆成单张镜头。",
   },
   {
@@ -155,7 +155,7 @@ const defaultFunctions: VideoCanvasFunctionDefinition[] = [
     route: "canvas-generate-image",
     enabled: true,
     prompt:
-      "以选中图片作为角色、服装、发型、场景光线和电影质感的唯一参考，生成一张 5x5 连贯分镜合成图。必须是 25 个连续镜头，体现从远景建立场景、人物动作推进、情绪变化、细节特写到结尾反应的完整微型叙事，不是同一张图重复裁切。保持同一角色身份和绿色服装一致，森林自然光线统一；每格景别、机位、主体大小或动作都要有明显差异。输出为单张 16:9 图片，必须严格 5x5 等分排版，25 个宫格的宽高一致、间距一致、边界完整，所有宫格之间使用清晰纯黑分割线，分割线等宽且横线竖线都要贯穿整张图；禁止使用白线、浅色线、柔边、渐变线、不规则拼贴、错位排版、跨格大图、留白边距、文字、水印、编号或额外边框说明。",
+      "以选中图片作为主体、外观、场景光线和画面质感的唯一参考，生成一张 5x5 连贯分镜合成图。必须是 25 个连续镜头，从建立场景、动作或事件推进、状态变化、细节特写到结尾反应形成完整微型叙事，不是同一张图重复裁切。保持参考图中的主体身份、数量、服装或产品结构、配色、材质和场景光线一致，不引入无依据的人物、指定服装颜色或环境。每格景别、机位、主体大小或动作要有合理差异。输出单张 16:9 图片，严格 5x5 等分排版，25 格宽高和间距一致、边界完整。使用清晰纯黑等宽分割线，横线竖线贯穿整张图。禁止白线、浅色线、柔边、渐变线、不规则拼贴、错位排版、跨格大图、留白边距、文字、水印、编号或额外边框说明。",
     description: "官方 25 宫格连贯分镜：生成 5x5 连续镜头合成图。",
   },
   {
@@ -696,14 +696,20 @@ function isRecord(value: unknown): value is Record<string, unknown> {
 }
 
 const legacyFunctionPrompts: Record<string, string[]> = {
+  "multi-view-nine-grid": [
+    "以选中图片作为角色、服装、发型、场景光线的唯一参考，生成一张 3x3 多机位分镜合成图。必须是 9 个明显不同的镜头画面，不是同一张图裁切或重复人脸。九格分别包含：1 森林环境远景/主体很小，2 全身远景，3 中景正面，4 侧面半身，5 低机位或俯拍，6 面部近景，7 手部/衣袖/动作细节，8 背面或侧背，9 仰视或情绪特写。保持同一角色身份、绿色服装、自然森林光线一致，每格构图、景别、机位和主体大小必须不同。输出为单张 16:9 图片，九格之间使用清晰黑色分割线，不能出现文字、水印、编号，不能白底棚拍。",
+  ],
   "story-beat-grid-4": [
+    "以选中图片作为角色身份、服装、发型、场景光线、镜头距离和画面质感的唯一参考，生成一张 2x2 剧情推演四宫格合成图。画面中只能有 4 个画面区域，只允许一条竖向分割线和一条横向分割线；严禁出现 3 行 3 列、9 个画面区域或九宫格布局。四格都必须保持参考图的近景/半身近中景人物构图，主体占画面主要面积，不要远景、全景、森林空镜或主体很小的画面。四格是同一角色在同一自然光场景中的连续轻微剧情变化：1 正面安静凝视，2 保留参考图抬手/袖摆动作的延续，3 视线或表情轻微变化，4 情绪收束的近中景反应。保持五官、发型、绿色服装、柔和森林逆光、色调和景深一致；每格角度和动作有轻微差异，但不能像 9 个不同机位分镜。输出单张 16:9 图片，2x2 等分排版，分割线清晰干净，不能出现文字、水印、编号、边框标签或 3x3/5x5 宫格。",
     "以选中图片作为角色身份、服装、发型、场景光线、镜头距离和画面质感的唯一参考，生成一张 2x2 剧情推演四宫格合成图。四格都必须保持参考图的近景/半身近中景人物构图，主体占画面主要面积，不要远景、全景、森林空镜或主体很小的画面。四格是同一角色在同一自然光场景中的连续轻微剧情变化：1 正面安静凝视，2 保留参考图抬手/袖摆动作的延续，3 视线或表情轻微变化，4 情绪收束的近中景反应。保持五官、发型、绿色服装、柔和森林逆光、色调和景深一致；每格角度和动作有轻微差异，但不能像 9 个不同机位分镜。输出单张 16:9 图片，2x2 等分排版，分割线清晰干净，不能出现文字、水印、编号、边框标签或 3x3/5x5 宫格。",
     "以选中图片作为角色、服装、发型、场景光线的唯一参考，生成一张 3x3 多机位分镜合成图。必须是 9 个明显不同的镜头画面，不是同一张图裁切或重复人脸。九格分别包含：1 森林环境远景/主体很小，2 全身远景，3 中景正面，4 侧面半身，5 低机位或俯拍，6 面部近景，7 手部/衣袖/动作细节，8 背面或侧背，9 仰视或情绪特写。保持同一角色身份、绿色服装、自然森林光线一致，每格构图、景别、机位和主体大小必须不同。输出为单张 16:9 图片，九格之间使用清晰黑色分割线，不能出现文字、水印、编号，不能白底棚拍。",
   ],
   "continuous-storyboard-grid-25": [
+    "以选中图片作为角色、服装、发型、场景光线和电影质感的唯一参考，生成一张 5x5 连贯分镜合成图。必须是 25 个连续镜头，体现从远景建立场景、人物动作推进、情绪变化、细节特写到结尾反应的完整微型叙事，不是同一张图重复裁切。保持同一角色身份和绿色服装一致，森林自然光线统一；每格景别、机位、主体大小或动作都要有明显差异。输出为单张 16:9 图片，必须严格 5x5 等分排版，25 个宫格的宽高一致、间距一致、边界完整，所有宫格之间使用清晰纯黑分割线，分割线等宽且横线竖线都要贯穿整张图；禁止使用白线、浅色线、柔边、渐变线、不规则拼贴、错位排版、跨格大图、留白边距、文字、水印、编号或额外边框说明。",
     "以选中图片作为角色、服装、发型、场景光线和电影质感的唯一参考，生成一张 5x5 连贯分镜合成图。必须是 25 个连续镜头，体现从远景建立场景、人物动作推进、情绪变化、细节特写到结尾反应的完整微型叙事，不是同一张图重复裁切。保持同一角色身份和绿色服装一致，森林自然光线统一；每格景别、机位、主体大小或动作都要有明显差异。输出为单张 16:9 图片，5x5 等分排版，使用清晰分割线，不能出现文字、水印、编号或多余边框说明。",
   ],
   "character-three-view": [
+    "以选中图片作为唯一角色参考，生成一张角色三视图设定图。只输出正面、侧面、背面 3 个完整站姿人物，三人横向并排在同一张 16:9 画面中，白色或浅灰干净背景。必须保持同一角色身份、五官、发型、绿色服装、材质、体型比例和配饰一致；正面、侧面、背面视角要清晰可辨，人物全身完整入画，站姿自然。禁止生成 3x3 九宫格、多机位分镜、黑色分割线、重复头像、文字、水印、编号、环境场景或额外人物。",
     "以输入角色图为唯一角色参考，生成正面、侧面、背面三视图角色设定图。保持五官、发型、服装、颜色和体型一致，干净白底，三视图水平排布，适合后续视频角色一致性使用。",
   ],
 };

lib/video-canvas/connections.ts

@@ -0,0 +1,116 @@
+/** Directed canvas edges are live inputs, not a cache of earlier request URLs. */
+export type CanvasConnection = { id: string; source: string; target: string };
+export type ConnectionNode = {
+  id: string;
+  title?: string;
+  kind?: string;
+  inputType?: string;
+  draftValues?: Record<string, any>;
+};
+
+export function connectionError(nodes: ConnectionNode[], edges: CanvasConnection[], source: string, target: string) {
+  if (source === target) return "不能连接到同一个节点";
+  const from = nodes.find(node => node.id === source), to = nodes.find(node => node.id === target);
+  if (!from || !to) return "连接的节点已不存在";
+  if ([from, to].some(node => node.kind === "group" || node.draftValues?.action === "TOOLBOX_TEMPLATE_GROUP"))
+    return "请连接分组内的具体节点，分组框不是素材输入";
+  if (edges.some(edge => edge.source === source && edge.target === target)) return "这两个节点已经连接";
+  const pending = [target], seen = new Set<string>();
+  while (pending.length) {
+    const id = pending.pop()!;
+    if (id === source) return "这条连线会形成循环依赖，请调整连接方向";
+    if (seen.has(id)) continue;
+    seen.add(id);
+    pending.push(...edges.filter(edge => edge.source === id).map(edge => edge.target));
+  }
+  return null;
+}
+
+export function connectedInputNodes<T extends ConnectionNode>(nodes: T[], edges: CanvasConnection[], target: string): T[] {
+  const seen = new Set<string>();
+  return edges.filter(edge => edge.target === target).flatMap(edge => {
+    if (seen.has(edge.source)) return [];
+    seen.add(edge.source);
+    const node = nodes.find(item => item.id === edge.source);
+    if (!node) throw Error("上游节点已删除，请移除失效连线后重试");
+    // Also reject cycles restored from an older saved canvas.
+    const remaining = edges.filter(item => item.id !== edge.id);
+    const error = connectionError(nodes, remaining, edge.source, target);
+    if (error && !error.includes("已经连接")) throw Error(error);
+    return [node];
+  });
+}
+
+export function connectedText(node: ConnectionNode) {
+  if (node.inputType !== "text" && node.kind !== "script" && node.kind !== "text") return "";
+  const content = node.draftValues?.content;
+  if (typeof content === "string") return content.trim();
+  if (Array.isArray(content)) return content.filter(value => typeof value === "string").join("\n").trim();
+  const rows = node.draftValues?.rows;
+  return Array.isArray(rows) ? rows.map(row => [row.shot, row.sceneDescription, row.dialogue, row.imagePrompt, row.motionPrompt].filter(Boolean).join(" · ")).join("\n").trim() : "";
+}
+
+export function assertConnectedInputsReady(nodes: ConnectionNode[]) {
+  for (const node of nodes) {
+    const draft = node.draftValues || {};
+    if (["pending", "running", "queued", "failed", "cancelled", "canceled", "interrupted"].includes(draft.generationStatus))
+      throw Error(`上游「${node.title || node.id}」尚未成功完成，请先完成该节点`);
+    if (node.inputType === "text" || node.kind === "script" || node.kind === "text") {
+      if (!connectedText(node)) throw Error(`上游「${node.title || node.id}」没有文本内容`);
+    } else if (["image", "video", "audio"].includes(node.inputType || "")) {
+      // sourceUrl/poster is an input preview; it is never proof of a generated output.
+      const output = draft.mediaUrl || draft.resultImage || draft.url?.[0] || draft.thumbnailUrl;
+      if (!output) throw Error(`上游「${node.title || node.id}」尚无可用素材，请先上传或生成`);
+    }
+  }
+}
+
+const inputStart = "\n\n<canvas-connected-text>\n";
+const inputEnd = "\n</canvas-connected-text>";
+export function withConnectedText(prompt: string, nodes: ConnectionNode[]) {
+  // Remove only a section created by this resolver, so editing/disconnecting an
+  // upstream text node replaces its previous content on a subsequent run.
+  const start = prompt.indexOf(inputStart);
+  const base = start >= 0 && prompt.endsWith(inputEnd) ? prompt.slice(0, start) : prompt;
+  const text = nodes.map(connectedText).filter(Boolean).join("\n\n");
+  return text ? `${base.trim()}${inputStart}${text}${inputEnd}` : base.trim();
+}
+
+export function connectedImageUrls(lists: string[][], extra: string[] = []) {
+  // Keep edge order and identical URLs in separate slots (first/last frame).
+  return [...lists.flat(), ...extra].map(url => url.trim()).filter(Boolean);
+}
+
+
+export function connectionTypeError(source: ConnectionNode, target: ConnectionNode) {
+  const type = source.inputType || source.draftValues?.mediaType || (source.kind === "text" || source.kind === "script" ? "text" : "");
+  const draft = target.draftValues || {};
+  const key = draft.functionKey || draft.params?.functionKey || draft.targetPreset;
+  const nodeType = draft.nodeType;
+  if (source.kind === "material-effect" || source.kind === "material-role")
+    return target.inputType === "video" && nodeType !== "smart-edit" ? null : "角色和视频特效素材只能连接到视频生成节点";
+  const allowed = nodeType === "smart-edit" || nodeType === "video-clip" || draft.action === "VIDEO_HD_UPSCALE" ? ["video"]
+    : target.inputType === "image" ? ["image", "text"]
+    : target.inputType === "video" ? (key === "audio-to-video" ? ["image", "text", "audio"] : ["image", "text"])
+    : target.inputType === "audio" ? ["text"]
+    : target.inputType === "text" || target.kind === "script" || target.kind === "text" ? ["image", "video", "text"]
+    : null;
+  return allowed && !allowed.includes(type) ? "此节点不能消费该类型的输入，请连接匹配的素材或文本节点" : null;
+}
+
+
+export function videoInputLimitError(mode: string, count: number) {
+  if (mode === "text-to-video" && count) return "已连接图片，请切换图生视频或全能参考模式后生成";
+  if (mode === "first-frame" && count > 1) return "首帧模式只能使用 1 张图片，请移除多余连线或切换模式";
+  if (mode === "first-last-frame" && count > 2) return "首尾帧模式只能使用 2 张图片，请移除多余连线或切换模式";
+  if (mode === "reference-images" && count > 8) return "全能参考最多使用 8 张图片，请减少输入";
+  return null;
+}
+
+
+export function requireResolvedReferenceImages(urls: string[], expectedCount: number) {
+  const images = urls.map(url => url.trim());
+  if (images.length !== expectedCount || images.some(url => !url))
+    throw Error("部分宫格参考图准备失败，尚未提交生成，请重试；不会使用整张原图代替");
+  return images;
+}

lib/video-canvas/group-execution.ts

@@ -0,0 +1,333 @@
+import {
+  assertConnectedInputsReady,
+  connectedInputNodes,
+  connectedText,
+  connectionTypeError,
+  videoInputLimitError,
+  withConnectedText,
+  type ConnectionNode,
+  type CanvasConnection,
+} from "./connections";
+import { groupMembers, isCanvasGroup, type GroupNode } from "./groups";
+import { hasRecoverableCanvasTask } from "./task-recovery";
+import type { CanvasTaskRequest, CanvasGenerationTask } from "./task-contract";
+export function executableGroupNode(node: GroupNode) {
+  const d = node.draftValues || {},
+    action = String(d.action || "");
+  if (isCanvasGroup(node) || node.kind?.startsWith("material-")) return false;
+  if (
+    d.generationTaskId ||
+    d.audioParams ||
+    /GENERATE|UPSCALE|STORY_ANALYSIS|SMART_EDIT/.test(action)
+  )
+    return true;
+  if (
+    /_RESOURCE$/.test(action) ||
+    (node.kind === "text" &&
+      connectedText(node) &&
+      !d.params?.prompt &&
+      !d.prompt)
+  )
+    return false;
+  return (
+    ["model", "workflow", "script", "preset"].includes(node.kind || "") ||
+    !!d.params?.prompt
+  );
+}
+export function compileGroupRun(
+  nodes: GroupNode[],
+  edges: CanvasConnection[],
+  groupId: string,
+  projectId: string,
+  prepare: (node: GroupNode) => Pick<CanvasTaskRequest, "kind" | "params">,
+  uuid: () => string
+) {
+  const members = groupMembers(nodes, groupId).filter(executableGroupNode);
+  if (!members.length) throw Error("编组没有可执行节点");
+  if (members.length > 40) throw Error("整组最多执行 40 个节点");
+  const runId = uuid(),
+    ids = new Map(members.map((n) => [n.id, uuid()]));
+  const requests: CanvasTaskRequest[] = members.map((node) => {
+    const inputs = connectedInputNodes(nodes, edges, node.id).map((source) => {
+      const error = connectionTypeError(source, node);
+      if (error) throw Error(`${node.title}：${error}`);
+      if (!ids.has(source.id)) assertConnectedInputsReady([source]);
+      if (source.draftValues?.gridSplit?.selectedIndices?.length)
+        throw Error("编组执行前请将选中的宫格保存为独立图片节点");
+      return { node: source, requestId: ids.get(source.id) };
+    });
+    return {
+      ...prepare(node),
+      requestId: ids.get(node.id)!,
+      projectId,
+      node: node as CanvasTaskRequest["node"],
+      edges: edges.filter((e) => e.target === node.id),
+      workflow: { runId, groupId, inputs },
+    };
+  });
+  const expanded = requests.flatMap((request) => {
+    if (request.kind !== "smart-edit" || request.params.autoRender !== true)
+      return [request];
+    const planNodeId = `group-plan-${uuid()}`,
+      planRequestId = uuid(),
+      planNode = {
+        ...request.node,
+        id: planNodeId,
+        title: `${request.node.title} · 选片方案`,
+      };
+    const { autoRender, ...params } = request.params;
+    const plan = {
+      ...request,
+      node: planNode,
+      requestId: planRequestId,
+      params,
+      hiddenNodeIds: [planNodeId],
+      edges: request.edges.map((e) => ({
+        ...e,
+        id: uuid(),
+        target: planNodeId,
+      })),
+    };
+    const render = {
+      ...request,
+      params: { operation: "render" },
+      hiddenNodeIds: [planNodeId],
+      workflow: {
+        ...request.workflow!,
+        inputs: [
+          {
+            node: planNode,
+            requestId: planRequestId,
+            output: "edit-plan" as const,
+          },
+        ],
+      },
+      edges: [{ id: uuid(), source: planNodeId, target: request.node.id }],
+    };
+    return [plan, render];
+  });
+  if (expanded.length > 40)
+    throw Error("包含选片阶段后整组超过 40 个任务，请拆分编组");
+  validateGroupBatch(expanded);
+  return expanded;
+}
+export function validateGroupBatch(requests: CanvasTaskRequest[]) {
+  if (!requests.length || requests.length > 40) throw Error("整组任务数量无效");
+  const first = requests[0],
+    map = new Map(requests.map((r) => [r.requestId, r]));
+  if (
+    map.size !== requests.length ||
+    new Set(requests.map((r) => r.node.id)).size !== requests.length
+  )
+    throw Error("重复的整组任务");
+  for (const r of requests) {
+    if (
+      !r.workflow ||
+      r.projectId !== first.projectId ||
+      r.workflow.runId !== first.workflow?.runId ||
+      r.workflow.groupId !== first.workflow?.groupId
+    )
+      throw Error("整组任务不属于同一画布和编组");
+    for (const input of r.workflow.inputs) {
+      if (
+        !r.edges.some(
+          (e) => e.source === input.node.id && e.target === r.node.id
+        )
+      )
+        throw Error("依赖缺少对应连线");
+      if (input.output === "edit-plan") {
+        const source = input.requestId ? map.get(input.requestId) : null;
+        if (
+          r.kind !== "smart-edit" ||
+          r.params.operation !== "render" ||
+          source?.kind !== "smart-edit" ||
+          source.params.operation !== "plan"
+        )
+          throw Error("剪辑方案依赖无效");
+      } else {
+        const error = connectionTypeError(input.node, r.node);
+        if (error) throw Error(error);
+      }
+      if (
+        input.requestId &&
+        map.get(input.requestId)?.node.id !== input.node.id
+      )
+        throw Error("依赖不属于本次编组执行");
+      if (!input.requestId) assertConnectedInputsReady([input.node]);
+    }
+    if (
+      r.edges.some(
+        (e) =>
+          e.target !== r.node.id ||
+          !r.workflow!.inputs.some((i) => i.node.id === e.source)
+      )
+    )
+      throw Error("连线缺少对应输入");
+  }
+  const visited = new Set<string>(),
+    visiting = new Set<string>();
+  const visit = (id: string) => {
+    if (visiting.has(id)) throw Error("编组存在循环依赖");
+    if (visited.has(id)) return;
+    visiting.add(id);
+    for (const input of map.get(id)!.workflow!.inputs)
+      if (input.requestId) visit(input.requestId);
+    visiting.delete(id);
+    visited.add(id);
+  };
+  for (const id of map.keys()) visit(id);
+}
+export function groupDependencyState(
+  request: CanvasTaskRequest,
+  tasks: CanvasGenerationTask[]
+) {
+  const dependencies =
+    request.workflow?.inputs
+      .filter((i) => i.requestId)
+      .map((i) =>
+        tasks.find(
+          (t) =>
+            t.requestId === i.requestId &&
+            t.request.workflow?.runId === request.workflow?.runId
+        )
+      ) || [];
+  if (
+    dependencies.some(
+      (t) =>
+        !t ||
+        ["failed", "cancelled"].includes(t.status) ||
+        (t.status === "interrupted" && !hasRecoverableCanvasTask(t))
+    )
+  )
+    return "failed";
+  return dependencies.every((t) => t?.status === "succeeded")
+    ? "ready"
+    : "waiting";
+}
+export function resolveGroupInputs(
+  request: CanvasTaskRequest,
+  tasks: CanvasGenerationTask[]
+): Record<string, unknown> {
+  if (!request.workflow) return request.params;
+  if (groupDependencyState(request, tasks) !== "ready")
+    throw Error("编组上游任务尚未成功完成");
+  const planInput = request.workflow.inputs.find(
+    (i) => i.output === "edit-plan"
+  );
+  if (planInput) {
+    const plan = tasks.find((t) => t.requestId === planInput.requestId)?.result
+      ?.plans?.[0];
+    if (!plan) throw Error("上游未返回剪辑方案");
+    return { ...request.params, plan };
+  }
+  const nodes = request.workflow.inputs.map((input) => {
+    if (!input.requestId) return input.node;
+    const task = tasks.find((t) => t.requestId === input.requestId)!;
+    const data = task.result || {};
+    const urls =
+      task.request.kind === "image" || task.request.kind === "image-operation"
+        ? (data.images || [])
+            .map((i: any) => i.originalUrl || i.thumbnailUrl)
+            .filter(Boolean)
+        : data.url
+          ? [data.url]
+          : [];
+    const text = data.raw || data.summary || (data.agentPlan || []).join("\n");
+    return {
+      ...input.node,
+      draftValues: {
+        ...input.node.draftValues,
+        generationStatus: "complete",
+        mediaUrl: urls[0],
+        url: urls,
+        content: text,
+        rows: data.analysisRows || data.shotIdeas,
+      },
+    };
+  });
+  assertConnectedInputsReady(nodes);
+  const urls = (type: string) =>
+    nodes
+      .filter((n) => n.inputType === type)
+      .flatMap((n) =>
+        n.draftValues?.url?.length
+          ? n.draftValues.url
+          : [
+              n.draftValues?.mediaUrl ||
+                n.draftValues?.resultImage ||
+                n.draftValues?.thumbnailUrl,
+            ]
+      )
+      .filter(Boolean);
+  const roleImages = nodes
+    .filter((n) => n.kind === "material-role" && n.inputType !== "image")
+    .map((n) => n.draftValues?.thumbnailUrl || n.draftValues?.mediaUrl)
+    .filter(Boolean);
+  const images = [...urls("image"), ...roleImages],
+    videos = urls("video"),
+    audio = urls("audio");
+  const p: Record<string, unknown> = {
+    ...request.params,
+    prompt: withConnectedText(String(request.params.prompt || ""), nodes),
+  };
+  if (request.kind === "image")
+    return {
+      ...p,
+      referenceImage: images[0],
+      referenceImages: images.slice(1),
+    };
+  if (request.kind === "video") {
+    const error = videoInputLimitError(String(p.mode), images.length);
+    if (error) throw Error(error);
+    return {
+      ...p,
+      prompt: [
+        p.prompt,
+        ...nodes.map(
+          (n) => n.draftValues?.rolePrompt || n.draftValues?.effectPrompt
+        ),
+      ]
+        .filter(Boolean)
+        .join("\n\n"),
+      images,
+      audioUrls: audio,
+    };
+  }
+  if (request.kind === "analysis")
+    return { ...p, imageUrls: images, videoUrls: videos };
+  if (request.kind === "audio")
+    return {
+      ...p,
+      text: [request.params.text, ...nodes.map(connectedText)]
+        .filter(Boolean)
+        .join("\n\n"),
+    };
+  if (request.kind === "video-upscale") {
+    if (videos.length !== 1) throw Error("视频高清需要一个上游视频");
+    return { ...p, url: videos[0] };
+  }
+  if (request.kind === "image-operation") {
+    if (images.length !== 1) throw Error("图片处理需要一个上游图片");
+    return { ...p, imageUrl: images[0] };
+  }
+  if (request.kind === "smart-edit") {
+    if (request.params.operation === "plan") {
+      if (!videos.length) throw Error("智能剪辑没有已就绪的视频输入");
+      return {
+        ...p,
+        assets: videos.map((url: string, i: number) => ({
+          id: `group-asset-${i + 1}`,
+          name:
+            nodes.filter((n) => n.inputType === "video")[i]?.title ||
+            `素材 ${i + 1}`,
+          url,
+        })),
+      };
+    }
+    const plan = request.params.plan as any;
+    if (videos.some((url) => !plan?.assets?.some((a: any) => a.url === url)))
+      throw Error("剪辑方案与当前连线素材不一致，请重新确认方案");
+    return p;
+  }
+  throw Error("编组节点暂不支持执行");
+}

lib/video-canvas/groups.ts

@@ -0,0 +1,398 @@
+import { connectionError } from "./connections";
+import type { CanvasConnection, ConnectionNode } from "./connections";
+export type GroupNode = ConnectionNode & {
+  x: number;
+  y: number;
+  width: number;
+  height: number;
+  title: string;
+};
+export const isCanvasGroup = (node: ConnectionNode) =>
+  node.draftValues?.action === "TOOLBOX_TEMPLATE_GROUP";
+export const parentGroup = (node: ConnectionNode) =>
+  String(node.draftValues?.parentGroupId || "");
+export function groupMembers<T extends GroupNode>(
+  nodes: T[],
+  groupId: string,
+  deep = true
+): T[] {
+  const result: T[] = [],
+    seen = new Set([groupId]);
+  const visit = (id: string) => {
+    for (const node of nodes.filter((n) => parentGroup(n) === id)) {
+      if (seen.has(node.id)) throw Error("编组包含循环成员关系");
+      seen.add(node.id);
+      result.push(node);
+      if (deep && isCanvasGroup(node)) visit(node.id);
+    }
+  };
+  visit(groupId);
+  return result;
+}
+export function groupBounds(nodes: GroupNode[]) {
+  if (!nodes.length) return { x: 0, y: 0, width: 400, height: 260 };
+  const x = Math.min(...nodes.map((n) => n.x)) - 32,
+    y = Math.min(...nodes.map((n) => n.y)) - 48;
+  return {
+    x,
+    y,
+    width: Math.max(...nodes.map((n) => n.x + n.width)) - x + 32,
+    height: Math.max(...nodes.map((n) => n.y + n.height)) - y + 32,
+  };
+}
+export function selectedRoots<T extends GroupNode>(nodes: T[], ids: string[]) {
+  const selected = new Set(ids);
+  return nodes.filter(
+    (node) =>
+      selected.has(node.id) &&
+      !nodes.some(
+        (group) =>
+          selected.has(group.id) &&
+          isCanvasGroup(group) &&
+          groupMembers(nodes, group.id).some((member) => member.id === node.id)
+      )
+  );
+}
+export function makeCanvasGroup<T extends GroupNode>(
+  nodes: T[],
+  ids: string[],
+  id: string,
+  title = "未命名编组"
+) {
+  const members = selectedRoots(nodes, ids);
+  if (members.length < 2) throw Error("请至少选择两个节点进行编组");
+  const parents = new Set(members.map(parentGroup)),
+    parent = parents.size === 1 ? parentGroup(members[0]) : "";
+  const group = {
+    id,
+    kind: "model",
+    title,
+    ...groupBounds(members),
+    accent: "#a0a0a0",
+    draftValues: {
+      action: "TOOLBOX_TEMPLATE_GROUP",
+      nodeType: "group",
+      parentGroupId: parent,
+      groupColor: "#a0a0a0",
+      groupVersion: 1,
+    },
+  };
+  const selected = new Set(members.map((n) => n.id));
+  return [
+    ...nodes.map((n) =>
+      selected.has(n.id)
+        ? { ...n, draftValues: { ...n.draftValues, parentGroupId: id } }
+        : n
+    ),
+    group as unknown as T,
+  ];
+}
+export function ungroupCanvasNodes<T extends GroupNode>(
+  nodes: T[],
+  groupId: string
+) {
+  const group = nodes.find((n) => n.id === groupId);
+  if (!group || !isCanvasGroup(group)) throw Error("未找到编组");
+  return nodes
+    .filter((n) => n.id !== groupId)
+    .map((n) =>
+      parentGroup(n) === groupId
+        ? {
+            ...n,
+            draftValues: {
+              ...n.draftValues,
+              parentGroupId: parentGroup(group),
+            },
+          }
+        : n
+    );
+}
+export function moveCanvasSelection<T extends GroupNode>(
+  nodes: T[],
+  ids: string[],
+  dx: number,
+  dy: number
+) {
+  const moving = new Set(ids);
+  for (const id of ids)
+    if (nodes.some((n) => n.id === id && isCanvasGroup(n)))
+      for (const child of groupMembers(nodes, id)) moving.add(child.id);
+  let result = nodes.map((n) =>
+    moving.has(n.id) ? { ...n, x: n.x + dx, y: n.y + dy } : n
+  );
+  // Expand stationary ancestor frames when a member is dragged outside them.
+  for (let pass = 0; pass < nodes.length; pass++) {
+    let changed = false;
+    result = result.map((group) => {
+      if (!isCanvasGroup(group) || moving.has(group.id)) return group;
+      const children = groupMembers(result, group.id, false);
+      if (!children.length) return group;
+      const bound = groupBounds(children),
+        x = Math.min(group.x, bound.x),
+        y = Math.min(group.y, bound.y);
+      const width = Math.max(group.x + group.width, bound.x + bound.width) - x,
+        height = Math.max(group.y + group.height, bound.y + bound.height) - y;
+      if (
+        x === group.x &&
+        y === group.y &&
+        width === group.width &&
+        height === group.height
+      )
+        return group;
+      changed = true;
+      return { ...group, x, y, width, height };
+    });
+    if (!changed) break;
+  }
+  return result;
+}
+export function layoutCanvasGroup<T extends GroupNode>(
+  nodes: T[],
+  groupId: string,
+  layout: "grid" | "horizontal" | "vertical"
+) {
+  const group = nodes.find((n) => n.id === groupId);
+  if (!group) throw Error("编组不存在");
+  const members = groupMembers(nodes, groupId, false);
+  const cols =
+    layout === "horizontal"
+      ? members.length
+      : layout === "vertical"
+        ? 1
+        : Math.max(1, Math.ceil(Math.sqrt(members.length)));
+  const cellWidth = Math.max(180, ...members.map((n) => n.width)) + 48,
+    cellHeight = Math.max(120, ...members.map((n) => n.height)) + 64;
+  let result = nodes;
+  for (const [i, member] of members.entries())
+    result = moveCanvasSelection(
+      result,
+      [member.id],
+      group.x + 32 + (i % cols) * cellWidth - member.x,
+      group.y + 48 + Math.floor(i / cols) * cellHeight - member.y
+    );
+  const bounds = groupBounds(groupMembers(result, groupId, false));
+  return result.map((n) => (n.id === groupId ? { ...n, ...bounds } : n));
+}
+export function migrateLegacyGroups<T extends GroupNode>(nodes: T[]) {
+  // Old preset frames stored geometry but no membership. Infer once on load.
+  const groups = nodes.filter(
+    (n) => isCanvasGroup(n) && !n.draftValues?.groupVersion
+  );
+  return nodes.map((n) => {
+    if (isCanvasGroup(n))
+      return { ...n, draftValues: { ...n.draftValues, groupVersion: 1 } };
+    if (parentGroup(n)) return n;
+    const enclosing = groups
+      .filter(
+        (g) =>
+          n.x >= g.x &&
+          n.y >= g.y &&
+          n.x + n.width <= g.x + g.width &&
+          n.y + n.height <= g.y + g.height
+      )
+      .sort((a, b) => a.width * a.height - b.width * b.height)[0];
+    return enclosing
+      ? { ...n, draftValues: { ...n.draftValues, parentGroupId: enclosing.id } }
+      : n;
+  });
+}
+export function exportCanvasGroup<T extends GroupNode>(
+  nodes: T[],
+  edges: CanvasConnection[],
+  groupId: string
+) {
+  const group = nodes.find((n) => n.id === groupId);
+  if (!group) throw Error("编组不存在");
+  let all = [group, ...groupMembers(nodes, groupId)];
+  const ids = new Set(all.map((n) => n.id));
+  const externalIds = [
+    ...new Set(
+      edges
+        .filter((e) => ids.has(e.target) && !ids.has(e.source))
+        .map((e) => e.source)
+    ),
+  ];
+  const inputs = externalIds.map((id, index) => {
+    const source = nodes.find((n) => n.id === id);
+    if (
+      !source ||
+      !["image", "video", "audio", "text"].includes(source.inputType || "")
+    )
+      throw Error("请先将组外角色、特效等上游节点加入编组，再保存模板");
+    return {
+      id,
+      title: `输入：${source.title}`.slice(0, 120),
+      kind: "input",
+      inputType: source.inputType,
+      x: group.x + 32,
+      y: group.y + 48 + index * 210,
+      width: 240,
+      height: 160,
+      accent: "#aaaaaa",
+      draftValues: {
+        action: `${source.inputType!.toUpperCase()}_RESOURCE`,
+        parentGroupId: group.id,
+        mediaType: source.inputType,
+        url: [],
+        content: [],
+        generated: false,
+        generationStatus: "idle",
+        templateInput: true,
+      },
+    } as unknown as T;
+  });
+  if (inputs.length) {
+    all = all.map((n) =>
+      n.id === group.id
+        ? {
+            ...n,
+            width: n.width + 280,
+            height: Math.max(n.height, inputs.length * 210 + 80),
+          }
+        : { ...n, x: n.x + 280 }
+    );
+    all.push(...inputs);
+    for (const input of inputs) ids.add(input.id);
+  }
+  return {
+    version: 1,
+    name: group.title,
+    nodes: all.map((n) => ({
+      ...n,
+      draftValues: {
+        ...n.draftValues,
+        params: n.draftValues?.params
+          ? {
+              ...n.draftValues.params,
+              taskId: undefined,
+              sourceNodeId: undefined,
+            }
+          : undefined,
+        ...(n.id === groupId ? { parentGroupId: "" } : {}),
+        generationTaskId: undefined,
+        taskRequestId: undefined,
+        appliedTaskStatus: undefined,
+        generationStatus:
+          n.draftValues?.mediaUrl || n.draftValues?.content
+            ? "complete"
+            : "idle",
+        generationError: undefined,
+        pendingLabel: undefined,
+      },
+    })),
+    edges: edges.filter((e) => ids.has(e.source) && ids.has(e.target)),
+  };
+}
+export function instantiateCanvasGroup<T extends GroupNode>(
+  template: { nodes: T[]; edges: CanvasConnection[] },
+  id: () => string,
+  x: number,
+  y: number
+) {
+  if (
+    !template ||
+    !Array.isArray(template.nodes) ||
+    !template.nodes.length ||
+    template.nodes.length > 100
+  )
+    throw Error("编组模板节点数量无效");
+  if (
+    !Array.isArray(template.edges) ||
+    template.edges.length > 300 ||
+    new Set(template.nodes.map((n) => n.id)).size !== template.nodes.length ||
+    template.nodes.some(
+      (n) =>
+        !n.id ||
+        ![n.x, n.y, n.width, n.height].every(Number.isFinite) ||
+        n.width <= 0 ||
+        n.height <= 0 ||
+        n.width > 10000 ||
+        n.height > 10000
+    )
+  )
+    throw Error("编组模板数据无效");
+  for (const node of template.nodes) {
+    if (
+      parentGroup(node) &&
+      !template.nodes.some(
+        (n) => n.id === parentGroup(node) && isCanvasGroup(n)
+      )
+    )
+      throw Error("编组模板成员关系无效");
+    if (isCanvasGroup(node)) groupMembers(template.nodes, node.id);
+  }
+  for (const edge of template.edges) {
+    const error = connectionError(
+      template.nodes,
+      template.edges.filter((e) => e !== edge),
+      edge.source,
+      edge.target
+    );
+    if (error) throw Error(error);
+  }
+  const map = new Map(template.nodes.map((n) => [n.id, id()]));
+  const root = template.nodes.find((n) => isCanvasGroup(n) && !parentGroup(n));
+  if (!root) throw Error("编组模板缺少根组");
+  if (
+    groupMembers(template.nodes, root.id).length !==
+    template.nodes.length - 1
+  )
+    throw Error("模板必须只有一个完整根组");
+  const nodes = template.nodes.map((n) => ({
+    ...n,
+    id: map.get(n.id)!,
+    x: n.x + x - root.x,
+    y: n.y + y - root.y,
+    draftValues: {
+      ...n.draftValues,
+      parentGroupId: map.get(parentGroup(n)) || "",
+      shotNodeIds: n.draftValues?.shotNodeIds
+        ?.map((key: string) => map.get(key))
+        .filter(Boolean),
+      sourceNodeId: map.get(n.draftValues?.sourceNodeId) || undefined,
+      sourceVideoNodeId: map.get(n.draftValues?.sourceVideoNodeId) || undefined,
+    },
+  }));
+  return {
+    nodes,
+    edges: template.edges
+      .map((e) => ({
+        ...e,
+        id: id(),
+        source: map.get(e.source)!,
+        target: map.get(e.target)!,
+      }))
+      .filter((e) => e.source && e.target),
+    groupId: map.get(root.id)!,
+  };
+}
+
+export function assignGroupMembers<T extends GroupNode>(
+  nodes: T[],
+  ids: string[],
+  groupId: string
+) {
+  if (groupId && !nodes.some((n) => n.id === groupId && isCanvasGroup(n)))
+    throw Error("编组不存在");
+  const roots = selectedRoots(nodes, ids);
+  if (
+    roots.some(
+      (n) =>
+        n.id === groupId ||
+        (isCanvasGroup(n) &&
+          groupMembers(nodes, n.id).some((child) => child.id === groupId))
+    )
+  )
+    throw Error("不能将编组加入自身或自己的子组");
+  const selected = new Set(roots.map((n) => n.id));
+  return moveCanvasSelection(
+    nodes.map((n) =>
+      selected.has(n.id)
+        ? { ...n, draftValues: { ...n.draftValues, parentGroupId: groupId } }
+        : n
+    ),
+    [],
+    0,
+    0
+  );
+}

lib/video-canvas/smart-edit/audio-process.ts

@@ -0,0 +1,178 @@
+import { execFile } from "node:child_process";
+import { promisify } from "node:util";
+import { createHash, randomUUID } from "node:crypto";
+import { readFile } from "node:fs/promises";
+import path from "node:path";
+import {
+  SECOND,
+  editPlanSchema,
+  planTimeline,
+  editableCaptions,
+  newCaptionStyle,
+  type EditPlan,
+  type TimelineCaption,
+} from "./contract";
+import {
+  materialize,
+  probeMedia,
+  ffmpeg,
+  mediaSafety,
+  prepareAudio,
+  pythonAnalysis,
+} from "./local";
+
+export async function processEditAudio(
+  operation: "separate" | "transcribe",
+  plan: EditPlan,
+  clipId: string | undefined,
+  userId: string,
+  dir: string,
+  progress: (stage: string) => Promise<void>,
+  upload: (file: string, suffix: string) => Promise<string>,
+  track: "video" | "audio" = "video",
+  keep: "both" | "vocals" | "background" = "both",
+): Promise<EditPlan> {
+  const timeline = planTimeline(plan);
+  if (operation === "separate") {
+    const index = plan.clips.findIndex((c) => c.id === clipId);
+    const audio = track === "audio" ? plan.audioClips?.find((c) => c.id === clipId) : undefined;
+    const clip = track === "audio" ? audio : plan.clips[index];
+    if (!clip || (track === "video" && plan.clips[index].audioDetached))
+      throw Error("请选择可分离的音视频片段");
+    if ((plan.audioClips?.length || 0) + (keep === "both" ? 2 : 1) - (audio ? 1 : 0) > 24)
+      throw Error("分离需要空闲音轨，最多支持 24 个音频片段");
+    const asset = plan.assets.find((a) => a.id === clip.assetId)!;
+    if (!asset.hasAudio) throw Error("当前视频没有音轨");
+    await progress("准备人声分离素材");
+    const media = await materialize(asset.url, userId);
+    if (media.hash !== asset.hash) throw Error("原素材内容已变化，请重新解析");
+    const actual = await probeMedia(media.file);
+    if (!actual.hasAudio || Math.abs(actual.duration - asset.duration) > 100000)
+      throw Error("原音轨参数已变化，请重新解析");
+    const inputAudio = path.join(dir, "separation-input.wav");
+    await ffmpeg([
+      ...mediaSafety,
+      "-ss",
+      String(clip.sourceIn / SECOND),
+      "-i",
+      media.file,
+      "-t",
+      String((clip.sourceOut - clip.sourceIn) / SECOND),
+      "-vn",
+      "-af",
+      "apad",
+      "-ac",
+      "2",
+      "-ar",
+      "44100",
+      "-c:a",
+      "pcm_f32le",
+      inputAudio,
+    ]);
+    await progress("正在分离人声与背景声");
+    try {
+      await promisify(execFile)(
+        process.env.SMART_EDIT_PYTHON || "python3",
+        [path.resolve("services/smart-edit/separate.py"), inputAudio, dir],
+        { timeout: 600000, maxBuffer: 2 * 1024 * 1024 },
+      );
+    } catch (e) {
+      const detail = String((e as { stderr?: string }).stderr || "");
+      throw Error(
+        detail.includes("尚未配置")
+          ? "人声分离模型尚未配置"
+          : "人声分离失败或超时，原音轨已保留",
+      );
+    }
+    await progress("保存人声与背景声音轨");
+    const audioClips = (plan.audioClips || []).filter((c) => !audio || c.id !== audio.id);
+    const used = new Set([...plan.clips, ...audioClips].map((c) => c.assetId));
+    const assets = plan.assets.filter((a) => used.has(a.id));
+    const t = timeline.clips[index],
+      next = timeline.clips[index + 1];
+    for (const [stem, title] of [
+      ["vocals", "人声"],
+      ["background", "背景声"],
+    ]) {
+      if (keep !== "both" && keep !== stem) continue;
+      const file = path.join(dir, stem + ".wav"),
+        metadata = await probeMedia(file);
+      if (Math.abs(metadata.duration - (clip.sourceOut - clip.sourceIn)) > 50000)
+        throw Error("分离音轨时长异常，原音轨已保留");
+      const id = randomUUID(),
+        bytes = await readFile(file);
+      assets.push({
+        id,
+        name: `${asset.name.slice(0, 180)} · ${title}`,
+        url: await upload(file, "wav"),
+        hash: createHash("sha256").update(bytes).digest("hex"),
+        subtitles: [],
+        ...metadata,
+      });
+      audioClips.push({
+        id: randomUUID(),
+        assetId: id,
+        start: audio ? audio.start : t.start,
+        sourceIn: 0,
+        sourceOut: clip.sourceOut - clip.sourceIn,
+        speed: clip.speed,
+        gain: clip.gain,
+        muted: clip.muted,
+        fadeIn: audio ? audio.fadeIn : (t.overlap - t.padBefore) / SECOND,
+        fadeInStart: audio ? audio.fadeInStart : t.overlap ? t.padBefore / t.overlap : 0,
+        fadeOut: audio ? audio.fadeOut : next ? (next.overlap - t.padAfter) / SECOND : 0,
+        fadeOutEnd: audio ? audio.fadeOutEnd : next?.overlap ? t.padAfter / next.overlap : 0,
+      });
+    }
+    return editPlanSchema.parse({
+      ...plan,
+      assets,
+      audioClips,
+      clips: plan.clips.map((c) =>
+        !audio && c.id === clip.id ? { ...c, audioDetached: true } : c,
+      ),
+    });
+  }
+  const captions: TimelineCaption[] = editableCaptions(plan).filter(
+    (c) => !c.id.startsWith("asr-"),
+  );
+  let recognized = 0;
+  for (const [index, clip] of timeline.clips.entries()) {
+    const asset = plan.assets.find((a) => a.id === clip.assetId)!;
+    if (!asset.hasAudio) continue;
+    await progress(`自动转写 ${index + 1}/${plan.clips.length}`);
+    const media = await materialize(asset.url, userId);
+    if (media.hash !== asset.hash) throw Error("原素材内容已变化，请重新解析");
+    const actual = await probeMedia(media.file);
+    if (!actual.hasAudio || Math.abs(actual.duration - asset.duration) > 100000)
+      throw Error("原音轨参数已变化，请重新解析");
+    const audio = await prepareAudio(media.file, path.join(media.dir, media.hash + ".wav"));
+    const analysis = await pythonAnalysis({ version: 1, audio, transcribe: true }, media.dir);
+    for (const [j, s] of analysis.subtitles.entries()) {
+      const from = Math.max(s.start, clip.sourceIn),
+        to = Math.min(s.end, clip.sourceOut);
+      if (to <= from) continue;
+      const start = Math.round(clip.start + (from - clip.sourceIn) / clip.speed);
+      const end = Math.min(
+        timeline.duration,
+        Math.round(clip.start + (to - clip.sourceIn) / clip.speed),
+      );
+      if (end <= start) continue;
+      captions.push({
+        id: `asr-${clip.id}-${j}`,
+        start,
+        end,
+        text: s.text,
+        style: { ...newCaptionStyle },
+      });
+      recognized++;
+    }
+  }
+  if (!recognized) throw Error("未识别到有效人声，原字幕已保留");
+  if (captions.length > 500) throw Error("转写后字幕超过 500 条，请缩短素材");
+  return editPlanSchema.parse({
+    ...plan,
+    subtitles: true,
+    captions: captions.sort((a, b) => a.start - b.start),
+  });
+}

lib/video-canvas/smart-edit/contract.ts

@@ -1,5 +1,9 @@
 import { z } from "zod";
+import { captionFont } from "./fonts";
 export const SECOND = 1_000_000;
+export const MAX_EDIT_SECONDS = 300;
+export const editRatios = ["21:9", "16:9", "9:16", "1:1", "4:3", "3:4"] as const;
+export const editResolutions = ["480p", "720p", "1080p"] as const;
 const time = z
   .number()
   .int()
@@ -14,6 +18,55 @@ const url = z
 export const subtitleSchema = z
   .object({ start: time, end: time, text: z.string().max(500) })
   .refine((x) => x.end > x.start);
+export const captionStyleSchema = z.object({
+  fontFamily: z
+    .string()
+    .max(180)
+    .refine((id) => Boolean(captionFont(id)), "未知字幕字体")
+    .optional(),
+  referenceHeight: z.union([z.literal(720), z.literal(1080)]).optional(),
+  fontSize: z.number().min(8).max(160).default(28),
+  color: z
+    .string()
+    .regex(/^#[0-9a-f]{6}$/i)
+    .default("#ffffff"),
+  outlineColor: z
+    .string()
+    .regex(/^#[0-9a-f]{6}$/i)
+    .default("#000000"),
+  outlineWidth: z.number().min(0).max(20).default(0),
+  opacity: z.number().min(0).max(1).default(1),
+  // Optional so existing edits keep their whole-caption opacity unchanged.
+  outlineOpacity: z.number().min(0).max(1).optional(),
+  spacing: z.number().min(-5).max(100).default(0),
+  x: z.number().min(0).max(1).optional(),
+  y: z.number().min(0).max(1).optional(),
+});
+export const timelineCaptionSchema = z
+  .object({
+    id,
+    start: time,
+    end: time,
+    text: z.string().max(500),
+    style: captionStyleSchema.optional(),
+  })
+  .refine((x) => x.end > x.start);
+export type TimelineCaption = z.infer<typeof timelineCaptionSchema>;
+export type CaptionStyle = z.infer<typeof captionStyleSchema>;
+export const defaultCaptionStyle: CaptionStyle = captionStyleSchema.parse({});
+export const newCaptionStyle: CaptionStyle = {
+  ...defaultCaptionStyle,
+  fontFamily: "source-han",
+  referenceHeight: 1080,
+  fontSize: 56,
+  outlineColor: "#ffffff",
+};
+const transitionSchema = z.object({
+  duration: z.number().min(0).max(4),
+  // Centered transitions preserve the edit duration; absent keeps saved legacy overlaps.
+  alignment: z.enum(["overlap", "center"]).optional(),
+  type: z.enum(["fade", "fadeblack", "fadewhite"]),
+});
 export const assetInputSchema = z.object({
   id,
   url,
@@ -22,14 +75,21 @@ export const assetInputSchema = z.object({
 });
 export const editAssetSchema = assetInputSchema.extend({
   duration: time,
-  width: z.number().int().positive().max(8192),
+  width: z.number().int().min(0).max(8192),
-  height: z.number().int().positive().max(8192),
+  height: z.number().int().min(0).max(8192),
   hasAudio: z.boolean(),
   hash: z.string().regex(/^[a-f0-9]{64}$/),
 });
 export const candidateSchema = z
   .object({ id, assetId: id, start: time, end: time, text: z.string().max(1500).default("") })
   .refine((x) => x.end - x.start >= 200000);
+export const videoTransformSchema = z.object({
+  x: z.number().min(-1).max(1).default(0),
+  y: z.number().min(-1).max(1).default(0),
+  scale: z.number().min(0.1).max(3).default(1),
+});
+export type VideoTransform = z.infer<typeof videoTransformSchema>;
+export const defaultVideoTransform: VideoTransform = { x: 0, y: 0, scale: 1 };
 export const editClipSchema = z
   .object({
     id,
@@ -37,19 +97,43 @@ export const editClipSchema = z
     sourceIn: time,
     sourceOut: time,
     speed: z.number().min(0.5).max(2).default(1),
-    gain: z.number().min(0).max(1).default(1),
+    gain: z.number().min(0).max(4).default(1),
+    muted: z.boolean().optional(),
     label: z.string().max(300).default(""),
+    // The transition belongs to the incoming clip, so it follows that clip when reordered.
+    transition: transitionSchema.optional(),
+    transform: videoTransformSchema.optional(),
+    audioDetached: z.boolean().optional(),
   })
   .refine((x) => x.sourceOut - x.sourceIn >= 200000);
+export const audioClipSchema = z
+  .object({
+    id,
+    assetId: id,
+    start: time,
+    sourceIn: time,
+    sourceOut: time,
+    speed: z.number().min(0.5).max(2).default(1),
+    gain: z.number().min(0).max(4).default(1),
+    muted: z.boolean().optional(),
+    fadeIn: z.number().min(0).max(5).default(0),
+    fadeOut: z.number().min(0).max(5).default(0),
+    fadeInStart: z.number().min(0).max(1).optional(),
+    fadeOutEnd: z.number().min(0).max(1).optional(),
+  })
+  .refine((c) => c.sourceOut - c.sourceIn >= 200000);
+export type AudioClip = z.infer<typeof audioClipSchema>;
 export const editPlanSchema = z.object({
   version: z.literal(1),
   title: z.string().max(200),
-  assets: z.array(editAssetSchema).min(1).max(6),
+  assets: z.array(editAssetSchema).min(1).max(36),
   clips: z.array(editClipSchema).min(1).max(12),
-  ratio: z.enum(["16:9", "9:16", "1:1"]),
+  ratio: z.enum(editRatios),
-  resolution: z.enum(["720p", "1080p"]).default("720p"),
+  resolution: z.enum(editResolutions).default("720p"),
   transition: z.number().min(0).max(0.5).default(0),
   subtitles: z.boolean().default(true),
+  captions: z.array(timelineCaptionSchema).max(500).optional(),
+  audioClips: z.array(audioClipSchema).max(24).optional(),
   music: z
     .object({ url, gain: z.number().min(0).max(1).default(0.2), start: time.default(0) })
     .optional(),
@@ -61,12 +145,12 @@ export type EditCandidate = z.infer<typeof candidateSchema>;
 export const smartEditRequestSchema = z.discriminatedUnion("operation", [
   z.object({
     operation: z.literal("plan"),
-    assets: z.array(assetInputSchema).min(1).max(6),
+    assets: z.array(assetInputSchema).min(1).max(12),
-    mode: z.enum(["montage", "talk", "explain", "ads"]).default("montage"),
+    mode: z.enum(["sequence", "auto", "montage", "talk", "explain", "ads"]).default("montage"),
     prompt: z.string().max(3000).default(""),
-    targetDuration: z.number().int().min(3).max(90).default(20),
+    targetDuration: z.number().int().min(0).max(MAX_EDIT_SECONDS).default(20),
-    ratio: z.enum(["16:9", "9:16", "1:1"]).default("16:9"),
+    ratio: z.enum(editRatios).default("16:9"),
-    resolution: z.enum(["720p", "1080p"]).default("720p"),
+    resolution: z.enum(editResolutions).default("720p"),
     removeSilence: z.boolean().default(false),
     transcribe: z.boolean().default(false),
     semantic: z.boolean().default(true),
@@ -78,6 +162,17 @@ export const smartEditRequestSchema = z.discriminatedUnion("operation", [
     beatsPerCut: z.enum(["2", "4"]).default("4"),
     variants: z.number().int().min(1).max(3).default(1),
   }),
+  z.object({
+    operation: z.literal("separate"),
+    plan: editPlanSchema,
+    clipId: id,
+    track: z.enum(["video", "audio"]).default("video"),
+    keep: z.enum(["both", "vocals", "background"]).default("both"),
+  }),
+  z.object({
+    operation: z.literal("transcribe"),
+    plan: editPlanSchema,
+  }),
   z.object({
     operation: z.literal("render"),
     plan: editPlanSchema,
@@ -85,38 +180,102 @@ export const smartEditRequestSchema = z.discriminatedUnion("operation", [
   }),
 ]);
 export type SmartEditRequest = z.infer<typeof smartEditRequestSchema>;
+export function editOutputSize(plan: Pick<EditPlan, "ratio" | "resolution">) {
+  const size = Number.parseInt(plan.resolution, 10),
+    [w, h] = plan.ratio.split(":").map(Number);
+  const even = (v: number) => Math.round(v / 2) * 2;
+  return w >= h
+    ? { width: even((size * w) / h), height: size }
+    : { width: size, height: even((size * h) / w) };
+}
 export function planTimeline(input: EditPlan) {
   const plan = editPlanSchema.parse(input),
     ids = new Set<string>();
+  const durations = plan.clips.map((c) =>
+    Math.round(
+      (Math.round(((c.sourceOut - c.sourceIn) / c.speed / SECOND) * 30) * SECOND) / 30,
+    ),
+  );
+  const centered = plan.clips.map((c) => c.transition?.alignment === "center");
+  const overlaps = plan.clips.map((c, i) =>
+    i
+      ? Math.min(
+          Math.round((c.transition?.duration ?? plan.transition) * SECOND),
+          centered[i] ? durations[i] * 2 : Math.floor(durations[i] / 3),
+          centered[i] ? durations[i - 1] * 2 : Math.floor(durations[i - 1] / 3),
+        )
+      : 0,
+  );
+  // Neighboring transitions cannot consume the same frames in a short middle clip.
+  for (let i = 0; i < durations.length; i++) {
+    const incoming = overlaps[i] / (centered[i] ? 2 : 1),
+      outgoing = (overlaps[i + 1] || 0) / (centered[i + 1] ? 2 : 1);
+    if (incoming + outgoing > durations[i]) {
+      const factor = durations[i] / (incoming + outgoing);
+      overlaps[i] = Math.floor(overlaps[i] * factor);
+      if (i + 1 < overlaps.length) overlaps[i + 1] = Math.floor(overlaps[i + 1] * factor);
+    }
+  }
+  const before = overlaps.map((d, i) =>
+    centered[i] ? Math.round((Math.floor(((d / SECOND) * 30) / 2) * SECOND) / 30) : 0,
+  );
+  const after = overlaps.map((d, i) => (centered[i] ? d - before[i] : 0));
   let cursor = 0;
   const clips = plan.clips.map((clip, index) => {
     if (ids.has(clip.id)) throw Error("片段 ID 重复");
     ids.add(clip.id);
     const asset = plan.assets.find((a) => a.id === clip.assetId);
-    if (!asset || clip.sourceOut > asset.duration) throw Error("片段超出原始素材范围");
+    if (!asset || !asset.width || !asset.height || clip.sourceOut > asset.duration)
-    const duration = Math.round(
+      throw Error("片段超出原始素材范围");
-      (Math.round(((clip.sourceOut - clip.sourceIn) / clip.speed / SECOND) * 30) * SECOND) / 30,
+    const duration = durations[index],
-    );
+      overlap = overlaps[index];
-    const previous = index ? plan.clips[index - 1] : null;
+    if (!centered[index]) cursor -= overlap;
-    const overlap = previous
-      ? Math.min(
-          Math.round(plan.transition * SECOND),
-          Math.floor(duration / 3),
-          Math.floor((previous.sourceOut - previous.sourceIn) / previous.speed / 3),
-        )
-      : 0;
-    cursor -= overlap;
     const start = cursor;
     cursor += duration;
-    return { ...clip, start, end: cursor, duration, overlap };
+    const padBefore = before[index],
+      padAfter = after[index + 1] || 0;
+    return {
+      ...clip,
+      start,
+      end: cursor,
+      duration,
+      overlap,
+      padBefore,
+      padAfter,
+      renderStart: start - padBefore,
+      renderEnd: cursor + padAfter,
+      transitionStart: start - padBefore,
+      transitionType: clip.transition?.type || "fade",
+    };
   });
   if (new Set(plan.assets.map((a) => a.id)).size !== plan.assets.length)
     throw Error("素材 ID 重复");
-  if (cursor > 90 * SECOND || cursor < 200000) throw Error("成片时长需在 0.2–90 秒内");
+  if (plan.captions && new Set(plan.captions.map((s) => s.id)).size !== plan.captions.length)
-  return { clips, duration: cursor };
+    throw Error("字幕 ID 重复");
+  if (cursor > MAX_EDIT_SECONDS * SECOND || cursor < 200000)
+    throw Error("成片时长需在 0.2–300 秒内");
+  const audioClips = (plan.audioClips || []).map((c) => {
+    const asset = plan.assets.find((a) => a.id === c.assetId);
+    if (!asset?.hasAudio || c.sourceOut > asset.duration)
+      throw Error("独立音轨素材无效或超出范围");
+    const duration = Math.round((c.sourceOut - c.sourceIn) / c.speed);
+    return { ...c, duration, end: c.start + duration };
+  });
+  if (new Set(audioClips.map((c) => c.id)).size !== audioClips.length)
+    throw Error("独立音轨 ID 重复");
+  cursor = Math.max(cursor, ...audioClips.map((c) => c.end));
+  if (cursor > MAX_EDIT_SECONDS * SECOND) throw Error("成片时长需在 0.2–300 秒内");
+  return { clips, audioClips, duration: cursor };
 }
 export function compiledSubtitles(plan: EditPlan) {
   if (!plan.subtitles) return [];
+  if (plan.captions) {
+    const duration = planTimeline(plan).duration;
+    return plan.captions.flatMap((s) => {
+      const end = Math.min(s.end, duration);
+      return end > s.start ? [{ ...s, end }] : [];
+    });
+  }
   return planTimeline(plan).clips.flatMap((clip) =>
     plan.assets
       .find((a) => a.id === clip.assetId)!
@@ -124,16 +283,63 @@ export function compiledSubtitles(plan: EditPlan) {
         const a = Math.max(s.start, clip.sourceIn),
           b = Math.min(s.end, clip.sourceOut);
         if (b <= a) return [];
-        // A crossfade's incoming subtitle starts when the outgoing subtitle ends.
+        // Only legacy overlapping edits delay incoming captions until the outgoing clip ends.
         const start = Math.max(
-          clip.start + clip.overlap,
+          clip.start + (clip.transition?.alignment === "center" ? 0 : clip.overlap),
           clip.start + Math.round((a - clip.sourceIn) / clip.speed),
         );
-        const end = Math.min(clip.end, clip.start + Math.round((b - clip.sourceIn) / clip.speed));
+        const end = Math.min(
+          clip.end,
+          clip.start + Math.round((b - clip.sourceIn) / clip.speed),
+        );
         return end > start ? [{ start, end, text: s.text }] : [];
       }),
   );
 }
+export function editableCaptions(plan: EditPlan): TimelineCaption[] {
+  if (plan.captions) return plan.captions;
+  return compiledSubtitles({ ...plan, subtitles: true }).map((s, i) => ({
+    ...s,
+    id: `caption-${i}`,
+  }));
+}
+export function subtitlesToAss(plan: EditPlan, width: number, height: number) {
+  const stamp = (v: number) => {
+    const cs = Math.round(v / 10000);
+    return `${Math.floor(cs / 360000)}:${String(Math.floor(cs / 6000) % 60).padStart(2, "0")}:${String(Math.floor(cs / 100) % 60).padStart(2, "0")}.${String(cs % 100).padStart(2, "0")}`;
+  };
+  const color = (hex: string, opacity: number) =>
+    `&H${Math.round((1 - opacity) * 255)
+      .toString(16)
+      .padStart(2, "0")}${hex.slice(5, 7)}${hex.slice(3, 5)}${hex.slice(1, 3)}`;
+  // Saved legacy styles use 720px; new reference-aligned styles explicitly use 1080px.
+  const rows = compiledSubtitles(plan) as Array<{
+    start: number;
+    end: number;
+    text: string;
+    style?: CaptionStyle;
+  }>;
+  const styles = rows.map((s, i) => {
+    const style = s.style || defaultCaptionStyle;
+    const scale = height / (style.referenceHeight || 720);
+    return `Style: c${i},${captionFont(style.fontFamily)?.family || "Noto Sans CJK SC"},${style.fontSize * scale},${color(style.color, style.opacity)},&H00FFFFFF,${color(style.outlineColor, style.opacity * (style.outlineOpacity ?? 1))},&HFF000000,0,0,0,0,100,100,${style.spacing * scale},0,1,${style.outlineWidth * scale},0,2,20,20,${Math.round(height * 0.05)},1`;
+  });
+  const events = rows.map((s, i) => {
+    const text = s.text
+      .replace(/\\/g, "＼")
+      .replace(/\{/g, "｛")
+      .replace(/\}/g, "｝")
+      .replace(/\r/g, "")
+      .replace(/\n/g, "\\N");
+    const style = s.style || defaultCaptionStyle;
+    const position =
+      style.x !== undefined || style.y !== undefined
+        ? `{\\pos(${Math.round((style.x ?? 0.5) * width)},${Math.round((style.y ?? 0.95) * height)})}`
+        : "";
+    return `Dialogue: 0,${stamp(s.start)},${stamp(s.end)},c${i},,0,0,0,,${position}${text}`;
+  });
+  return `[Script Info]\nScriptType: v4.00+\nPlayResX: ${width}\nPlayResY: ${height}\nWrapStyle: 0\nScaledBorderAndShadow: yes\n\n[V4+ Styles]\nFormat: Name, Fontname, Fontsize, PrimaryColour, SecondaryColour, OutlineColour, BackColour, Bold, Italic, Underline, StrikeOut, ScaleX, ScaleY, Spacing, Angle, BorderStyle, Outline, Shadow, Alignment, MarginL, MarginR, MarginV, Encoding\n${styles.join("\n")}\n\n[Events]\nFormat: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text\n${events.join("\n")}\n`;
+}
 export function subtitlesToSrt(subtitles: z.infer<typeof subtitleSchema>[]) {
   const stamp = (v: number) => {
     const ms = Math.round(v / 1000);
@@ -194,22 +400,39 @@ export function buildEditPlans(
   if (!ordered.length) throw Error("没有选中可用片段");
   return Array.from({ length: params.mode === "ads" ? params.variants : 1 }, (_, variant) => {
     const list = variant
-      ? [...ordered.slice(variant % ordered.length), ...ordered.slice(0, variant % ordered.length)]
+      ? [
+          ...ordered.slice(variant % ordered.length),
+          ...ordered.slice(0, variant % ordered.length),
+        ]
       : ordered;
     const clips: EditPlan["clips"] = [];
     let duration = 0;
+    const target =
+      params.mode === "sequence"
+        ? MAX_EDIT_SECONDS
+        : params.targetDuration ||
+          Math.min(
+            MAX_EDIT_SECONDS,
+            assets.reduce((sum, a) => sum + a.duration / SECOND, 0),
+          );
+    const visualLength = Math.max(5, Math.min(30, target / 12)) * SECOND;
     for (const candidate of list) {
-      if (duration >= params.targetDuration * SECOND || clips.length >= 12) break;
+      if (duration >= target * SECOND || clips.length >= 12) break;
       const available = candidate.end - candidate.start,
-        remaining = params.targetDuration * SECOND - duration;
+        remaining = target * SECOND - duration;
       const speech = params.mode === "talk" || params.mode === "explain";
       // Speech selections keep their sentence boundary even if it slightly misses requested length.
-      let length = speech ? available : Math.min(available, remaining, 5 * SECOND);
+      let length =
-      if (speech && duration + length > 90 * SECOND) continue;
+        params.mode === "sequence"
-      if (params.beatSync && !speech && beats.length) {
+          ? Math.min(available, remaining)
+          : speech
+            ? available
+            : Math.min(available, remaining, visualLength);
+      if (speech && duration + length > MAX_EDIT_SECONDS * SECOND) continue;
+      if (params.mode !== "sequence" && params.beatSync && !speech && beats.length) {
         const stride = Number(params.beatsPerCut);
         const desired = Math.min(
-          5 * SECOND,
+          visualLength,
           available,
           remaining / Math.max(1, Math.min(12 - clips.length, list.length - clips.length)),
         );
@@ -217,9 +440,12 @@ export function buildEditPlans(
           .filter((_, i) => i % stride === 0)
           .map((t) => t - params.music!.start)
           .filter(
-            (t) => t - duration >= 0.5 * SECOND && t - duration <= Math.min(available, remaining),
+            (t) =>
+              t - duration >= 0.5 * SECOND && t - duration <= Math.min(available, remaining),
           )
-          .sort((a, b) => Math.abs(a - duration - desired) - Math.abs(b - duration - desired))[0];
+          .sort(
+            (a, b) => Math.abs(a - duration - desired) - Math.abs(b - duration - desired),
+          )[0];
         if (next !== undefined) length = next - duration;
       }
       if (length < 200000) continue;
@@ -236,7 +462,7 @@ export function buildEditPlans(
     }
     const plan: EditPlan = {
       version: 1,
-      title: `${{ montage: "素材混剪", talk: "口播视频", explain: "讲解视频", ads: "批量广告" }[params.mode]}${variant ? ` ${variant + 1}` : ""}`,
+      title: `${{ sequence: "智能剪辑", auto: "默认模式", montage: "素材混剪", talk: "口播视频", explain: "讲解视频", ads: "批量广告" }[params.mode]}${variant ? ` ${variant + 1}` : ""}`,
       assets,
       clips,
       ratio: params.ratio,
@@ -247,7 +473,16 @@ export function buildEditPlans(
       warnings: [...warnings],
     };
     const actual = planTimeline(plan).duration / SECOND;
-    if (Math.abs(actual - params.targetDuration) > 0.5)
+    if (
+      params.mode === "sequence" &&
+      assets.reduce((total, a) => total + a.duration, 0) > duration
+    )
+      plan.warnings.push("时间轴已载入前 300 秒，超出的素材未加入；可在设置中另建剪辑方案。");
+    if (
+      params.mode !== "sequence" &&
+      params.targetDuration > 0 &&
+      Math.abs(actual - params.targetDuration) > 0.5
+    )
       plan.warnings.push(
         `目标 ${params.targetDuration} 秒，实际 ${actual.toFixed(2)} 秒；保留完整语句或受素材长度限制。`,
       );

lib/video-canvas/smart-edit/execute.ts

@@ -1,4 +1,5 @@
 import "server-only";
+import { processEditAudio } from "./audio-process";
 import { sql } from "drizzle-orm";
 import { db } from "@/lib/db";
 import { uploadFile } from "@/lib/storage/service";
@@ -31,7 +32,7 @@ import { remoteRender, remoteRenderConfig, type RemoteRenderCheckpoint } from ".
 import { ProviderTaskUncertainError } from "../video-task-lifecycle";
 export type SmartEditCheckpoint = {
   operation: "smart-edit";
-  action: "plan" | "render";
+  action: "plan" | "render" | "separate" | "transcribe";
   phase: "local" | "analyzed" | "selecting" | "selected" | "remote-render";
   remote?: RemoteRenderCheckpoint;
   stage: string;
@@ -67,8 +68,9 @@ export async function executeSmartEdit(
   const upload = async (file: string, suffix: string) => {
     await lifecycle.assertActive();
     const bytes = await readFile(file);
-    return (await uploadFile(`smart-edit-${randomUUID()}.${suffix}`, Uint8Array.from(bytes).buffer))
+    return (
-      .url;
+      await uploadFile(`smart-edit-${randomUUID()}.${suffix}`, Uint8Array.from(bytes).buffer)
+    ).url;
   };
   if (params.operation === "render" && (state.remote || remoteRenderConfig())) {
     const dir = await mkdtemp(path.join(tmpdir(), "remote-render-"));
@@ -78,9 +80,10 @@ export async function executeSmartEdit(
         active: lifecycle.assertActive,
         save: (remote, stage) => save({ phase: "remote-render", remote, stage }),
       });
-      const actual = await probeMedia(rendered.file);
+      const actual = await probeMedia(rendered.file, 300.2);
       if (
-        Math.abs(actual.duration / SECOND - planTimeline(params.plan).duration / SECOND) > 0.15 ||
+        Math.abs(actual.duration / SECOND - planTimeline(params.plan).duration / SECOND) >
+          0.15 ||
         actual.width !== rendered.width ||
         actual.height !== rendered.height
       )
@@ -135,22 +138,33 @@ export async function executeSmartEdit(
             if (!meta.width || !meta.height) throw Error(`${input.name} 不包含视频画面`);
             total += meta.duration;
             if (total > 600 * SECOND) throw Error("素材总时长不能超过 10 分钟");
-            const prepared = await prepareAnalysis(
+            // Opening a timeline only probes source files; it must not run scene analysis or models.
-              media.file,
+            const analysis =
-              media.hash,
+              params.mode === "sequence"
-              media.dir,
+                ? {
-              meta.hasAudio,
+                    scenes: [[0, meta.duration]],
-            );
+                    active: [[0, meta.duration]],
-            const analysis = await pythonAnalysis(
+                    subtitles: input.subtitles,
-              {
+                    warnings: [],
-                version: 1,
+                  }
-                video: prepared.proxy,
+                : await (async () => {
-                audio: prepared.audio,
+                    const prepared = await prepareAnalysis(
-                duration: meta.duration,
+                      media.file,
-                transcribe: params.transcribe && !input.subtitles.length,
+                      media.hash,
-              },
+                      media.dir,
-              media.dir,
+                      meta.hasAudio,
-            );
+                    );
+                    return pythonAnalysis(
+                      {
+                        version: 1,
+                        video: prepared.proxy,
+                        audio: prepared.audio,
+                        duration: meta.duration,
+                        transcribe: params.transcribe && !input.subtitles.length,
+                      },
+                      media.dir,
+                    );
+                  })();
             const subtitles = input.subtitles.length ? input.subtitles : analysis.subtitles;
             if (subtitles.some((s) => s.end > meta.duration + 100000))
               throw Error(`${input.name} 的字幕时间超出视频长度`);
@@ -183,9 +197,14 @@ export async function executeSmartEdit(
                 // Keep full speech sentences; split long visual scenes into bounded candidates.
                 for (let start = range.start; start < range.end; ) {
                   const end =
-                    speech && subtitles.length
+                    params.mode === "sequence" || (speech && subtitles.length)
                       ? range.end
-                      : Math.min(range.end, start + 5 * SECOND);
+                      : Math.min(
+                          range.end,
+                          start +
+                            Math.max(5, Math.min(30, (params.targetDuration || 20) / 12)) *
+                              SECOND,
+                        );
                   if (end - start >= 200000)
                     candidates.push({
                       id: `candidate-${candidates.length}`,
@@ -205,7 +224,9 @@ export async function executeSmartEdit(
           for (let i = 0; fair.length < 24 && grouped.some((g) => i < g.length); i++)
             for (const group of grouped) if (group[i] && fair.length < 24) fair.push(group[i]);
           if (fair.length < candidates.length)
-            warnings.push(`共识别 ${candidates.length} 个候选，本轮按素材轮询选取 24 个进入选片。`);
+            warnings.push(
+              `共识别 ${candidates.length} 个候选，本轮按素材轮询选取 24 个进入选片。`,
+            );
           candidates = fair;
           if (params.beatSync && params.music) {
             await save({ stage: "检测音乐节拍" });
@@ -213,8 +234,14 @@ export async function executeSmartEdit(
               meta = await probeMedia(music.file);
             if (!meta.hasAudio) throw Error("背景音乐没有音轨");
             if (params.music.start >= meta.duration) throw Error("音乐开始时间超出文件");
-            const audio = await prepareAudio(music.file, path.join(music.dir, music.hash + ".wav"));
+            const audio = await prepareAudio(
-            const analysis = await pythonAnalysis({ version: 1, audio, beats: true }, music.dir);
+              music.file,
+              path.join(music.dir, music.hash + ".wav"),
+            );
+            const analysis = await pythonAnalysis(
+              { version: 1, audio, beats: true },
+              music.dir,
+            );
             beats = analysis.beats;
             warnings.push(...analysis.warnings);
           } else if (params.beatSync) warnings.push("未选择音乐，跳过音乐卡点。");
@@ -229,7 +256,7 @@ export async function executeSmartEdit(
         }
         if (!candidates.length) throw Error("没有可用片段，请关闭去静音或更换素材");
         let order = candidates.map((c) => c.id);
-        if (params.semantic) {
+        if (params.semantic && params.mode !== "sequence") {
           const images: string[] = [];
           for (const candidate of candidates) {
             await lifecycle.assertActive();
@@ -271,10 +298,26 @@ export async function executeSmartEdit(
         await save({ phase: "selected", stage: "剪辑方案已生成", plans });
         return { plans, summary: "剪辑方案已生成" };
       }
+      if (params.operation === "separate" || params.operation === "transcribe") {
+        if (state.phase === "selected" && state.plans) return { plans: state.plans };
+        const plan = await processEditAudio(
+          params.operation,
+          params.plan,
+          params.operation === "separate" ? params.clipId : undefined,
+          userId,
+          dir,
+          (stage) => save({ stage }),
+          upload,
+          params.operation === "separate" ? params.track : "video",
+          params.operation === "separate" ? params.keep : "both",
+        );
+        await save({ phase: "selected", stage: "音频处理完成", plans: [plan] });
+        return { plans: [plan] };
+      }
       const plan = editPlanSchema.parse(params.plan);
       planTimeline(plan);
       const rendered = await renderJob(plan, userId, dir, (stage) => save({ stage }));
-      const actual = await probeMedia(rendered.file);
+      const actual = await probeMedia(rendered.file, 300.2);
       if (Math.abs(actual.duration / SECOND - rendered.duration) > 0.15)
         throw Error("导出时长与剪辑方案不一致，未发布结果");
       await save({ stage: "转存成片" });

lib/video-canvas/smart-edit/font-files.ts

@@ -0,0 +1,88 @@
+import path from "node:path";
+import { readFile, writeFile, mkdir, rename, rm } from "node:fs/promises";
+import { createHash, randomUUID } from "node:crypto";
+import { execFile } from "node:child_process";
+import { promisify } from "node:util";
+import { captionFont } from "./fonts";
+const pending = new Map<string, Promise<string>>();
+const sha = (data: Uint8Array) => createHash("sha256").update(data).digest("hex");
+export function captionFontFile(id: string): Promise<string> {
+  const font = captionFont(id);
+  if (!font) return Promise.reject(Error("未知字幕字体"));
+  if (!font.url)
+    return Promise.resolve(
+      path.join(
+        process.env.SMART_EDIT_FONTS_DIR || path.resolve("public/fonts/smart-edit"),
+        font.file,
+      ),
+    );
+  if (pending.has(id)) return pending.get(id)!;
+  const work = (async () => {
+    const root = path.resolve(
+      process.env.SMART_EDIT_FONT_CACHE_DIR || ".cache/smart-edit-fonts",
+    );
+    await mkdir(root, { recursive: true });
+    const output = path.join(root, font.file);
+    try {
+      const data = await readFile(output);
+      if (font.preparedSha256 && sha(data) === font.preparedSha256) return output;
+    } catch {}
+    if (pending.size > 4) throw Error("字体正在加载，请稍后重试");
+    const source = new URL(font.url!);
+    if (
+      source.protocol !== "https:" ||
+      !["fonts.gstatic.com", "fonts.gstatic.cn", "libtv-res.liblib.art"].includes(
+        source.hostname,
+      )
+    )
+      throw Error("字体来源无效");
+    const response = await fetch(source, {
+      signal: AbortSignal.timeout(60000),
+      redirect: "error",
+      cache: "no-store",
+    });
+    if (!response.ok || Number(response.headers.get("content-length")) > 40 * 1024 * 1024)
+      throw Error("字体下载失败");
+    const reader = response.body!.getReader(),
+      chunks: Uint8Array[] = [];
+    let bytes = 0;
+    try {
+      while (true) {
+        const { done, value } = await reader.read();
+        if (done) break;
+        bytes += value.length;
+        if (bytes > 40 * 1024 * 1024) throw Error("字体文件过大");
+        chunks.push(value);
+      }
+    } finally {
+      await reader.cancel();
+    }
+    const data = Buffer.concat(chunks);
+    if (!font.sha256 || sha(data) !== font.sha256)
+      throw Error("字体资源校验失败，请更新字体目录");
+    const temp = path.join(root, randomUUID()),
+      prepared = temp + ".ttf";
+    try {
+      await writeFile(temp, data);
+      await promisify(execFile)(
+        process.env.SMART_EDIT_FONT_PYTHON || process.env.SMART_EDIT_PYTHON || "python3",
+        [
+          process.env.SMART_EDIT_FONT_SCRIPT || path.resolve("services/smart-edit/font.py"),
+          temp,
+          prepared,
+          font.family,
+        ],
+        { timeout: 120000, maxBuffer: 512 * 1024 },
+      );
+      if (sha(await readFile(prepared)) !== font.preparedSha256)
+        throw Error("字体处理版本不一致，请安装指定字体依赖");
+      await rename(prepared, output);
+      return output;
+    } finally {
+      await rm(temp, { force: true });
+      await rm(prepared, { force: true });
+    }
+  })().finally(() => pending.delete(id));
+  pending.set(id, work);
+  return work;
+}

lib/video-canvas/smart-edit/fonts.ts

@@ -0,0 +1,64 @@
+import catalog from "./font-catalog.json";
+export type CaptionFont = {
+  id: string;
+  name: string;
+  family: string;
+  file: string;
+  common: boolean;
+  recommended: boolean;
+  google: boolean;
+  url?: string;
+  sha256?: string;
+  preparedSha256?: string;
+};
+/** Same packaged font bytes are used in browser previews and FFmpeg exports. */
+const bundledFonts: CaptionFont[] = [
+  {
+    id: "source-han",
+    name: "Source Han Sans CN",
+    family: "Source Han Sans SC",
+    file: "source-han.otf",
+    common: true,
+    recommended: true,
+    google: false,
+  },
+  {
+    id: "noto-sans-sc",
+    name: "Noto Sans SC",
+    family: "Noto Sans SC",
+    file: "noto-sans-sc.ttf",
+    common: true,
+    recommended: true,
+    google: true,
+  },
+  {
+    id: "noto-serif-sc",
+    name: "Noto Serif SC",
+    family: "Noto Serif SC",
+    file: "noto-serif-sc.ttf",
+    common: true,
+    recommended: true,
+    google: true,
+  },
+  {
+    id: "roboto",
+    name: "Roboto",
+    family: "Roboto",
+    file: "roboto.ttf",
+    common: true,
+    recommended: false,
+    google: true,
+  },
+  {
+    id: "inter",
+    name: "Inter",
+    family: "Inter",
+    file: "inter.ttf",
+    common: true,
+    recommended: false,
+    google: true,
+  },
+];
+export const captionFonts: CaptionFont[] = [...bundledFonts, ...catalog];
+const byId = new Map(captionFonts.map((font) => [font.id, font]));
+export const captionFont = (id?: string) => byId.get(id || "");

lib/video-canvas/smart-edit/intent.ts

@@ -0,0 +1,20 @@
+/** Explicit switches take precedence over the common beat-sync phrase shortcut. */
+export function resolveSmartEditIntent<
+  T extends {
+    mode?: string;
+    prompt?: string;
+    beatSync?: boolean;
+    variants?: number;
+    music?: { url: string };
+  },
+>(draft: T) {
+  const prompt = draft.prompt || "";
+  const beatSync =
+    draft.beatSync ??
+    (/卡点|beat[ -]?sync/i.test(prompt) &&
+      !/(?:不|不要|无需|不用|取消|关闭).{0,5}卡点|no beat[ -]?sync/i.test(prompt));
+  return { ...draft, beatSync, variants: draft.variants ?? (draft.mode === "ads" ? 3 : 1) };
+}
+export function assertBeatMusic(draft: { beatSync?: boolean; music?: { url: string } }) {
+  if (draft.beatSync && !draft.music?.url) throw Error("音乐卡点需要先选择或上传背景音乐");
+}

lib/video-canvas/smart-edit/local.ts

@@ -12,7 +12,7 @@ export const mediaSafety = [
   "-format_whitelist",
   "mov,matroska,webm,avi,mpeg,mpegts,flv,ogg,wav,mp3,flac,aac",
 ];
-export async function ffmpeg(args: string[], cwd?: string) {
+export async function ffmpeg(args: string[], cwd?: string, timeout = 240000) {
   await exec(
     "ffmpeg",
     [
@@ -28,10 +28,10 @@ export async function ffmpeg(args: string[], cwd?: string) {
       "1",
       ...args,
     ],
-    { cwd, timeout: 240000, maxBuffer: 2 * 1024 * 1024 },
+    { cwd, timeout, maxBuffer: 2 * 1024 * 1024 },
   );
 }
-export async function probeMedia(file: string) {
+export async function probeMedia(file: string, maxSeconds = 180) {
   const { stdout } = await exec(
     "ffprobe",
     ["-v", "error", ...mediaSafety, "-show_streams", "-show_format", "-of", "json", file],
@@ -40,8 +40,8 @@ export async function probeMedia(file: string) {
   const info = JSON.parse(stdout),
     v = info.streams?.find((s: any) => s.codec_type === "video");
   const duration = Math.round(Number(info.format?.duration || v?.duration) * SECOND);
-  if (!Number.isFinite(duration) || duration <= 0 || duration > 180 * SECOND)
+  if (!Number.isFinite(duration) || duration <= 0 || duration > maxSeconds * SECOND)
-    throw Error("单个素材需在 0–180 秒以内");
+    throw Error(`媒体时长需在 0–${maxSeconds} 秒以内`);
   if (v && (!v.width || !v.height || v.width * v.height > 33554432))
     throw Error("素材画面尺寸超过处理上限");
   return {
@@ -114,7 +114,10 @@ export async function pythonAnalysis(
   data: Record<string, unknown>,
   dir: string,
 ): Promise<LocalAnalysis> {
-  const key = createHash("sha256").update(JSON.stringify(data)).digest("hex"),
+  const key = createHash("sha256")
+      .update("analysis-v3:")
+      .update(JSON.stringify(data))
+      .digest("hex"),
     input = path.join(dir, key + ".analysis-input.json"),
     output = path.join(dir, key + ".analysis.json");
   try {
@@ -147,7 +150,12 @@ export async function pythonAnalysis(
     .slice(0, 500);
   return value;
 }
-export async function prepareAnalysis(file: string, hash: string, dir: string, hasAudio: boolean) {
+export async function prepareAnalysis(
+  file: string,
+  hash: string,
+  dir: string,
+  hasAudio: boolean,
+) {
   const proxy = path.join(dir, hash + ".proxy.mp4"),
     audio = path.join(dir, hash + ".wav");
   try {

lib/video-canvas/smart-edit/remote-render.ts

@@ -4,7 +4,7 @@ import { Readable, Transform } from "node:stream";
 import { pipeline } from "node:stream/promises";
 import path from "node:path";
 import { z } from "zod";
-import type { EditPlan } from "./contract";
+import { MAX_EDIT_SECONDS, type EditPlan } from "./contract";
 import { ProviderTaskUncertainError } from "../video-task-lifecycle";
 
 export type RemoteRenderCheckpoint = { id: string; endpointHash: string };
@@ -33,9 +33,12 @@ const statusSchema = z.object({
   error: z.string().max(500).optional(),
   result: z
     .object({
-      width: z.number().int().positive().max(1920),
+      width: z.number().int().positive().max(2520),
-      height: z.number().int().positive().max(1920),
+      height: z.number().int().positive().max(2520),
-      durationSeconds: z.number().positive().max(90.2),
+      durationSeconds: z
+        .number()
+        .positive()
+        .max(MAX_EDIT_SECONDS + 0.2),
     })
     .optional(),
 });

lib/video-canvas/smart-edit/render-job.ts

@@ -11,8 +11,8 @@ export async function renderJob(
 ) {
   const plan = editPlanSchema.parse(value);
   planTimeline(plan);
-  if (plan.assets.reduce((n, a) => n + a.duration, 0) > 600 * SECOND)
+  if (plan.assets.reduce((n, a) => n + a.duration, 0) > 1800 * SECOND)
-    throw Error("素材总时长不能超过 10 分钟");
+    throw Error("素材与分离音轨总时长不能超过 30 分钟");
   const files: Record<string, string> = {};
   for (const asset of plan.assets) {
     await progress("读取原始素材");
@@ -32,11 +32,12 @@ export async function renderJob(
   if (plan.music) {
     const media = await materialize(plan.music.url, userId),
       meta = await probeMedia(media.file);
-    if (!meta.hasAudio || plan.music.start >= meta.duration) throw Error("背景音乐或起始时间无效");
+    if (!meta.hasAudio || plan.music.start >= meta.duration)
+      throw Error("背景音乐或起始时间无效");
     musicFile = media.file;
   }
   const rendered = await renderEditPlan(plan, files, musicFile, dir, progress);
-  const actual = await probeMedia(rendered.file);
+  const actual = await probeMedia(rendered.file, 300.2);
   if (Math.abs(actual.duration / SECOND - rendered.duration) > 0.15)
     throw Error("导出时长与剪辑方案不一致，未发布结果");
   return { ...rendered, duration: actual.duration / SECOND };

lib/video-canvas/smart-edit/render.ts

@@ -1,7 +1,19 @@
-import { readFile, writeFile } from "node:fs/promises";
+import { readFile, writeFile, mkdir, copyFile } from "node:fs/promises";
 import path from "node:path";
-import { type EditPlan, planTimeline, compiledSubtitles, subtitlesToSrt, SECOND } from "./contract";
+import { captionFontFile } from "./font-files";
+import { captionFont } from "./fonts";
+import {
+  type EditPlan,
+  editOutputSize,
+  planTimeline,
+  compiledSubtitles,
+  subtitlesToSrt,
+  subtitlesToAss,
+  SECOND,
+  defaultVideoTransform,
+} from "./contract";
 import { ffmpeg, mediaSafety } from "./local";
+const renderFfmpeg = (args: string[], cwd?: string) => ffmpeg(args, cwd, 600000);
 export async function renderEditPlan(
   plan: EditPlan,
   files: Record<string, string>,
@@ -11,13 +23,36 @@ export async function renderEditPlan(
 ) {
   const timeline = planTimeline(plan),
     duration = timeline.duration / SECOND;
-  const size = plan.resolution === "1080p" ? 1080 : 720;
+  const { width, height } = editOutputSize(plan);
-  const [width, height] =
+  const selectedFonts = [
-    plan.ratio === "16:9"
+    ...new Set(
-      ? [Math.round((size * 16) / 9), size]
+      compiledSubtitles(plan)
-      : plan.ratio === "9:16"
+        .map(
-        ? [size, Math.round((size * 16) / 9)]
+          (caption) =>
-        : [size, size];
+            captionFont((caption as { style?: { fontFamily?: string } }).style?.fontFamily)?.id,
+        )
+        .filter((file): file is NonNullable<typeof file> => Boolean(file)),
+    ),
+  ];
+  if (selectedFonts.length) {
+    const fontDir = path.join(dir, "fonts");
+    await mkdir(fontDir, { recursive: true });
+    const source =
+      process.env.SMART_EDIT_FONTS_DIR ||
+      path.resolve(process.cwd(), "public/fonts/smart-edit");
+    for (const id of selectedFonts) {
+      const file = captionFont(id)!.file;
+      try {
+        await progress("准备字幕字体");
+        await copyFile(await captionFontFile(id), path.join(fontDir, file));
+      } catch {
+        throw Error(`字幕字体文件缺失：${file}，请安装剪辑字体包后重新导出`);
+      }
+    }
+    // Latin faces use the same packaged Chinese fallback in both renderers.
+    if (!selectedFonts.includes("source-han"))
+      await copyFile(path.join(source, "source-han.otf"), path.join(fontDir, "source-han.otf"));
+  }
   const normalized: string[] = [];
   for (let i = 0; i < timeline.clips.length; i++) {
     await progress(`处理片段 ${i + 1}/${timeline.clips.length}`);
@@ -35,24 +70,43 @@ export async function renderEditPlan(
       files[clip.assetId],
     ];
     if (!asset.hasAudio) inputs.push("-f", "lavfi", "-i", "anullsrc=r=48000:cl=stereo");
-    await ffmpeg([
+    const t = clip.transform || defaultVideoTransform;
+    const transformWidth = Math.max(2, Math.round((width * t.scale) / 2) * 2),
+      transformHeight = Math.max(2, Math.round((height * t.scale) / 2) * 2);
+    const transformX = Math.round((width - transformWidth) / 2 + t.x * width),
+      transformY = Math.round((height - transformHeight) / 2 + t.y * height);
+    const pad =
+      clip.padBefore || clip.padAfter
+        ? `,tpad=start_mode=clone:start_duration=${clip.padBefore / SECOND}:stop_mode=clone:stop_duration=${clip.padAfter / SECOND}`
+        : "";
+    const normalize = `setpts=(PTS-STARTPTS)/${clip.speed},scale=${width}:${height}:force_original_aspect_ratio=decrease,pad=${width}:${height}:(ow-iw)/2:(oh-ih)/2,setsar=1,fps=30,trim=duration=${clip.duration / SECOND},format=yuv420p${pad}`;
+    const transformed = t.scale !== 1 || t.x !== 0 || t.y !== 0;
+    await renderFfmpeg([
       ...inputs,
       "-map",
-      "0:v:0",
+      transformed ? "[v]" : "0:v:0",
       "-map",
       asset.hasAudio ? "0:a:0" : "1:a:0",
-      "-vf",
+      ...(transformed
-      `setpts=(PTS-STARTPTS)/${clip.speed},scale=${width}:${height}:force_original_aspect_ratio=decrease,pad=${width}:${height}:(ow-iw)/2:(oh-ih)/2,setsar=1,fps=30,format=yuv420p`,
+        ? [
+            "-filter_complex",
+            `[0:v]${normalize},scale=${transformWidth}:${transformHeight}[fg];color=c=black:s=${width}x${height}:r=30[bg];[bg][fg]overlay=x=${transformX}:y=${transformY}:shortest=1[v]`,
+          ]
+        : ["-vf", normalize]),
       "-af",
-      `asetpts=PTS-STARTPTS,atempo=${clip.speed},aresample=48000,aformat=channel_layouts=stereo,volume=${clip.gain},apad`,
+      `asetpts=PTS-STARTPTS,atempo=${clip.speed},aresample=48000,aformat=channel_layouts=stereo,volume=${clip.audioDetached || clip.muted ? 0 : clip.gain},atrim=duration=${clip.duration / SECOND}${clip.padBefore ? `,adelay=${Math.round(clip.padBefore / 1000)}:all=1` : ""},apad`,
       "-t",
-      String(clip.duration / SECOND),
+      String((clip.duration + clip.padBefore + clip.padAfter) / SECOND),
       "-c:v",
       "libx264",
       "-preset",
       "veryfast",
       "-crf",
       "21",
+      "-maxrate",
+      "6M",
+      "-bufsize",
+      "12M",
       "-threads",
       "2",
       "-c:a",
@@ -67,78 +121,192 @@ export async function renderEditPlan(
   }
   await progress("合成时间轴");
   const joined = path.join(dir, "joined.mp4");
-  if (plan.transition === 0 || normalized.length === 1) {
+  if (!timeline.clips.some((clip) => clip.overlap) || normalized.length === 1) {
     await writeFile(
       path.join(dir, "concat.txt"),
       normalized.map((_, i) => `file 'clip-${i}.mp4'`).join("\n"),
     );
-    await ffmpeg(["-f", "concat", "-safe", "1", "-i", "concat.txt", "-c", "copy", joined], dir);
+    await renderFfmpeg(
+      ["-f", "concat", "-safe", "1", "-i", "concat.txt", "-c", "copy", joined],
+      dir,
+    );
   } else {
     // Normalize first: all xfade inputs have identical pixel format, timebase and dimensions.
-    const filters: string[] = [];
+    const filters: string[] = normalized.map(
-    let video = "0:v",
+      (_, i) => `[${i}:v]format=yuv444p,settb=AVTB,setpts=PTS-STARTPTS[vin${i}]`,
-      audio = "0:a";
+    );
+    const audioFilters = timeline.clips.map(
+      (c, i) =>
+        `[${i}:a]atrim=duration=${(c.duration + c.padBefore + c.padAfter) / SECOND},asetpts=PTS-STARTPTS[ain${i}]`,
+    );
+    let video = "vin0",
+      audio = "ain0";
     for (let i = 1; i < normalized.length; i++) {
       const clip = timeline.clips[i];
-      filters.push(
+      if (clip.overlap) {
-        `[${video}][${i}:v]xfade=transition=fade:duration=${clip.overlap / SECOND}:offset=${clip.start / SECOND}[v${i}]`,
+        const length = clip.overlap / SECOND,
-        `[${audio}][${i}:a]acrossfade=d=${clip.overlap / SECOND}:c1=tri:c2=tri[a${i}]`,
+          offset = clip.transitionStart / SECOND;
-      );
+        if (clip.transitionType === "fade") {
+          filters.push(
+            `[${video}][vin${i}]xfade=transition=fade:duration=${length}:offset=${offset}[v${i}]`,
+          );
+        } else {
+          // Two native fades meet at a solid-color midpoint, matching the browser overlay.
+          // Avoid custom per-pixel expressions: they are prohibitively slow at 1080P.
+          const half = length / 2,
+            color = clip.transitionType === "fadeblack" ? "black" : "white";
+          filters.push(
+            `color=c=${color}:s=${width}x${height}:r=30:d=${length},format=yuv444p,settb=AVTB[solid${i}]`,
+            `[vin${i}]trim=start=${half},setpts=PTS-STARTPTS[tail${i}]`,
+            `[${video}][solid${i}]xfade=transition=fade:duration=${half}:offset=${offset}[dip${i}]`,
+            `[dip${i}][tail${i}]xfade=transition=fade:duration=${half}:offset=${offset + half}[v${i}]`,
+          );
+        }
+        audioFilters.push(`[${audio}][ain${i}]acrossfade=d=${length}:c1=tri:c2=tri[a${i}]`);
+      } else {
+        filters.push(`[${video}][vin${i}]concat=n=2:v=1:a=0[v${i}]`);
+        audioFilters.push(`[${audio}][ain${i}]concat=n=2:v=0:a=1[a${i}]`);
+      }
       video = `v${i}`;
       audio = `a${i}`;
     }
-    await ffmpeg([
+    const videoFile = path.join(dir, "joined-video.mp4"),
-      ...normalized.flatMap((file) => ["-i", file]),
+      audioFile = path.join(dir, "joined-audio.m4a");
+    await renderFfmpeg([
+      ...normalized.flatMap((file) => ["-threads", "2", "-i", file]),
       "-filter_complex",
       filters.join(";"),
       "-map",
       `[${video}]`,
-      "-map",
+      "-an",
-      `[${audio}]`,
-      "-t",
-      String(duration),
       "-c:v",
       "libx264",
       "-preset",
       "veryfast",
       "-crf",
       "21",
+      "-maxrate",
+      "6M",
+      "-bufsize",
+      "12M",
       "-threads",
       "2",
+      "-pix_fmt",
+      "yuv420p",
+      videoFile,
+    ]);
+    // Keep framesync video and full-length audio fades in separate passes. Otherwise
+    // FFmpeg 4 can finish the audio graph before emitting buffered video at offset 0.
+    await renderFfmpeg([
+      ...normalized.flatMap((file) => ["-i", file]),
+      "-filter_complex",
+      audioFilters.join(";"),
+      "-map",
+      `[${audio}]`,
+      "-vn",
       "-c:a",
       "aac",
+      audioFile,
+    ]);
+    await renderFfmpeg([
+      "-i",
+      videoFile,
+      "-i",
+      audioFile,
+      "-map",
+      "0:v:0",
+      "-map",
+      "1:a:0",
+      "-c",
+      "copy",
       joined,
     ]);
   }
   const subtitles = compiledSubtitles(plan);
   const srt = path.join(dir, "subtitles.srt");
   await writeFile(srt, subtitlesToSrt(subtitles));
+  await writeFile(path.join(dir, "subtitles.ass"), subtitlesToAss(plan, width, height));
   const output = path.join(dir, "result.mp4");
   const inputs = ["-i", joined];
   const filters: string[] = [];
-  if (subtitles.length)
+  const videoTail = Math.max(
-    filters.push(
+    0,
-      "[0:v]subtitles=filename=subtitles.srt:force_style='FontSize=22,Outline=2,MarginV=30'[v]",
+    duration - timeline.clips[timeline.clips.length - 1].end / SECOND,
-    );
+  );
+  const videoFilters = [
+    ...(videoTail ? [`tpad=stop_mode=add:stop_duration=${videoTail}`] : []),
+    ...(subtitles.length
+      ? [
+          selectedFonts.length
+            ? "ass=filename=subtitles.ass:fontsdir=fonts"
+            : "ass=filename=subtitles.ass",
+        ]
+      : []),
+  ];
+  if (videoFilters.length) filters.push(`[0:v]${videoFilters.join(",")}[v]`);
+  filters.push(`[0:a]apad,atrim=duration=${duration}[baseaudio]`);
+  const mixLabels = ["[baseaudio]"];
+  let audioInput = 1;
   if (musicFile && plan.music) {
     inputs.push("-ss", String(plan.music.start / SECOND), "-i", musicFile);
     filters.push(
-      `[1:a]asetpts=PTS-STARTPTS,aresample=48000,aformat=channel_layouts=stereo,volume=${plan.music.gain},apad,atrim=duration=${duration}[music]`,
+      `[${audioInput}:a]asetpts=PTS-STARTPTS,aresample=48000,aformat=channel_layouts=stereo,volume=${plan.music.gain},apad,atrim=duration=${duration}[music]`,
-      `[0:a][music]amix=inputs=2:duration=first:normalize=0,alimiter=limit=0.95:level=0[a]`,
     );
+    mixLabels.push("[music]");
+    audioInput++;
   }
+  for (const [index, clip] of timeline.audioClips.entries()) {
+    inputs.push(
+      ...mediaSafety,
+      "-ss",
+      String(clip.sourceIn / SECOND),
+      "-t",
+      String((clip.sourceOut - clip.sourceIn) / SECOND),
+      "-i",
+      files[clip.assetId],
+    );
+    const length = clip.duration / SECOND,
+      fadeIn = Math.min(clip.fadeIn, length),
+      fadeOut = Math.min(clip.fadeOut, length);
+    const offsetEnvelope =
+      clip.fadeInStart || clip.fadeOutEnd
+        ? `,volume='${fadeIn ? `min(1,${clip.fadeInStart || 0}+(1-${clip.fadeInStart || 0})*t/${fadeIn})` : "1"}*${fadeOut ? `min(1,${clip.fadeOutEnd || 0}+(1-${clip.fadeOutEnd || 0})*(${length}-t)/${fadeOut})` : "1"}':eval=frame`
+        : `${fadeIn ? `,afade=t=in:d=${fadeIn}` : ""}${fadeOut ? `,afade=t=out:st=${length - fadeOut}:d=${fadeOut}` : ""}`;
+    filters.push(
+      `[${audioInput}:a]asetpts=PTS-STARTPTS,atempo=${clip.speed},aresample=48000,aformat=channel_layouts=stereo,volume=${clip.muted ? 0 : clip.gain},atrim=duration=${length}${offsetEnvelope},adelay=${Math.round(clip.start / 1000)}:all=1[track${index}]`,
+    );
+    mixLabels.push(`[track${index}]`);
+    audioInput++;
+  }
+  if (mixLabels.length > 1)
+    filters.push(
+      `${mixLabels.join("")}amix=inputs=${mixLabels.length}:duration=first:normalize=0,alimiter=limit=0.95:level=0[a]`,
+    );
   await progress("输出成片");
-  await ffmpeg(
+  await renderFfmpeg(
     [
       ...inputs,
       ...(filters.length ? ["-filter_complex", filters.join(";")] : []),
       "-map",
-      subtitles.length ? "[v]" : "0:v:0",
+      videoFilters.length ? "[v]" : "0:v:0",
       "-map",
-      musicFile && plan.music ? "[a]" : "0:a:0",
+      mixLabels.length > 1 ? "[a]" : "[baseaudio]",
       "-c:v",
-      subtitles.length ? "libx264" : "copy",
+      videoFilters.length ? "libx264" : "copy",
-      ...(subtitles.length ? ["-preset", "veryfast", "-crf", "21", "-threads", "2"] : []),
+      ...(videoFilters.length
+        ? [
+            "-preset",
+            "veryfast",
+            "-crf",
+            "21",
+            "-maxrate",
+            "6M",
+            "-bufsize",
+            "12M",
+            "-threads",
+            "2",
+          ]
+        : []),
       "-c:a",
       "aac",
       "-t",

lib/video-canvas/task-contract.ts

@@ -1,12 +1,23 @@
 import { hasRecoverableCanvasTask } from "./task-recovery";
 import { normalizeVideoAnalysisRows } from "./video-story";
-import { storyTaskAttachmentSchema, mergeStoryTaskResult } from "./story-task-attachment";
+import {
+  storyTaskAttachmentSchema,
+  mergeStoryTaskResult,
+} from "./story-task-attachment";
 import { z } from "zod";
 
 export const canvasTaskRequestSchema = z.object({
   requestId: z.string().uuid(),
   projectId: z.string().uuid(),
-  kind: z.enum(["video", "image", "analysis", "image-operation", "audio", "smart-edit"]),
+  kind: z.enum([
+    "video",
+    "image",
+    "analysis",
+    "image-operation",
+    "audio",
+    "smart-edit",
+    "video-upscale",
+  ]),
   node: z
     .object({
       id: z.string().min(1).max(180),
@@ -21,6 +32,27 @@ export const canvasTaskRequestSchema = z.object({
     })
     .passthrough(),
   params: z.record(z.unknown()),
+  workflow: z
+    .object({
+      runId: z.string().uuid(),
+      groupId: z.string().min(1).max(180),
+      inputs: z
+        .array(
+          z.object({
+            output: z.literal("edit-plan").optional(),
+            requestId: z.string().uuid().optional(),
+            node: z.object({
+              id: z.string().min(1).max(180),
+              title: z.string().optional(),
+              kind: z.string().optional(),
+              inputType: z.string().optional(),
+              draftValues: z.record(z.any()).optional(),
+            }),
+          })
+        )
+        .max(30),
+    })
+    .optional(),
   attachment: storyTaskAttachmentSchema.optional(),
   hiddenNodeIds: z.array(z.string()).max(20).optional(),
   edges: z
@@ -47,9 +79,15 @@ export function canCancelCanvasTask(status: string) {
   return status === "pending";
 }
 
-export function requireAnalysisResult(data: Record<string, any> | undefined, mode: string) {
+export function requireAnalysisResult(
+  data: Record<string, any> | undefined,
+  mode: string
+) {
   if (!data) throw Error("分析没有返回结果，原内容已保留");
-  if (mode === "video-analysis" && !normalizeVideoAnalysisRows(data.analysisRows).length)
+  if (
+    mode === "video-analysis" &&
+    !normalizeVideoAnalysisRows(data.analysisRows).length
+  )
     throw Error("视频解析没有返回有效镜头，原内容已保留");
   if (
     mode === "director-reconstruct" &&
@@ -66,8 +104,8 @@ export function requireAnalysisResult(data: Record<string, any> | undefined, mod
       !rows.length ||
       !rows.some((row) =>
         [row.sceneDescription, row.description, row.imagePrompt].some(
-          (value) => typeof value === "string" && value.trim(),
+          (value) => typeof value === "string" && value.trim()
-        ),
+        )
       ))
   ) {
     throw Error("脚本没有返回有效镜头，原内容已保留");
@@ -76,7 +114,7 @@ export function requireAnalysisResult(data: Record<string, any> | undefined, mod
     mode !== "story-script" &&
     mode !== "video-analysis" &&
     ![data.raw, data.summary, ...(data.agentPlan || [])].some(
-      (value) => typeof value === "string" && value.trim(),
+      (value) => typeof value === "string" && value.trim()
     )
   ) {
     throw Error("分析返回了空内容，原内容已保留");
@@ -87,13 +125,20 @@ export function requireAnalysisResult(data: Record<string, any> | undefined, mod
 // Stable result IDs and request IDs prevent repeat polling from duplicating images
 // or an old task from overwriting a later generation on the same node.
 export function reconcileCanvasTasks<
-  T extends { id: string; draftValues?: Record<string, any>; [key: string]: any },
+  T extends {
+    id: string;
+    draftValues?: Record<string, any>;
+    [key: string]: any;
+  },
 >(nodes: T[], tasks: CanvasGenerationTask[], projectRevision: string) {
   let result = nodes;
   const latestByNode = new Map<string, CanvasGenerationTask>();
   for (const task of tasks) {
     const previous = latestByNode.get(task.request.node.id);
-    if (!previous || Date.parse(task.createdAt) >= Date.parse(previous.createdAt))
+    if (
+      !previous ||
+      Date.parse(task.createdAt) >= Date.parse(previous.createdAt)
+    )
       latestByNode.set(task.request.node.id, task);
   }
   for (const task of latestByNode.values()) {
@@ -109,7 +154,9 @@ export function reconcileCanvasTasks<
     const node = result[index];
     const draft = node.draftValues || {};
     if (draft.taskRequestId && draft.taskRequestId !== task.requestId) {
-      const savedTask = tasks.find((item) => item.requestId === draft.taskRequestId);
+      const savedTask = tasks.find(
+        (item) => item.requestId === draft.taskRequestId
+      );
       if (
         !savedTask ||
         Date.parse(task.createdAt) <= Date.parse(savedTask.createdAt) ||
@@ -117,7 +164,7 @@ export function reconcileCanvasTasks<
       )
         continue;
     }
-    const appliedStatus = `${task.id}:${task.status}${task.request.kind === "smart-edit" ? `:${task.result?.smartEdit?.stage || ""}` : ""}`;
+    const appliedStatus = `${task.id}:${task.status}${["smart-edit", "video-upscale"].includes(task.request.kind) ? `:${task.result?.upscale?.stage || task.result?.smartEdit?.stage || ""}` : ""}`;
     if (draft.appliedTaskStatus === appliedStatus) continue;
     const nextDraft = {
       ...draft,
@@ -128,7 +175,11 @@ export function reconcileCanvasTasks<
       generationProgress: 0,
       generationStatus: "pending",
       pendingLabel:
-        task.status === "pending" ? "排队中" : task.result?.smartEdit?.stage || "生成中",
+        task.status === "pending"
+          ? "排队中"
+          : task.result?.upscale?.stage ||
+            task.result?.smartEdit?.stage ||
+            "生成中",
     };
     let subtitle = nextDraft.pendingLabel;
     if (task.status === "succeeded") {
@@ -141,13 +192,33 @@ export function reconcileCanvasTasks<
       });
       subtitle = "生成结果";
       if (task.request.kind === "smart-edit") {
-        if (task.request.params.operation === "plan")
+        if (task.request.params.operation !== "render") {
-          Object.assign(nextDraft, { smartEditPlans: data.plans, smartEditPlanTaskId: task.id });
+          const saved = draft.smartEditDraft;
-        else
+          const processing = ["separate", "transcribe"].includes(
+            String(task.request.params.operation)
+          );
+          const plans =
+            processing && saved?.plans?.length
+              ? saved.plans.map((plan: unknown, i: number) =>
+                  i === (saved.selectedPlanIndex || 0)
+                    ? data.plans?.[0] || plan
+                    : plan
+                )
+              : data.plans;
+          Object.assign(nextDraft, {
+            smartEditPlans: plans,
+            smartEditPlanTaskId: task.id,
+          });
+          if (processing && saved)
+            Object.assign(nextDraft, {
+              smartEditDraft: { ...saved, plans, planTaskId: task.id },
+            });
+        } else
           Object.assign(nextDraft, {
             mediaUrl: data.url,
             url: [data.url],
             videoMeta: data.metadata,
+            durationSeconds: data.metadata?.durationSeconds,
             subtitleUrl: data.subtitleUrl,
             editPlan: data.plan,
             action: "VIDEO_RESOURCE",
@@ -161,18 +232,29 @@ export function reconcileCanvasTasks<
           action: "AUDIO_RESOURCE",
           mediaType: "audio",
           audioMeta: data.metadata,
-          audioParams: task.request.params,
+          audioParams:
+            task.request.node.draftValues?.audioParams || task.request.params,
         });
-      if (task.request.kind === "video")
+      if (["video", "video-upscale"].includes(task.request.kind))
-        Object.assign(nextDraft, { mediaUrl: data.url, url: [data.url], videoMeta: data.metadata });
+        Object.assign(nextDraft, {
-      if (task.request.kind === "image" || task.request.kind === "image-operation") {
+          mediaUrl: data.url,
+          url: [data.url],
+          videoMeta: data.metadata,
+        });
+      if (
+        task.request.kind === "image" ||
+        task.request.kind === "image-operation"
+      ) {
         const urls = (data.images || [])
           .map((image: any) => image.originalUrl || image.thumbnailUrl)
           .filter(Boolean);
         Object.assign(nextDraft, {
           mediaUrl: urls[0],
           url: urls,
-          action: draft.action === "IMAGE_GENERATE_NINE_GRID" ? draft.action : "IMAGE_RESOURCE",
+          action:
+            draft.action === "IMAGE_GENERATE_NINE_GRID"
+              ? draft.action
+              : "IMAGE_RESOURCE",
         });
         for (let i = 1; i < urls.length; i++) {
           const id = `${source.id}-result-${task.id}-${i}`;
@@ -184,19 +266,28 @@ export function reconcileCanvasTasks<
                 id,
                 title: `${node.title} ${i + 1}`,
                 y: node.y + i * (node.height + 40),
-                draftValues: { ...nextDraft, mediaUrl: urls[i], url: [urls[i]] },
+                draftValues: {
+                  ...nextDraft,
+                  mediaUrl: urls[i],
+                  url: [urls[i]],
+                },
               } as T,
             ];
         }
       }
-      if (task.request.kind === "analysis" && task.request.params.mode === "director-reconstruct")
+      if (
+        task.request.kind === "analysis" &&
+        task.request.params.mode === "director-reconstruct"
+      )
         Object.assign(nextDraft, {
           directorScene: data.directorScene,
           reconstructionSummary: data.summary,
         });
       if (task.request.kind === "analysis")
         Object.assign(nextDraft, {
-          content: [data.raw || data.summary || data.agentPlan?.join("\n") || ""],
+          content: [
+            data.raw || data.summary || data.agentPlan?.join("\n") || "",
+          ],
           ...(Array.isArray(data.analysisRows || data.shotIdeas)
             ? { rows: data.analysisRows || data.shotIdeas }
             : {}),
@@ -208,10 +299,13 @@ export function reconcileCanvasTasks<
         pendingLabel: "",
         generationError:
           task.error ||
-          (task.status === "cancelled" ? "已取消排队，未提交供应商" : "任务失败，请查看任务状态"),
+          (task.status === "cancelled"
+            ? "已取消排队，未提交供应商"
+            : "任务失败，请查看任务状态"),
       });
       if (task.status === "interrupted" && hasRecoverableCanvasTask(task))
-        nextDraft.generationError = "任务查询中断，稍后会自动查询原任务，请勿重复生成。";
+        nextDraft.generationError =
+          "任务查询中断，稍后会自动查询原任务，请勿重复生成。";
       subtitle =
         task.status === "cancelled"
           ? "已取消"
@@ -221,7 +315,10 @@ export function reconcileCanvasTasks<
               : "待核实"
             : "生成失败";
     }
-    if (task.request.kind === "analysis" && task.request.params.mode === "video-analysis") {
+    if (
+      task.request.kind === "analysis" &&
+      task.request.params.mode === "video-analysis"
+    ) {
       Object.assign(nextDraft, {
         status:
           task.status === "succeeded"
@@ -236,7 +333,7 @@ export function reconcileCanvasTasks<
           rows: normalizeVideoAnalysisRows(
             task.result?.analysisRows,
             task.result?.framePreviews || [],
-            source.id,
+            source.id
           ),
           framePreviews: task.result?.framePreviews || [],
           summary: task.result?.summary || "",
@@ -249,11 +346,15 @@ export function reconcileCanvasTasks<
             ...item,
             ...(task.status === "succeeded" &&
             task.request.kind !== "analysis" &&
-            !(task.request.kind === "smart-edit" && task.request.params.operation === "plan")
+            !(
+              task.request.kind === "smart-edit" &&
+              task.request.params.operation !== "render"
+            )
               ? {
                   kind: "media",
                   inputType:
-                    task.request.kind === "video" || task.request.kind === "smart-edit"
+                    ["video", "video-upscale"].includes(task.request.kind) ||
+                    task.request.kind === "smart-edit"
                       ? "video"
                       : task.request.kind === "audio"
                         ? "audio"
@@ -263,7 +364,7 @@ export function reconcileCanvasTasks<
             subtitle,
             draftValues: nextDraft,
           }
-        : item,
+        : item
     );
   }
   for (const task of tasks) {
@@ -275,14 +376,21 @@ export function reconcileCanvasTasks<
       !result.some((node) => node.id === task.request.node.id)
     )
       continue;
-    const targetIndex = result.findIndex((node) => node.id === attachment.nodeId);
+    const targetIndex = result.findIndex(
+      (node) => node.id === attachment.nodeId
+    );
     if (targetIndex < 0) continue;
     const node = result[targetIndex],
       draft = node.draftValues || {};
-    const next = mergeStoryTaskResult(draft, attachment, task.result || {}, task.id);
+    const next = mergeStoryTaskResult(
+      draft,
+      attachment,
+      task.result || {},
+      task.id
+    );
     if (next !== draft)
       result = result.map((item, index) =>
-        index === targetIndex ? { ...item, draftValues: next } : item,
+        index === targetIndex ? { ...item, draftValues: next } : item
       );
   }
   return result;

lib/video-canvas/task-recovery.ts

@@ -10,7 +10,9 @@ export type RecoveryTask = {
   request: { kind: string; params: Record<string, unknown> };
   result?: Record<string, any> | null;
 };
-export function getEnhanceCheckpoint(task: Pick<RecoveryTask, "request" | "result">) {
+export function getEnhanceCheckpoint(
+  task: Pick<RecoveryTask, "request" | "result">
+) {
   const upstream = task.result?.upstreamTask;
   if (
     task.request.kind !== "image-operation" ||
@@ -27,7 +29,9 @@ export function getEnhanceCheckpoint(task: Pick<RecoveryTask, "request" | "resul
     sourceImageUrl: String(upstream.sourceImageUrl),
   };
 }
-export function getVideoCheckpoint(task: Pick<RecoveryTask, "request" | "result">) {
+export function getVideoCheckpoint(
+  task: Pick<RecoveryTask, "request" | "result">
+) {
   if (task.request.kind !== "video") return null;
   const parsed = videoCheckpointSchema.safeParse(task.result?.upstreamTask);
   if (!parsed.success) return null;
@@ -41,7 +45,9 @@ export function getVideoCheckpoint(task: Pick<RecoveryTask, "request" | "result"
     return null;
   return parsed.data;
 }
-export function getAudioCheckpoint(task: Pick<RecoveryTask, "request" | "result">) {
+export function getAudioCheckpoint(
+  task: Pick<RecoveryTask, "request" | "result">
+) {
   if (task.request.kind !== "audio") return null;
   const parsed = audioCheckpointSchema.safeParse(task.result?.upstreamAudio);
   return parsed.success && parsed.data.modelRef === task.request.params.modelRef
@@ -49,7 +55,7 @@ export function getAudioCheckpoint(task: Pick<RecoveryTask, "request" | "result"
     : null;
 }
 export function getSmartEditCheckpoint(
-  task: Pick<RecoveryTask, "request" | "result">,
+  task: Pick<RecoveryTask, "request" | "result">
 ): SmartEditCheckpoint | null {
   const value = task.result?.smartEdit;
   if (
@@ -75,7 +81,8 @@ export function hasRecoverableCanvasTask(task: RecoveryTask, now = Date.now()) {
       getEnhanceCheckpoint(task) ||
         getVideoCheckpoint(task) ||
         getAudioCheckpoint(task) ||
-        getSmartEditCheckpoint(task),
+        getSmartEditCheckpoint(task) ||
+        getUpscaleCheckpoint(task)
     ) &&
     Number.isFinite(created) &&
     now >= created &&
@@ -84,8 +91,23 @@ export function hasRecoverableCanvasTask(task: RecoveryTask, now = Date.now()) {
   );
 }
 export function canResumeCanvasTask(task: RecoveryTask, now = Date.now()) {
-  if (task.status !== "interrupted" || !hasRecoverableCanvasTask(task, now)) return false;
+  if (task.status !== "interrupted" || !hasRecoverableCanvasTask(task, now))
+    return false;
   const updated = Date.parse(task.updatedAt || "");
-  const delay = Math.min(300000, 30000 * 2 ** Math.min(4, Math.max(0, (task.attempts || 1) - 1)));
+  const delay = Math.min(
+    300000,
+    30000 * 2 ** Math.min(4, Math.max(0, (task.attempts || 1) - 1))
+  );
   return Number.isFinite(updated) && now - updated >= delay;
 }
+
+export function getUpscaleCheckpoint(
+  task: Pick<RecoveryTask, "request" | "result">
+) {
+  const value = task.result?.upscale?.remote;
+  return task.request.kind === "video-upscale" &&
+    /^[a-f0-9]{64}$/.test(value?.id || "") &&
+    /^[a-f0-9]{64}$/.test(value?.endpointHash || "")
+    ? (value as { id: string; endpointHash: string })
+    : null;
+}

lib/video-canvas/task-store.ts

@@ -1,16 +1,20 @@
+import { validateGroupBatch, groupDependencyState } from "./group-execution";
 import { canResumeCanvasTask, hasRecoverableCanvasTask } from "./task-recovery";
 import "server-only";
-import { and, asc, desc, eq, gte, lt, sql } from "drizzle-orm";
+import { and, asc, desc, eq, gte, lt, sql, inArray } from "drizzle-orm";
 import { db } from "@/lib/db";
 import { canvasCapabilityTask as table, canvasProject } from "@/lib/db/schema";
-import { type CanvasTaskRequest, type CanvasGenerationTask } from "./task-contract";
+import {
+  type CanvasTaskRequest,
+  type CanvasGenerationTask,
+} from "./task-contract";
 
 const capabilityId = "video-canvas.generation";
 const scope = (userId: string, projectId: string) =>
   and(
     eq(table.userId, userId),
     eq(table.projectId, projectId),
-    eq(table.capabilityId, capabilityId),
+    eq(table.capabilityId, capabilityId)
   );
 const serialize = (row: typeof table.$inferSelect): CanvasGenerationTask => ({
   id: row.id,
@@ -29,26 +33,37 @@ export async function ownsTaskProject(userId: string, projectId: string) {
       await db
         .select({ id: canvasProject.id })
         .from(canvasProject)
-        .where(and(eq(canvasProject.id, projectId), eq(canvasProject.userId, userId)))
+        .where(
+          and(eq(canvasProject.id, projectId), eq(canvasProject.userId, userId))
+        )
         .limit(1)
-    )[0],
+    )[0]
   );
 }
-export async function createCanvasGenerationTask(userId: string, request: CanvasTaskRequest) {
+export async function createCanvasGenerationTask(
+  userId: string,
+  request: CanvasTaskRequest
+) {
   return db.transaction(async (tx) => {
     await tx.execute(
-      sql`select pg_advisory_xact_lock(hashtext(${capabilityId}), hashtext(${request.projectId}))`,
+      sql`select pg_advisory_xact_lock(hashtext(${capabilityId}), hashtext(${request.projectId}))`
     );
     const existing = (
       await tx
         .select()
         .from(table)
-        .where(and(scope(userId, request.projectId), eq(table.requestId, request.requestId)))
+        .where(
+          and(
+            scope(userId, request.projectId),
+            eq(table.requestId, request.requestId)
+          )
+        )
         .limit(1)
     )[0];
     if (existing) {
       if (
-        JSON.stringify(existing.request) !== JSON.stringify(JSON.parse(JSON.stringify(request)))
+        JSON.stringify(existing.request) !==
+        JSON.stringify(JSON.parse(JSON.stringify(request)))
       ) {
         // JSONB property order is not stable. Compare the immutable logical request.
         const old = existing.request as CanvasTaskRequest;
@@ -57,33 +72,46 @@ export async function createCanvasGenerationTask(userId: string, request: Canvas
           old.kind !== request.kind ||
           JSON.stringify(sortObject(old.attachment)) !==
             JSON.stringify(sortObject(request.attachment)) ||
-          JSON.stringify(sortObject(old.params)) !== JSON.stringify(sortObject(request.params))
+          JSON.stringify(sortObject(old.params)) !==
+            JSON.stringify(sortObject(request.params))
         )
-          throw Object.assign(Error("同一任务 ID 不能用于不同请求"), { status: 409 });
+          throw Object.assign(Error("同一任务 ID 不能用于不同请求"), {
+            status: 409,
+          });
       }
       return serialize(existing);
     }
     const active = await tx
       .select()
       .from(table)
-      .where(and(scope(userId, request.projectId), sql`${table.status} in ('pending','running')`));
+      .where(
+        and(
+          scope(userId, request.projectId),
+          sql`${table.status} in ('pending','running')`
+        )
+      );
     const duplicate = active.find(
-      (row) => (row.request as CanvasTaskRequest).node.id === request.node.id,
+      (row) => (row.request as CanvasTaskRequest).node.id === request.node.id
     );
     if (duplicate) return serialize(duplicate);
     const interrupted = await tx
       .select()
       .from(table)
-      .where(and(scope(userId, request.projectId), eq(table.status, "interrupted")));
+      .where(
+        and(scope(userId, request.projectId), eq(table.status, "interrupted"))
+      );
     const recovering = interrupted.find(
       (row) =>
         (row.request as CanvasTaskRequest).node.id === request.node.id &&
-        hasRecoverableCanvasTask(serialize(row)),
+        hasRecoverableCanvasTask(serialize(row))
     );
     if (recovering) return serialize(recovering);
 
     if (active.length >= 40)
-      throw Object.assign(Error("当前画布排队任务已达 40 个，请等待完成后继续"), { status: 429 });
+      throw Object.assign(
+        Error("当前画布排队任务已达 40 个，请等待完成后继续"),
+        { status: 429 }
+      );
     const [created] = await tx
       .insert(table)
       .values({
@@ -107,11 +135,14 @@ function sortObject(value: any): any {
       ? Object.fromEntries(
           Object.keys(value)
             .sort()
-            .map((key) => [key, sortObject(value[key])]),
+            .map((key) => [key, sortObject(value[key])])
         )
       : value;
 }
-export async function listCanvasGenerationTasks(userId: string, projectId: string) {
+export async function listCanvasGenerationTasks(
+  userId: string,
+  projectId: string
+) {
   // A worker lost to process shutdown must never silently re-submit a paid request.
   await db
     .update(table)
@@ -125,8 +156,8 @@ export async function listCanvasGenerationTasks(userId: string, projectId: strin
       and(
         scope(userId, projectId),
         eq(table.status, "running"),
-        lt(table.updatedAt, new Date(Date.now() - 120000)),
+        lt(table.updatedAt, new Date(Date.now() - 120000))
-      ),
+      )
     );
   const rows = await db
     .select()
@@ -136,10 +167,13 @@ export async function listCanvasGenerationTasks(userId: string, projectId: strin
     .limit(240);
   return rows.map(serialize).reverse();
 }
-export async function claimCanvasGenerationTask(userId: string, projectId: string) {
+export async function claimCanvasGenerationTask(
+  userId: string,
+  projectId: string
+) {
   return db.transaction(async (tx) => {
     await tx.execute(
-      sql`select pg_advisory_xact_lock(hashtext(${capabilityId}), hashtext(${projectId}))`,
+      sql`select pg_advisory_xact_lock(hashtext(${capabilityId}), hashtext(${projectId}))`
     );
     if (
       (
@@ -151,12 +185,57 @@ export async function claimCanvasGenerationTask(userId: string, projectId: strin
       ).length
     )
       return null;
-    let [next] = await tx
+    const pending = await tx
       .select()
       .from(table)
       .where(and(scope(userId, projectId), eq(table.status, "pending")))
       .orderBy(asc(table.createdAt))
-      .limit(1);
+      .limit(40);
+    const dependencyIds = [
+      ...new Set(
+        pending.flatMap(
+          (row) =>
+            (row.request as CanvasTaskRequest).workflow?.inputs
+              .map((input) => input.requestId)
+              .filter((id): id is string => Boolean(id)) || []
+        )
+      ),
+    ];
+    const predecessors = dependencyIds.length
+      ? await tx
+          .select()
+          .from(table)
+          .where(
+            and(
+              scope(userId, projectId),
+              inArray(table.requestId, dependencyIds)
+            )
+          )
+      : [];
+    const tasks = [
+      ...new Map(
+        [...pending, ...predecessors].map((row) => [row.id, serialize(row)])
+      ).values(),
+    ];
+    let next: typeof table.$inferSelect | undefined;
+    for (const candidate of pending) {
+      const state = groupDependencyState(
+        candidate.request as CanvasTaskRequest,
+        tasks
+      );
+      if (state === "failed") {
+        await tx
+          .update(table)
+          .set({
+            status: "failed",
+            error: "上游任务失败或取消，未提交此节点",
+            updatedAt: new Date(),
+            completedAt: new Date(),
+          })
+          .where(eq(table.id, candidate.id));
+        tasks.find((t) => t.id === candidate.id)!.status = "failed";
+      } else if (state === "ready" && !next) next = candidate;
+    }
     if (!next) {
       const interrupted = await tx
         .select()
@@ -166,8 +245,8 @@ export async function claimCanvasGenerationTask(userId: string, projectId: strin
             scope(userId, projectId),
             eq(table.status, "interrupted"),
             gte(table.createdAt, new Date(Date.now() - 86400000)),
-            lt(table.attempts, 20),
+            lt(table.attempts, 20)
-          ),
+          )
         )
         .orderBy(asc(table.updatedAt));
       next = interrupted.find((row) => canResumeCanvasTask(serialize(row)))!;
@@ -187,7 +266,10 @@ export async function claimCanvasGenerationTask(userId: string, projectId: strin
     return claimed ? serialize(claimed) : null;
   });
 }
-export async function heartbeatCanvasGenerationTask(id: string, attempt?: number) {
+export async function heartbeatCanvasGenerationTask(
+  id: string,
+  attempt?: number
+) {
   await db
     .update(table)
     .set({ updatedAt: new Date() })
@@ -195,15 +277,15 @@ export async function heartbeatCanvasGenerationTask(id: string, attempt?: number
       and(
         eq(table.id, id),
         eq(table.status, "running"),
-        attempt === undefined ? undefined : eq(table.attempts, attempt),
+        attempt === undefined ? undefined : eq(table.attempts, attempt)
-      ),
+      )
     );
 }
 export async function finishCanvasGenerationTask(
   id: string,
   result?: Record<string, any>,
   error?: string,
-  attempt?: number,
+  attempt?: number
 ) {
   // A late real result can recover an interrupted task, but cannot replace a newer request ID.
   await db
@@ -219,11 +301,15 @@ export async function finishCanvasGenerationTask(
       and(
         eq(table.id, id),
         sql`${table.status} in ('running','interrupted')`,
-        attempt === undefined ? undefined : eq(table.attempts, attempt),
+        attempt === undefined ? undefined : eq(table.attempts, attempt)
-      ),
+      )
     );
 }
-export async function cancelCanvasGenerationTask(userId: string, projectId: string, id: string) {
+export async function cancelCanvasGenerationTask(
+  userId: string,
+  projectId: string,
+  id: string
+) {
   const [cancelled] = await db
     .update(table)
     .set({
@@ -232,7 +318,13 @@ export async function cancelCanvasGenerationTask(userId: string, projectId: stri
       updatedAt: new Date(),
       completedAt: new Date(),
     })
-    .where(and(scope(userId, projectId), eq(table.id, id), eq(table.status, "pending")))
+    .where(
+      and(
+        scope(userId, projectId),
+        eq(table.id, id),
+        eq(table.status, "pending")
+      )
+    )
     .returning();
   if (cancelled) return { ok: true, task: serialize(cancelled) };
   const [current] = await db
@@ -240,13 +332,16 @@ export async function cancelCanvasGenerationTask(userId: string, projectId: stri
     .from(table)
     .where(and(scope(userId, projectId), eq(table.id, id)))
     .limit(1);
-  return { ok: current?.status === "cancelled", task: current ? serialize(current) : null };
+  return {
+    ok: current?.status === "cancelled",
+    task: current ? serialize(current) : null,
+  };
 }
 export async function hideCanvasGenerationNode(
   userId: string,
   projectId: string,
   id: string,
-  nodeId: string,
+  nodeId: string
 ) {
   return db.transaction(async (tx) => {
     const [task] = await tx
@@ -255,15 +350,22 @@ export async function hideCanvasGenerationNode(
       .where(and(scope(userId, projectId), eq(table.id, id)))
       .for("update");
     if (!task) return false;
-    const request = task.request as CanvasTaskRequest & { hiddenNodeIds?: string[] };
+    const request = task.request as CanvasTaskRequest & {
-    if (nodeId !== request.node.id && !nodeId.startsWith(`${request.node.id}-result-`))
+      hiddenNodeIds?: string[];
+    };
+    if (
+      nodeId !== request.node.id &&
+      !nodeId.startsWith(`${request.node.id}-result-`)
+    )
       return false;
     await tx
       .update(table)
       .set({
         request: {
           ...request,
-          hiddenNodeIds: [...new Set([...(request.hiddenNodeIds || []), nodeId])],
+          hiddenNodeIds: [
+            ...new Set([...(request.hiddenNodeIds || []), nodeId]),
+          ],
         },
       })
       .where(eq(table.id, id));
@@ -274,7 +376,7 @@ export async function hideCanvasGenerationNode(
 export async function checkpointCanvasGenerationTask(
   id: string,
   result: Record<string, unknown>,
-  attempt?: number,
+  attempt?: number
 ) {
   await db
     .update(table)
@@ -283,11 +385,15 @@ export async function checkpointCanvasGenerationTask(
       and(
         eq(table.id, id),
         eq(table.status, "running"),
-        attempt === undefined ? undefined : eq(table.attempts, attempt),
+        attempt === undefined ? undefined : eq(table.attempts, attempt)
-      ),
+      )
     );
 }
-export async function interruptCanvasGenerationTask(id: string, error: string, attempt?: number) {
+export async function interruptCanvasGenerationTask(
+  id: string,
+  error: string,
+  attempt?: number
+) {
   await db
     .update(table)
     .set({ status: "interrupted", error, updatedAt: new Date() })
@@ -295,8 +401,8 @@ export async function interruptCanvasGenerationTask(id: string, error: string, a
       and(
         eq(table.id, id),
         eq(table.status, "running"),
-        attempt === undefined ? undefined : eq(table.attempts, attempt),
+        attempt === undefined ? undefined : eq(table.attempts, attempt)
-      ),
+      )
     );
 }
 
@@ -304,7 +410,114 @@ export async function ownsCanvasTaskLease(id: string, attempt: number) {
   const [row] = await db
     .select({ id: table.id })
     .from(table)
-    .where(and(eq(table.id, id), eq(table.attempts, attempt), eq(table.status, "running")))
+    .where(
+      and(
+        eq(table.id, id),
+        eq(table.attempts, attempt),
+        eq(table.status, "running")
+      )
+    )
     .limit(1);
   return Boolean(row);
 }
+
+export async function createCanvasGroupTasks(
+  userId: string,
+  requests: CanvasTaskRequest[]
+) {
+  validateGroupBatch(requests);
+  return db.transaction(async (tx) => {
+    const projectId = requests[0].projectId;
+    await tx.execute(
+      sql`select pg_advisory_xact_lock(hashtext(${capabilityId}), hashtext(${projectId}))`
+    );
+    const same = await tx
+      .select()
+      .from(table)
+      .where(
+        and(
+          scope(userId, projectId),
+          sql`${table.request}->'workflow'->>'runId' = ${requests[0].workflow!.runId}`
+        )
+      );
+    if (same.length) {
+      if (
+        same.length !== requests.length ||
+        requests.some((r) => {
+          const old = same.find((t) => t.requestId === r.requestId);
+          return (
+            !old ||
+            JSON.stringify(sortObject(old.request)) !==
+              JSON.stringify(sortObject(r))
+          );
+        })
+      )
+        throw Object.assign(Error("同一编组执行编号不能用于不同参数"), {
+          status: 409,
+        });
+      return requests.map((request) =>
+        serialize(same.find((row) => row.requestId === request.requestId)!)
+      );
+    }
+    const candidates = await tx
+      .select()
+      .from(table)
+      .where(
+        and(
+          scope(userId, projectId),
+          sql`(${table.status} in ('pending','running') or (${table.status}='interrupted' and ${table.createdAt} > now() - interval '24 hours'))`
+        )
+      );
+    const active = candidates.filter(
+      (t) =>
+        ["pending", "running"].includes(t.status) ||
+        (t.status === "interrupted" && hasRecoverableCanvasTask(serialize(t)))
+    );
+    if (active.length + requests.length > 40)
+      throw Object.assign(Error("画布任务队列已满"), { status: 429 });
+    if (
+      active.some((t) =>
+        requests.some(
+          (r) => r.node.id === (t.request as CanvasTaskRequest).node.id
+        )
+      )
+    )
+      throw Object.assign(Error("编组中已有节点正在执行，请等待完成"), {
+        status: 409,
+      });
+    const now = Date.now();
+    const rows = await tx
+      .insert(table)
+      .values(
+        requests.map((request, i) => ({
+          requestId: request.requestId,
+          userId,
+          projectId,
+          capabilityId,
+          source: "video-canvas",
+          request,
+          createdAt: new Date(now + i),
+          updatedAt: new Date(now),
+        }))
+      )
+      .returning();
+    return rows.map(serialize);
+  });
+}
+export async function groupTaskInputs(
+  userId: string,
+  request: CanvasTaskRequest
+) {
+  const rows = await db
+    .select()
+    .from(table)
+    .where(
+      and(
+        scope(userId, request.projectId),
+        sql`${table.request}->'workflow'->>'runId' = ${request.workflow?.runId || ""}`
+      )
+    );
+  return rows
+    .map(serialize)
+    .filter((t) => t.request.workflow?.runId === request.workflow?.runId);
+}

lib/video-canvas/task-worker.ts

@@ -1,3 +1,7 @@
+import { resolveGroupInputs } from "./group-execution";
+import { groupTaskInputs } from "./task-store";
+import { executeVideoUpscale } from "./video-upscale";
+import { getUpscaleCheckpoint } from "./task-recovery";
 import { executeSmartEdit } from "./smart-edit/execute";
 import { executeCanvasAudio } from "./audio-generation";
 import { ProviderTaskUncertainError } from "./video-task-lifecycle";
@@ -13,7 +17,10 @@ import {
   ImageOperationUncertainError,
   CanvasTaskLeaseLostError,
 } from "./image-operation";
-import { executeEnhancePOST, executeEnhanceGET } from "@/lib/canvas/operations/enhance";
+import {
+  executeEnhancePOST,
+  executeEnhanceGET,
+} from "@/lib/canvas/operations/enhance";
 import { executeExtendPOST } from "@/lib/canvas/operations/extend";
 import { executeRemoveBackgroundPOST } from "@/lib/canvas/operations/rmbg";
 import { executeAngleEditPOST } from "@/lib/canvas/operations/angle-edit";
@@ -26,7 +33,10 @@ import "server-only";
 import { executeDirectVideoGeneration } from "@/lib/ai/tools/generate-video";
 import { executeCanvasImageResponse } from "@/lib/canvas/generate-image-response";
 import { executeWorkflowAnalysisResponse } from "./workflow-analysis-response";
-import { requireAnalysisResult, type CanvasGenerationTask } from "./task-contract";
+import {
+  requireAnalysisResult,
+  type CanvasGenerationTask,
+} from "./task-contract";
 import { validateVideoReferences } from "./reference-input";
 import { probeRemoteUrl } from "@/lib/net/url-access";
 import {
@@ -46,21 +56,40 @@ async function execute(task: CanvasGenerationTask, userId: string) {
   const videoResume = getVideoCheckpoint(task);
   const audioResume = getAudioCheckpoint(task);
   const editResume = getSmartEditCheckpoint(task);
-  if ((task.attempts || 1) > 1 && !resume && !videoResume && !audioResume && !editResume)
+  const upscaleResume = getUpscaleCheckpoint(task);
+  if (
+    (task.attempts || 1) > 1 &&
+    !resume &&
+    !videoResume &&
+    !audioResume &&
+    !editResume &&
+    !upscaleResume
+  )
     throw Error("缺少上游任务编号，禁止重新提交");
-  const params = request.params;
+  const params = request.workflow
+    ? resolveGroupInputs(request, await groupTaskInputs(userId, request))
+    : request.params;
+  if (request.kind === "video-upscale")
+    return executeVideoUpscale(params, userId, task.id, {
+      resume: upscaleResume || undefined,
+      active: assertLease,
+      save: (value) =>
+        checkpointCanvasGenerationTask(task.id, value, task.attempts),
+    });
   if (request.kind === "smart-edit")
     return executeSmartEdit(params, userId, {
       taskId: task.id,
       resume: editResume,
       assertActive: assertLease,
-      checkpoint: (value) => checkpointCanvasGenerationTask(task.id, value, task.attempts),
+      checkpoint: (value) =>
+        checkpointCanvasGenerationTask(task.id, value, task.attempts),
     });
   if (request.kind === "audio")
     return executeCanvasAudio(params, {
       resume: audioResume,
       assertActive: assertLease,
-      checkpoint: (value) => checkpointCanvasGenerationTask(task.id, value, task.attempts),
+      checkpoint: (value) =>
+        checkpointCanvasGenerationTask(task.id, value, task.attempts),
     });
   if (request.kind === "image-operation")
     return executeImageOperation(
@@ -76,14 +105,15 @@ async function execute(task: CanvasGenerationTask, userId: string) {
       {
         resume,
         assertLease,
-        checkpoint: (result) => checkpointCanvasGenerationTask(task.id, result, task.attempts),
+        checkpoint: (result) =>
-      },
+          checkpointCanvasGenerationTask(task.id, result, task.attempts),
+      }
     );
   if (request.kind === "video") {
     const checked = videoResume
       ? { images: params.images as string[] | undefined, error: undefined }
       : await validateVideoReferences(params.images, (url) =>
-          probeRemoteUrl(url, { timeoutMs: 8000 }),
+          probeRemoteUrl(url, { timeoutMs: 8000 })
         );
     if (checked.error) throw Error(checked.error);
     const result = await executeDirectVideoGeneration({
@@ -95,7 +125,11 @@ async function execute(task: CanvasGenerationTask, userId: string) {
         resume: videoResume,
         assertActive: assertLease,
         save: (checkpoint) =>
-          checkpointCanvasGenerationTask(task.id, { upstreamTask: checkpoint }, task.attempts),
+          checkpointCanvasGenerationTask(
+            task.id,
+            { upstreamTask: checkpoint },
+            task.attempts
+          ),
       },
     });
     if (!result.success || !result.videoUrl) throw Error("视频生成未返回结果");
@@ -105,7 +139,9 @@ async function execute(task: CanvasGenerationTask, userId: string) {
         ...result.metadata,
         temporaryVideoUrl: result.videoUrl,
         storageStatus:
-          result.metadata?.permanentVideoUrl === result.videoUrl ? "archived" : "temporary",
+          result.metadata?.permanentVideoUrl === result.videoUrl
+            ? "archived"
+            : "temporary",
       },
       modelId: result.modelId,
       modelRef: result.modelRef,
@@ -115,35 +151,44 @@ async function execute(task: CanvasGenerationTask, userId: string) {
     request.kind === "image"
       ? await executeCanvasImageResponse(
           { ...params, prompt: String(params.prompt || ""), async: false },
-          userId,
+          userId
         )
       : await executeWorkflowAnalysisResponse(params);
   const body = await response.json();
   if (!response.ok || !body.success) throw Error(body.error || "任务执行失败");
   if (
     request.kind === "image" &&
-    !body.data?.images?.some((image: any) => image.originalUrl || image.thumbnailUrl)
+    !body.data?.images?.some(
+      (image: any) => image.originalUrl || image.thumbnailUrl
+    )
   )
     throw Error("图片生成没有返回结果");
-  if (request.kind === "analysis") requireAnalysisResult(body.data, String(params.mode));
+  if (request.kind === "analysis")
+    requireAnalysisResult(body.data, String(params.mode));
   requireStoryPromptResult(body.data, request.attachment);
   return body.data;
 }
-export async function runCanvasGenerationQueue(userId: string, projectId: string) {
+export async function runCanvasGenerationQueue(
+  userId: string,
+  projectId: string
+) {
   // Work is independent of the browser request. Atomic DB claiming also covers multiple processes.
   for (let i = 0; i < 40; i++) {
     const task = await claimCanvasGenerationTask(userId, projectId);
     if (!task) return;
     const timer = setInterval(
-      () => void heartbeatCanvasGenerationTask(task.id, task.attempts).catch(() => {}),
+      () =>
-      20000,
+        void heartbeatCanvasGenerationTask(task.id, task.attempts).catch(
+          () => {}
+        ),
+      20000
     );
     try {
       await finishCanvasGenerationTask(
         task.id,
         await execute(task, userId),
         undefined,
-        task.attempts,
+        task.attempts
       );
     } catch (error) {
       if (error instanceof CanvasTaskLeaseLostError) return;
@@ -151,14 +196,18 @@ export async function runCanvasGenerationQueue(userId: string, projectId: string
         error instanceof ImageOperationUncertainError ||
         error instanceof ProviderTaskUncertainError
       ) {
-        await interruptCanvasGenerationTask(task.id, error.message, task.attempts);
+        await interruptCanvasGenerationTask(
+          task.id,
+          error.message,
+          task.attempts
+        );
         continue;
       }
       await finishCanvasGenerationTask(
         task.id,
         undefined,
         error instanceof Error ? error.message : "任务执行失败",
-        task.attempts,
+        task.attempts
       );
     } finally {
       clearInterval(timer);

lib/video-canvas/video-upscale-contract.ts

@@ -0,0 +1,33 @@
+import { z } from "zod";
+export const videoUpscaleSchema = z.object({
+  url: z
+    .string()
+    .url()
+    .refine((v) => /^https?:\/\//.test(v)),
+  resolution: z.enum(["1080P", "2K", "4K"]),
+});
+export type VideoUpscaleInput = z.infer<typeof videoUpscaleSchema>;
+export function upscaleDimensions(
+  width: number,
+  height: number,
+  resolution: VideoUpscaleInput["resolution"]
+) {
+  if (![width, height].every((v) => Number.isInteger(v) && v > 0))
+    throw Error("视频尺寸无效");
+  const [long, short] =
+    resolution === "4K"
+      ? [3840, 2160]
+      : resolution === "2K"
+        ? [2560, 1440]
+        : [1920, 1080];
+  const ratio = Math.min(
+    (width >= height ? long : short) / width,
+    (width >= height ? short : long) / height
+  );
+  if (ratio <= 1) throw Error("源视频已达到所选分辨率，请选择更高分辨率");
+  if (ratio > 4) throw Error("单次 AI 超分最大 4 倍，请先选择较低目标分辨率");
+  return {
+    width: Math.floor((width * ratio) / 2) * 2,
+    height: Math.floor((height * ratio) / 2) * 2,
+  };
+}

lib/video-canvas/video-upscale-remote.ts

@@ -0,0 +1,181 @@
+import { createHash } from "node:crypto";
+import { createWriteStream } from "node:fs";
+import { Readable, Transform } from "node:stream";
+import { pipeline } from "node:stream/promises";
+import path from "node:path";
+import { z } from "zod";
+import type { VideoUpscaleInput } from "./video-upscale-contract";
+import { ProviderTaskUncertainError } from "./video-task-lifecycle";
+
+export type UpscaleCheckpoint = { id: string; endpointHash: string };
+const sha = (value: string) => createHash("sha256").update(value).digest("hex");
+export function upscaleConfig() {
+  const value = process.env.VIDEO_UPSCALE_URL;
+  if (!value) return null;
+  const url = new URL(value);
+  if (
+    !["http:", "https:"].includes(url.protocol) ||
+    url.username ||
+    url.password ||
+    url.search ||
+    url.hash
+  )
+    throw Error("视频超分服务地址配置无效");
+  const base = value.replace(/\/$/, ""),
+    token = process.env.VIDEO_UPSCALE_TOKEN;
+  if (!token || token.length < 32) throw Error("视频超分服务认证未配置");
+  return { base, token };
+}
+const statusSchema = z.object({
+  id: z.string().regex(/^[a-f0-9]{64}$/),
+  status: z.enum(["queued", "running", "succeeded", "failed"]),
+  stage: z.string().max(300),
+  error: z.string().max(500).optional(),
+  result: z
+    .object({
+      width: z.number().int().positive().max(3840),
+      height: z.number().int().positive().max(3840),
+      durationSeconds: z.number().positive().max(120.2),
+      fps: z.number().positive().max(60),
+      frames: z.number().int().positive().max(7200),
+      audio: z.boolean(),
+    })
+    .optional(),
+});
+export async function remoteUpscale(
+  plan: VideoUpscaleInput,
+  userId: string,
+  taskId: string,
+  dir: string,
+  lifecycle: {
+    resume?: UpscaleCheckpoint;
+    active: () => Promise<void>;
+    save: (checkpoint: UpscaleCheckpoint, stage: string) => Promise<void>;
+  }
+) {
+  const config = upscaleConfig();
+  if (!config)
+    throw Error(
+      lifecycle.resume
+        ? "远程超分任务的服务配置已移除，不能回退本机重复超分"
+        : "视频高清服务尚未接通"
+    );
+  const checkpoint = lifecycle.resume || {
+    id: sha("video-upscale-v1:" + userId + ":" + taskId),
+    endpointHash: sha(config.base),
+  };
+  if (
+    !/^[a-f0-9]{64}$/.test(checkpoint.id) ||
+    checkpoint.endpointHash !== sha(config.base)
+  )
+    throw Error("超分任务所属服务器已变化，请恢复原服务配置后继续");
+  await lifecycle.save(checkpoint, "提交算力服务器");
+  const headers = {
+    authorization: "Bearer " + config.token,
+    "content-type": "application/json",
+  };
+  async function request(route: string, body?: unknown) {
+    await lifecycle.active();
+    let response: Response;
+    try {
+      response = await fetch(config!.base + route, {
+        method: body ? "POST" : "GET",
+        headers,
+        body: body ? JSON.stringify(body) : undefined,
+        signal: AbortSignal.timeout(20000),
+        redirect: "error",
+        cache: "no-store",
+      });
+    } catch {
+      throw new ProviderTaskUncertainError(
+        "算力服务器连接中断，任务已保存，可继续查询"
+      );
+    }
+    if ([429, 500, 502, 503, 504].includes(response.status))
+      throw new ProviderTaskUncertainError(
+        "算力服务器繁忙，任务已保存等待继续"
+      );
+    if (!response.ok)
+      throw Error(
+        `超分服务请求失败（${response.status}），请检查服务状态或任务参数`
+      );
+    let data: unknown;
+    try {
+      data = await response.json();
+    } catch {
+      throw new ProviderTaskUncertainError("超分服务响应中断，将查询同一任务");
+    }
+    const result = statusSchema.parse(data);
+    if (result.id !== checkpoint.id)
+      throw Error("超分服务返回了不匹配的任务编号");
+    return result;
+  }
+  let status = await request("/v1/upscales", {
+    id: checkpoint.id,
+    tenant: sha(userId),
+    plan,
+  });
+  const deadline = Date.now() + 8 * 60000;
+  while (status.status === "queued" || status.status === "running") {
+    await lifecycle.save(checkpoint, status.stage);
+    if (Date.now() > deadline)
+      throw new ProviderTaskUncertainError("远程超分仍在运行，已保存任务进度");
+    await new Promise((resolve) => setTimeout(resolve, 3000));
+    status = await request("/v1/upscales/" + checkpoint.id);
+  }
+  if (status.status === "failed")
+    throw Error(status.error || "算力服务器超分失败");
+  if (!status.result) throw Error("超分服务未返回成片参数");
+  await lifecycle.save(checkpoint, "读取远程成片");
+  async function download(kind: string, destination: string, maxBytes: number) {
+    await lifecycle.active();
+    try {
+      const response = await fetch(
+        config!.base + "/v1/upscales/" + checkpoint.id + "/" + kind,
+        {
+          headers,
+          signal: AbortSignal.timeout(180000),
+          redirect: "error",
+          cache: "no-store",
+        }
+      );
+      if (
+        !response.ok ||
+        !response.body ||
+        Number(response.headers.get("content-length")) > maxBytes
+      )
+        throw Error("远程成片不可读取或超过处理上限");
+      let size = 0;
+      const limit = new Transform({
+        transform(chunk, _encoding, callback) {
+          size += chunk.length;
+          callback(
+            size > maxBytes ? Error("远程成片超过处理上限") : null,
+            chunk
+          );
+        },
+      });
+      await pipeline(
+        Readable.fromWeb(response.body as any),
+        limit,
+        createWriteStream(destination, { mode: 0o600 })
+      );
+    } catch {
+      throw new ProviderTaskUncertainError(
+        "成片已生成但读取中断，将继续读取同一任务"
+      );
+    }
+  }
+  const file = path.join(dir, "result.mp4");
+  await download("video", file, 300 * 1024 * 1024);
+
+  return {
+    file,
+    width: status.result.width,
+    height: status.result.height,
+    duration: status.result.durationSeconds,
+    fps: status.result.fps,
+    frames: status.result.frames,
+    audio: status.result.audio,
+  };
+}

lib/video-canvas/video-upscale.ts

@@ -0,0 +1,62 @@
+import { ProviderTaskUncertainError } from "./video-task-lifecycle";
+import "server-only";
+import { mkdtemp, readFile, rm } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import path from "node:path";
+import { uploadFile } from "@/lib/storage/service";
+import { videoUpscaleSchema } from "./video-upscale-contract";
+import { remoteUpscale, type UpscaleCheckpoint } from "./video-upscale-remote";
+export async function executeVideoUpscale(
+  params: Record<string, unknown>,
+  userId: string,
+  taskId: string,
+  lifecycle: {
+    resume?: UpscaleCheckpoint;
+    active: () => Promise<void>;
+    save: (value: Record<string, unknown>) => Promise<void>;
+  }
+) {
+  const input = videoUpscaleSchema.parse(params),
+    dir = await mkdtemp(path.join(tmpdir(), "upscale-"));
+  try {
+    const output = await remoteUpscale(input, userId, taskId, dir, {
+      resume: lifecycle.resume,
+      active: lifecycle.active,
+      save: (remote, stage) => lifecycle.save({ upscale: { remote, stage } }),
+    });
+    await lifecycle.active();
+    const file = await readFile(output.file);
+    let stored;
+    try {
+      stored = await uploadFile(
+        `video-upscale-${taskId}.mp4`,
+        (file.byteOffset === 0 && file.byteLength === file.buffer.byteLength
+          ? file.buffer
+          : file.buffer.slice(
+              file.byteOffset,
+              file.byteOffset + file.byteLength
+            )) as ArrayBuffer
+      );
+    } catch {
+      throw new ProviderTaskUncertainError(
+        "高清结果已生成，归档暂时失败，将继续归档同一结果"
+      );
+    }
+    return {
+      url: stored.url,
+      metadata: {
+        width: output.width,
+        height: output.height,
+        durationSeconds: output.duration,
+        fps: output.fps,
+        frames: output.frames,
+        hasAudio: output.audio,
+        engine: "Real-ESRGAN",
+        model: "realesr-general-x4v3",
+        resolution: input.resolution,
+      },
+    };
+  } finally {
+    await rm(dir, { recursive: true, force: true });
+  }
+}

next.config.ts

@@ -37,8 +37,9 @@ const nextConfig: NextConfig = {
   // 外部包配置（Next.js 16+）
   serverExternalPackages: ["sharp"],
   outputFileTracingIncludes: {
-    "/api/canvas/generation-tasks": ["./services/smart-edit/analyze.py"],
+    "/api/canvas/generation-tasks": ["./services/smart-edit/analyze.py", "./services/smart-edit/separate.py", "./services/smart-edit/font.py"],
-    "/api/canvas/smart-edit/capabilities": ["./services/smart-edit/analyze.py"],
+    "/api/canvas/smart-edit/fonts/[id]": ["./services/smart-edit/font.py"],
+    "/api/canvas/smart-edit/capabilities": ["./services/smart-edit/analyze.py", "./services/smart-edit/separate.py", "./services/smart-edit/font.py"],
   },
   transpilePackages: ["@sparkjsdev/spark"],
   typescript: {

playwright.helpers.config.ts

@@ -3,6 +3,8 @@ import { defineConfig } from "@playwright/test";
 export default defineConfig({
   testDir: "./tests",
   testMatch: [
+    /routes\/video-canvas-groups-upscale\.test\.ts/,
+    /routes\/video-canvas-connections\.test\.ts/,
     /routes\/video-canvas-smart-edit\.test\.ts/,
     /routes\/video-canvas-tasks\.test\.ts/,
     /routes\/video-canvas-director\.test\.ts/,

public/fonts/smart-edit/MANIFEST.json

@@ -0,0 +1,36 @@
+[
+  {
+    "file": "roboto.ttf",
+    "upstreamSha256": "d7598e12c5dbef095ff8272cfc55da0250bd07fbdecbac8a530b9b277872a134",
+    "sha256": "b55c3f5427be8b29e743ebc6a64128af5824c0cffd198fcfc643e4c5eafde0c8",
+    "family": "Roboto",
+    "weight": 400
+  },
+  {
+    "file": "noto-serif-sc.ttf",
+    "upstreamSha256": "050080d9255a86808f2945bffac582b31ef32bc36411ce29563b4961670c66f9",
+    "sha256": "08ac21d3d82d929cb493f2523532dd32dce1d797964de493071b64df87afa6ed",
+    "family": "Noto Serif SC ExtraLight",
+    "weight": 400
+  },
+  {
+    "file": "noto-sans-sc.ttf",
+    "upstreamSha256": "a3041811a78c361b1de50f953c805e0244951c21c5bd412f7232ef0d899af0da",
+    "sha256": "36716fbd3b964480d238030cab171b0ea756e959adc3d323995e2433a06c61e5",
+    "family": "Noto Sans SC Thin",
+    "weight": 400
+  },
+  {
+    "file": "inter.ttf",
+    "upstreamSha256": "29160a80ff49ddcab2c97711247e08b1fab27a484a329ce8b813d820dc559031",
+    "sha256": "54d85e725cf1f763acda27f27bda373f0d0b9117fbf6f1d9015f04004d1ebd3d",
+    "family": "Inter",
+    "weight": 400
+  },
+  {
+    "file": "source-han.otf",
+    "sha256": "f1d8611151880c6c336aabeac4640ef434fa13cbfbf1ffe82d0a71b2a5637256",
+    "family": "Source Han Sans SC",
+    "weight": 400
+  }
+]

public/fonts/smart-edit/README.md

@@ -0,0 +1,9 @@
+# 智能剪辑字幕字体
+
+来源：Adobe 官方 [Source Han Sans](https://github.com/adobe-fonts/source-han-sans/tree/release/OTF/SimplifiedChinese) 与 [Google Fonts](https://github.com/google/fonts/tree/main/ofl)。各字体的完整 SIL Open Font License 文本随文件保存。`SOURCES.json` 保存下载来源；`MANIFEST.json` 保存文件 SHA-256 和上游下载摘要。
+
+Google 的可变 TTF 使用 fontTools 4.59.1 实例化为 400 字重，其他轴保持上游默认值，以让浏览器与 libass 使用相同静态轮廓。Adobe 文件为未修改的 Regular OTF。Source Han 在选单保留参考名称 Source Han Sans CN，实际文件的内置家族名为 Source Han Sans SC；Google CJK 文件保留上游 name 表，其排版家族名由 `nameID=16` 定义，旧版家族名可能包含 Thin/ExtraLight，不代表当前实例的字重。
+
+浏览器字体按需加载，使用所选 TTF/OTF 原文件；导出把同一文件和中文回退 Source Han 复制到任务的 `fonts/` 目录，并通过 libass `fontsdir` 引用。构建独立合成服务时会把整个字体包复制到 `dist/fonts`。
+
+这五个实际可用字体不等于参考站完整目录；未添加只有名称、没有字体文件的假选项。

public/fonts/smart-edit/SOURCES.json

@@ -0,0 +1,37 @@
+[
+  [
+    "source-han",
+    "adobe-fonts/source-han-sans",
+    "release",
+    "OTF/SimplifiedChinese/SourceHanSansSC-Regular.otf",
+    "LICENSE.txt"
+  ],
+  [
+    "noto-sans-sc",
+    "google/fonts",
+    "main",
+    "ofl/notosanssc/NotoSansSC[wght].ttf",
+    "ofl/notosanssc/OFL.txt"
+  ],
+  [
+    "noto-serif-sc",
+    "google/fonts",
+    "main",
+    "ofl/notoserifsc/NotoSerifSC[wght].ttf",
+    "ofl/notoserifsc/OFL.txt"
+  ],
+  [
+    "roboto",
+    "google/fonts",
+    "main",
+    "ofl/roboto/Roboto[wdth,wght].ttf",
+    "ofl/roboto/OFL.txt"
+  ],
+  [
+    "inter",
+    "google/fonts",
+    "main",
+    "ofl/inter/Inter[opsz,wght].ttf",
+    "ofl/inter/OFL.txt"
+  ]
+]

public/fonts/smart-edit/inter-LICENSE.txt

@@ -0,0 +1,93 @@
+Copyright 2020 The Inter Project Authors (https://github.com/rsms/inter)
+
+This Font Software is licensed under the SIL Open Font License, Version 1.1.
+This license is copied below, and is also available with a FAQ at:
+https://scripts.sil.org/OFL
+
+
+-----------------------------------------------------------
+SIL OPEN FONT LICENSE Version 1.1 - 26 February 2007
+-----------------------------------------------------------
+
+PREAMBLE
+The goals of the Open Font License (OFL) are to stimulate worldwide
+development of collaborative font projects, to support the font creation
+efforts of academic and linguistic communities, and to provide a free and
+open framework in which fonts may be shared and improved in partnership
+with others.
+
+The OFL allows the licensed fonts to be used, studied, modified and
+redistributed freely as long as they are not sold by themselves. The
+fonts, including any derivative works, can be bundled, embedded,
+redistributed and/or sold with any software provided that any reserved
+names are not used by derivative works. The fonts and derivatives,
+however, cannot be released under any other type of license. The
+requirement for fonts to remain under this license does not apply
+to any document created using the fonts or their derivatives.
+
+DEFINITIONS
+"Font Software" refers to the set of files released by the Copyright
+Holder(s) under this license and clearly marked as such. This may
+include source files, build scripts and documentation.
+
+"Reserved Font Name" refers to any names specified as such after the
+copyright statement(s).
+
+"Original Version" refers to the collection of Font Software components as
+distributed by the Copyright Holder(s).
+
+"Modified Version" refers to any derivative made by adding to, deleting,
+or substituting -- in part or in whole -- any of the components of the
+Original Version, by changing formats or by porting the Font Software to a
+new environment.
+
+"Author" refers to any designer, engineer, programmer, technical
+writer or other person who contributed to the Font Software.
+
+PERMISSION & CONDITIONS
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of the Font Software, to use, study, copy, merge, embed, modify,
+redistribute, and sell modified and unmodified copies of the Font
+Software, subject to the following conditions:
+
+1) Neither the Font Software nor any of its individual components,
+in Original or Modified Versions, may be sold by itself.
+
+2) Original or Modified Versions of the Font Software may be bundled,
+redistributed and/or sold with any software, provided that each copy
+contains the above copyright notice and this license. These can be
+included either as stand-alone text files, human-readable headers or
+in the appropriate machine-readable metadata fields within text or
+binary files as long as those fields can be easily viewed by the user.
+
+3) No Modified Version of the Font Software may use the Reserved Font
+Name(s) unless explicit written permission is granted by the corresponding
+Copyright Holder. This restriction only applies to the primary font name as
+presented to the users.
+
+4) The name(s) of the Copyright Holder(s) or the Author(s) of the Font
+Software shall not be used to promote, endorse or advertise any
+Modified Version, except to acknowledge the contribution(s) of the
+Copyright Holder(s) and the Author(s) or with their explicit written
+permission.
+
+5) The Font Software, modified or unmodified, in part or in whole,
+must be distributed entirely under this license, and must not be
+distributed under any other license. The requirement for fonts to
+remain under this license does not apply to any document created
+using the Font Software.
+
+TERMINATION
+This license becomes null and void if any of the above conditions are
+not met.
+
+DISCLAIMER
+THE FONT SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT
+OF COPYRIGHT, PATENT, TRADEMARK, OR OTHER RIGHT. IN NO EVENT SHALL THE
+COPYRIGHT HOLDER BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
+INCLUDING ANY GENERAL, SPECIAL, INDIRECT, INCIDENTAL, OR CONSEQUENTIAL
+DAMAGES, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+FROM, OUT OF THE USE OR INABILITY TO USE THE FONT SOFTWARE OR FROM
+OTHER DEALINGS IN THE FONT SOFTWARE.

public/fonts/smart-edit/noto-sans-sc-LICENSE.txt

@@ -0,0 +1,93 @@
+Copyright 2014-2021 Adobe (http://www.adobe.com/), with Reserved Font Name 'Source'
+
+This Font Software is licensed under the SIL Open Font License, Version 1.1.
+This license is copied below, and is also available with a FAQ at:
+https://scripts.sil.org/OFL
+
+
+-----------------------------------------------------------
+SIL OPEN FONT LICENSE Version 1.1 - 26 February 2007
+-----------------------------------------------------------
+
+PREAMBLE
+The goals of the Open Font License (OFL) are to stimulate worldwide
+development of collaborative font projects, to support the font creation
+efforts of academic and linguistic communities, and to provide a free and
+open framework in which fonts may be shared and improved in partnership
+with others.
+
+The OFL allows the licensed fonts to be used, studied, modified and
+redistributed freely as long as they are not sold by themselves. The
+fonts, including any derivative works, can be bundled, embedded,
+redistributed and/or sold with any software provided that any reserved
+names are not used by derivative works. The fonts and derivatives,
+however, cannot be released under any other type of license. The
+requirement for fonts to remain under this license does not apply
+to any document created using the fonts or their derivatives.
+
+DEFINITIONS
+"Font Software" refers to the set of files released by the Copyright
+Holder(s) under this license and clearly marked as such. This may
+include source files, build scripts and documentation.
+
+"Reserved Font Name" refers to any names specified as such after the
+copyright statement(s).
+
+"Original Version" refers to the collection of Font Software components as
+distributed by the Copyright Holder(s).
+
+"Modified Version" refers to any derivative made by adding to, deleting,
+or substituting -- in part or in whole -- any of the components of the
+Original Version, by changing formats or by porting the Font Software to a
+new environment.
+
+"Author" refers to any designer, engineer, programmer, technical
+writer or other person who contributed to the Font Software.
+
+PERMISSION & CONDITIONS
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of the Font Software, to use, study, copy, merge, embed, modify,
+redistribute, and sell modified and unmodified copies of the Font
+Software, subject to the following conditions:
+
+1) Neither the Font Software nor any of its individual components,
+in Original or Modified Versions, may be sold by itself.
+
+2) Original or Modified Versions of the Font Software may be bundled,
+redistributed and/or sold with any software, provided that each copy
+contains the above copyright notice and this license. These can be
+included either as stand-alone text files, human-readable headers or
+in the appropriate machine-readable metadata fields within text or
+binary files as long as those fields can be easily viewed by the user.
+
+3) No Modified Version of the Font Software may use the Reserved Font
+Name(s) unless explicit written permission is granted by the corresponding
+Copyright Holder. This restriction only applies to the primary font name as
+presented to the users.
+
+4) The name(s) of the Copyright Holder(s) or the Author(s) of the Font
+Software shall not be used to promote, endorse or advertise any
+Modified Version, except to acknowledge the contribution(s) of the
+Copyright Holder(s) and the Author(s) or with their explicit written
+permission.
+
+5) The Font Software, modified or unmodified, in part or in whole,
+must be distributed entirely under this license, and must not be
+distributed under any other license. The requirement for fonts to
+remain under this license does not apply to any document created
+using the Font Software.
+
+TERMINATION
+This license becomes null and void if any of the above conditions are
+not met.
+
+DISCLAIMER
+THE FONT SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT
+OF COPYRIGHT, PATENT, TRADEMARK, OR OTHER RIGHT. IN NO EVENT SHALL THE
+COPYRIGHT HOLDER BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
+INCLUDING ANY GENERAL, SPECIAL, INDIRECT, INCIDENTAL, OR CONSEQUENTIAL
+DAMAGES, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+FROM, OUT OF THE USE OR INABILITY TO USE THE FONT SOFTWARE OR FROM
+OTHER DEALINGS IN THE FONT SOFTWARE.

public/fonts/smart-edit/noto-serif-sc-LICENSE.txt

@@ -0,0 +1,94 @@
+Copyright 2012 Google Inc. All Rights Reserved.
+
+This Font Software is licensed under the SIL Open Font License,
+Version 1.1.
+
+This license is copied below, and is also available with a FAQ at:
+http://scripts.sil.org/OFL
+
+-----------------------------------------------------------
+SIL OPEN FONT LICENSE Version 1.1 - 26 February 2007
+-----------------------------------------------------------
+
+PREAMBLE
+The goals of the Open Font License (OFL) are to stimulate worldwide
+development of collaborative font projects, to support the font
+creation efforts of academic and linguistic communities, and to
+provide a free and open framework in which fonts may be shared and
+improved in partnership with others.
+
+The OFL allows the licensed fonts to be used, studied, modified and
+redistributed freely as long as they are not sold by themselves. The
+fonts, including any derivative works, can be bundled, embedded,
+redistributed and/or sold with any software provided that any reserved
+names are not used by derivative works. The fonts and derivatives,
+however, cannot be released under any other type of license. The
+requirement for fonts to remain under this license does not apply to
+any document created using the fonts or their derivatives.
+
+DEFINITIONS
+"Font Software" refers to the set of files released by the Copyright
+Holder(s) under this license and clearly marked as such. This may
+include source files, build scripts and documentation.
+
+"Reserved Font Name" refers to any names specified as such after the
+copyright statement(s).
+
+"Original Version" refers to the collection of Font Software
+components as distributed by the Copyright Holder(s).
+
+"Modified Version" refers to any derivative made by adding to,
+deleting, or substituting -- in part or in whole -- any of the
+components of the Original Version, by changing formats or by porting
+the Font Software to a new environment.
+
+"Author" refers to any designer, engineer, programmer, technical
+writer or other person who contributed to the Font Software.
+
+PERMISSION & CONDITIONS
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of the Font Software, to use, study, copy, merge, embed,
+modify, redistribute, and sell modified and unmodified copies of the
+Font Software, subject to the following conditions:
+
+1) Neither the Font Software nor any of its individual components, in
+Original or Modified Versions, may be sold by itself.
+
+2) Original or Modified Versions of the Font Software may be bundled,
+redistributed and/or sold with any software, provided that each copy
+contains the above copyright notice and this license. These can be
+included either as stand-alone text files, human-readable headers or
+in the appropriate machine-readable metadata fields within text or
+binary files as long as those fields can be easily viewed by the user.
+
+3) No Modified Version of the Font Software may use the Reserved Font
+Name(s) unless explicit written permission is granted by the
+corresponding Copyright Holder. This restriction only applies to the
+primary font name as presented to the users.
+
+4) The name(s) of the Copyright Holder(s) or the Author(s) of the Font
+Software shall not be used to promote, endorse or advertise any
+Modified Version, except to acknowledge the contribution(s) of the
+Copyright Holder(s) and the Author(s) or with their explicit written
+permission.
+
+5) The Font Software, modified or unmodified, in part or in whole,
+must be distributed entirely under this license, and must not be
+distributed under any other license. The requirement for fonts to
+remain under this license does not apply to any document created using
+the Font Software.
+
+TERMINATION
+This license becomes null and void if any of the above conditions are
+not met.
+
+DISCLAIMER
+THE FONT SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT
+OF COPYRIGHT, PATENT, TRADEMARK, OR OTHER RIGHT. IN NO EVENT SHALL THE
+COPYRIGHT HOLDER BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
+INCLUDING ANY GENERAL, SPECIAL, INDIRECT, INCIDENTAL, OR CONSEQUENTIAL
+DAMAGES, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+FROM, OUT OF THE USE OR INABILITY TO USE THE FONT SOFTWARE OR FROM
+OTHER DEALINGS IN THE FONT SOFTWARE.

public/fonts/smart-edit/roboto-LICENSE.txt

@@ -0,0 +1,93 @@
+Copyright 2011 The Roboto Project Authors (https://github.com/googlefonts/roboto-classic)
+
+This Font Software is licensed under the SIL Open Font License, Version 1.1.
+This license is copied below, and is also available with a FAQ at:
+https://openfontlicense.org
+
+
+-----------------------------------------------------------
+SIL OPEN FONT LICENSE Version 1.1 - 26 February 2007
+-----------------------------------------------------------
+
+PREAMBLE
+The goals of the Open Font License (OFL) are to stimulate worldwide
+development of collaborative font projects, to support the font creation
+efforts of academic and linguistic communities, and to provide a free and
+open framework in which fonts may be shared and improved in partnership
+with others.
+
+The OFL allows the licensed fonts to be used, studied, modified and
+redistributed freely as long as they are not sold by themselves. The
+fonts, including any derivative works, can be bundled, embedded,
+redistributed and/or sold with any software provided that any reserved
+names are not used by derivative works. The fonts and derivatives,
+however, cannot be released under any other type of license. The
+requirement for fonts to remain under this license does not apply
+to any document created using the fonts or their derivatives.
+
+DEFINITIONS
+"Font Software" refers to the set of files released by the Copyright
+Holder(s) under this license and clearly marked as such. This may
+include source files, build scripts and documentation.
+
+"Reserved Font Name" refers to any names specified as such after the
+copyright statement(s).
+
+"Original Version" refers to the collection of Font Software components as
+distributed by the Copyright Holder(s).
+
+"Modified Version" refers to any derivative made by adding to, deleting,
+or substituting -- in part or in whole -- any of the components of the
+Original Version, by changing formats or by porting the Font Software to a
+new environment.
+
+"Author" refers to any designer, engineer, programmer, technical
+writer or other person who contributed to the Font Software.
+
+PERMISSION & CONDITIONS
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of the Font Software, to use, study, copy, merge, embed, modify,
+redistribute, and sell modified and unmodified copies of the Font
+Software, subject to the following conditions:
+
+1) Neither the Font Software nor any of its individual components,
+in Original or Modified Versions, may be sold by itself.
+
+2) Original or Modified Versions of the Font Software may be bundled,
+redistributed and/or sold with any software, provided that each copy
+contains the above copyright notice and this license. These can be
+included either as stand-alone text files, human-readable headers or
+in the appropriate machine-readable metadata fields within text or
+binary files as long as those fields can be easily viewed by the user.
+
+3) No Modified Version of the Font Software may use the Reserved Font
+Name(s) unless explicit written permission is granted by the corresponding
+Copyright Holder. This restriction only applies to the primary font name as
+presented to the users.
+
+4) The name(s) of the Copyright Holder(s) or the Author(s) of the Font
+Software shall not be used to promote, endorse or advertise any
+Modified Version, except to acknowledge the contribution(s) of the
+Copyright Holder(s) and the Author(s) or with their explicit written
+permission.
+
+5) The Font Software, modified or unmodified, in part or in whole,
+must be distributed entirely under this license, and must not be
+distributed under any other license. The requirement for fonts to
+remain under this license does not apply to any document created
+using the Font Software.
+
+TERMINATION
+This license becomes null and void if any of the above conditions are
+not met.
+
+DISCLAIMER
+THE FONT SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT
+OF COPYRIGHT, PATENT, TRADEMARK, OR OTHER RIGHT. IN NO EVENT SHALL THE
+COPYRIGHT HOLDER BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
+INCLUDING ANY GENERAL, SPECIAL, INDIRECT, INCIDENTAL, OR CONSEQUENTIAL
+DAMAGES, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+FROM, OUT OF THE USE OR INABILITY TO USE THE FONT SOFTWARE OR FROM
+OTHER DEALINGS IN THE FONT SOFTWARE.

public/fonts/smart-edit/source-han-LICENSE.txt

@@ -0,0 +1,96 @@
+Copyright 2014-2025 Adobe (http://www.adobe.com/), with Reserved Font
+Name 'Source'. Source is a trademark of Adobe in the United States
+and/or other countries.
+
+This Font Software is licensed under the SIL Open Font License,
+Version 1.1.
+
+This license is copied below, and is also available with a FAQ at:
+http://scripts.sil.org/OFL
+
+-----------------------------------------------------------
+SIL OPEN FONT LICENSE Version 1.1 - 26 February 2007
+-----------------------------------------------------------
+
+PREAMBLE
+The goals of the Open Font License (OFL) are to stimulate worldwide
+development of collaborative font projects, to support the font
+creation efforts of academic and linguistic communities, and to
+provide a free and open framework in which fonts may be shared and
+improved in partnership with others.
+
+The OFL allows the licensed fonts to be used, studied, modified and
+redistributed freely as long as they are not sold by themselves. The
+fonts, including any derivative works, can be bundled, embedded,
+redistributed and/or sold with any software provided that any reserved
+names are not used by derivative works. The fonts and derivatives,
+however, cannot be released under any other type of license. The
+requirement for fonts to remain under this license does not apply to
+any document created using the fonts or their derivatives.
+
+DEFINITIONS
+"Font Software" refers to the set of files released by the Copyright
+Holder(s) under this license and clearly marked as such. This may
+include source files, build scripts and documentation.
+
+"Reserved Font Name" refers to any names specified as such after the
+copyright statement(s).
+
+"Original Version" refers to the collection of Font Software
+components as distributed by the Copyright Holder(s).
+
+"Modified Version" refers to any derivative made by adding to,
+deleting, or substituting -- in part or in whole -- any of the
+components of the Original Version, by changing formats or by porting
+the Font Software to a new environment.
+
+"Author" refers to any designer, engineer, programmer, technical
+writer or other person who contributed to the Font Software.
+
+PERMISSION & CONDITIONS
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of the Font Software, to use, study, copy, merge, embed,
+modify, redistribute, and sell modified and unmodified copies of the
+Font Software, subject to the following conditions:
+
+1) Neither the Font Software nor any of its individual components, in
+Original or Modified Versions, may be sold by itself.
+
+2) Original or Modified Versions of the Font Software may be bundled,
+redistributed and/or sold with any software, provided that each copy
+contains the above copyright notice and this license. These can be
+included either as stand-alone text files, human-readable headers or
+in the appropriate machine-readable metadata fields within text or
+binary files as long as those fields can be easily viewed by the user.
+
+3) No Modified Version of the Font Software may use the Reserved Font
+Name(s) unless explicit written permission is granted by the
+corresponding Copyright Holder. This restriction only applies to the
+primary font name as presented to the users.
+
+4) The name(s) of the Copyright Holder(s) or the Author(s) of the Font
+Software shall not be used to promote, endorse or advertise any
+Modified Version, except to acknowledge the contribution(s) of the
+Copyright Holder(s) and the Author(s) or with their explicit written
+permission.
+
+5) The Font Software, modified or unmodified, in part or in whole,
+must be distributed entirely under this license, and must not be
+distributed under any other license. The requirement for fonts to
+remain under this license does not apply to any document created using
+the Font Software.
+
+TERMINATION
+This license becomes null and void if any of the above conditions are
+not met.
+
+DISCLAIMER
+THE FONT SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT
+OF COPYRIGHT, PATENT, TRADEMARK, OR OTHER RIGHT. IN NO EVENT SHALL THE
+COPYRIGHT HOLDER BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
+INCLUDING ANY GENERAL, SPECIAL, INDIRECT, INCIDENTAL, OR CONSEQUENTIAL
+DAMAGES, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+FROM, OUT OF THE USE OR INABILITY TO USE THE FONT SOFTWARE OR FROM
+OTHER DEALINGS IN THE FONT SOFTWARE.

public/video-canvas/transitions/README.md

@@ -0,0 +1,13 @@
+# 转场库动态预览素材
+
+按用户明确要求从参考站可见的转场卡片取得原始动态 WebP，原字节保存，不截成静态图片、不生成替代图，也不使用运行时外链。每张为 75 帧、3 秒循环。
+
+来源页面：https://www.liblib.tv/canvas
+
+| 本地文件 | 参考页面实际图片地址 | SHA-256 |
+| --- | --- | --- |
+| fade.webp | https://libtv-res.liblib.art/upload-images/7769a090dda6411dbed73c8da22ea01e/52744c87f20f2d682118c51360237c8ecf2e5cf0.webp | bee8d24aa6db9c4898645e4540fa647fbd720d2640ecda823be26b1c7ef59579 |
+| fadeblack.webp | https://libtv-res.liblib.art/upload-images/7769a090dda6411dbed73c8da22ea01e/33b60ac24835284581b432ec3dc6770c28c4cdfd.webp | dbfdb0f9b17389a5239fecee2322d7eef64a9cb387600d7b77e9ec0a05408f66 |
+| fadewhite.webp | https://libtv-res.liblib.art/upload-images/7769a090dda6411dbed73c8da22ea01e/887e2a7ed6539e258a98fa01d003ca60cb1a4c79.webp | 648d6025e39206ddbc5f63b8a6415d62b5a44f1ca9d58822d9df58928dec9ce9 |
+
+卡片图片只用于展示效果类别。右侧预览与成片合成使用用户选中的相邻素材，不把示例图作为用户成片素材。

scripts/qa/smart-edit-main-chain.cjs

@@ -0,0 +1,214 @@
+// Real local Next/auth/database/storage/task/model/render chain. Requires an isolated account.json and speech.mp4 in SMART_EDIT_QA_DIR. Never target production.
+const fs = require("fs"),
+  crypto = require("crypto"),
+  path = require("path"),
+  assert = require("assert/strict");
+const q = path.resolve(process.env.SMART_EDIT_QA_DIR || ".qa/smart-edit/main-chain"),
+  base = process.env.SMART_EDIT_QA_URL || "http://127.0.0.1:3025";
+assert(
+  ["127.0.0.1", "localhost"].includes(new URL(base).hostname),
+  "Use an isolated local server only",
+);
+fs.mkdirSync(q, { recursive: true });
+let cookies = {};
+let report = [];
+async function request(p, o = {}) {
+  const r = await fetch(base + p, {
+    ...o,
+    headers: {
+      ...o.headers,
+      Cookie: Object.entries(cookies)
+        .map(([k, v]) => k + "=" + v)
+        .join("; "),
+    },
+    redirect: "manual",
+    signal: AbortSignal.timeout(180000),
+  });
+  for (const c of r.headers.getSetCookie()) {
+    let t = c.split(";")[0],
+      i = t.indexOf("=");
+    cookies[t.slice(0, i)] = t.slice(i + 1);
+  }
+  let raw = await r.text();
+  let data;
+  try {
+    data = JSON.parse(raw);
+  } catch {
+    data = raw.slice(0, 300);
+  }
+  return { status: r.status, data };
+}
+const post = (p, b) =>
+  request(p, {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify(b),
+  });
+(async () => {
+  const a = JSON.parse(fs.readFileSync(path.join(q, "account.json")));
+  let csrf = await request("/api/auth/csrf");
+  assert.equal(csrf.status, 200, JSON.stringify(csrf));
+  await request("/api/auth/callback/credentials", {
+    method: "POST",
+    headers: {
+      "Content-Type": "application/x-www-form-urlencoded",
+      "X-Auth-Return-Redirect": "1",
+    },
+    body: new URLSearchParams({
+      csrfToken: csrf.data.csrfToken,
+      email: a.email,
+      password: a.password,
+      callbackUrl: base + "/video-canvas",
+    }),
+  });
+  let session = await request("/api/auth/session");
+  assert.equal(session.data.user?.id, a.id, "real login");
+  fs.writeFileSync(path.join(q, "cookies.json"), JSON.stringify(cookies), { mode: 0o600 });
+  console.log("LOGIN_OK");
+  let cap = await request("/api/canvas/smart-edit/capabilities");
+  assert.equal(cap.data.capabilities.asr, true);
+  assert.equal(cap.data.capabilities.separation, true);
+  console.log("CAPABILITIES_OK");
+  let project = await post("/api/canvas", {
+    title: "智能剪辑本地全链路验收0918",
+    workspaceType: "video",
+  });
+  assert.equal(project.status, 200, JSON.stringify(project));
+  let id = project.data.data.id;
+  assert(
+    Math.abs(Date.now() - new Date(project.data.data.createdAt).getTime()) < 60000,
+    "new canvas timestamp must agree with task timestamps",
+  );
+  fs.writeFileSync(path.join(q, "project-id.txt"), id);
+  let form = new FormData();
+  form.append(
+    "file",
+    new Blob([fs.readFileSync(path.join(q, "speech.mp4"))], { type: "video/mp4" }),
+    "测试中文视频.mp4",
+  );
+  let uploaded = await request("/api/canvas/media/upload", { method: "POST", body: form });
+  assert.equal(uploaded.status, 200, JSON.stringify(uploaded));
+  let asset = {
+    id: "qa-video",
+    name: "中文实测视频",
+    url: uploaded.data.data.url,
+    subtitles: [],
+  };
+  console.log("UPLOAD_OK");
+  const node = {
+    id: "qa-edit",
+    title: "智能剪辑全链路验收",
+    x: 200,
+    y: 200,
+    width: 440,
+    height: 350,
+    kind: "text",
+    inputType: "text",
+    draftValues: { action: "SMART_EDIT_WORKFLOW", nodeType: "smart-edit" },
+  };
+  async function task(params, target = node) {
+    let req = {
+      requestId: crypto.randomUUID(),
+      projectId: id,
+      kind: "smart-edit",
+      node: target,
+      params,
+      edges: [],
+    };
+    let start = Date.now();
+    let r = await post("/api/canvas/generation-tasks", req);
+    assert.equal(r.status, 202, JSON.stringify(r));
+    let repeat = await post("/api/canvas/generation-tasks", req);
+    assert.equal(repeat.data.task.id, r.data.task.id, "idempotent task");
+    while (Date.now() - start < 900000) {
+      await new Promise((r) => setTimeout(r, 1500));
+      let all = await request("/api/canvas/generation-tasks?projectId=" + id);
+      assert.equal(all.status, 200);
+      let t = all.data.tasks.find((t) => t.requestId === req.requestId);
+      if (t && ["succeeded", "failed", "interrupted"].includes(t.status)) {
+        console.log(
+          params.operation,
+          t.status,
+          t.error || "",
+          ((Date.now() - start) / 1000).toFixed(1) + "s",
+        );
+        report.push({
+          operation: params.operation,
+          status: t.status,
+          seconds: (Date.now() - start) / 1000,
+        });
+        assert.equal(t.status, "succeeded", t.error);
+        fs.writeFileSync(
+          path.join(q, params.operation + "-task.json"),
+          JSON.stringify(t, null, 2),
+        );
+        return t.result;
+      }
+    }
+    throw Error("timeout " + params.operation);
+  }
+  let result = await task({
+    operation: "plan",
+    mode: "sequence",
+    assets: [asset],
+    targetDuration: 300,
+    semantic: false,
+    resolution: "480p",
+  });
+  let plan = result.plans[0];
+  result = await task({ operation: "separate", plan, clipId: plan.clips[0].id });
+  plan = result.plans[0];
+  assert.equal(plan.audioClips.length, 2);
+  assert.equal(plan.clips[0].audioDetached, true);
+  assert.equal(plan.assets.length, 3);
+  result = await task({ operation: "transcribe", plan });
+  plan = result.plans[0];
+  assert(plan.captions.some((s) => s.text.includes("智能剪辑")));
+  result = await task({
+    operation: "separate",
+    plan,
+    clipId: plan.audioClips[0].id,
+    track: "audio",
+  });
+  plan = result.plans[0];
+  assert.equal(plan.audioClips.length, 3, "independent audio separation");
+  for (const keep of ["vocals", "background"]) {
+    const processed = await task({
+      operation: "separate",
+      plan,
+      clipId: plan.audioClips[0].id,
+      track: "audio",
+      keep,
+    });
+    const next = processed.plans[0];
+    assert.equal(next.audioClips.length, plan.audioClips.length);
+    const added = next.audioClips.filter(
+      (c) => !plan.audioClips.some((old) => old.id === c.id),
+    );
+    assert.equal(added.length, 1);
+    assert(
+      next.assets
+        .find((a) => a.id === added[0].assetId)
+        .name.endsWith(keep === "vocals" ? "人声" : "背景声"),
+    );
+  }
+  plan.captions[0].style = { ...plan.captions[0].style, fontFamily: "catalog-0006-font" };
+  let render = await task(
+    { operation: "render", plan, sourceNodeId: "qa-edit" },
+    { ...node, id: "qa-render", title: "本地验收成片", kind: "media", inputType: "video" },
+  );
+  let media = await fetch(render.url);
+  assert.equal(media.status, 200);
+  fs.writeFileSync(path.join(q, "main-result.mp4"), Buffer.from(await media.arrayBuffer()));
+  let srt = await fetch(render.subtitleUrl);
+  assert.equal(srt.status, 200);
+  assert((await srt.text()).includes("智能剪辑"));
+  fs.writeFileSync(
+    path.join(q, "main-result.json"),
+    JSON.stringify({ projectId: id, plan, render, report }, null, 2),
+  );
+  console.log("MAIN_CHAIN_OK", JSON.stringify(report));
+})().catch((e) => {
+  console.error(e);
+  process.exitCode = 1;
+});

scripts/qa/smart-edit-models.py

@@ -0,0 +1,30 @@
+"""Real local ASR + separation regression. Pass an actual Chinese speech WAV fixture.
+Usage: SMART_EDIT_*_MODEL=... python scripts/qa/smart-edit-models.py speech.wav output-dir
+No provider calls; the input and all derived files stay local.
+"""
+import json,sys,subprocess,os
+from pathlib import Path
+import numpy as np
+import soundfile as sf
+sys.path.insert(0,str(Path('services/smart-edit').resolve()))
+from analyze import analyze
+from separate import separate
+source=Path(sys.argv[1]).resolve();root=Path(sys.argv[2]).resolve();root.mkdir(parents=True,exist_ok=True)
+subprocess.run(['ffmpeg','-v','error','-y','-i',str(source),'-ac','1','-ar','16000',str(root/'speech.wav')],check=True)
+result=analyze({'audio':str(root/'speech.wav'),'transcribe':True})
+text=''.join(s['text'] for s in result['subtitles']);assert '智能剪辑' in text and '人声分离' in text,text
+assert len(result['subtitles'])>=2,'Sentence timestamps must be enabled'
+assert all(0<=s['start']<s['end'] for s in result['subtitles'])
+subprocess.run(['ffmpeg','-v','error','-y','-i',str(source),'-ac','2','-ar','44100',str(root/'clean.wav')],check=True)
+clean,sr=sf.read(root/'clean.wav',dtype='float32',always_2d=True)
+t=np.arange(len(clean))/sr
+music=(.06*np.sin(2*np.pi*220*t)+.04*np.sin(2*np.pi*330*t))*(.6+.4*np.sin(2*np.pi*2*t)**2)
+mix=clean+music[:,None];sf.write(root/'mixed.wav',mix,sr,subtype='FLOAT')
+separate(str(root/'mixed.wav'),str(root))
+v,_=sf.read(root/'vocals.wav',dtype='float32',always_2d=True);b,_=sf.read(root/'background.wav',dtype='float32',always_2d=True)
+assert v.shape==b.shape==mix.shape
+error=float(np.max(np.abs(v+b-mix)));assert error<1e-6,error
+assert float(np.sqrt(np.mean(v*v)))>.005 and float(np.sqrt(np.mean(b*b)))>.005
+assert not np.allclose(v,mix),'Separation must not return the original track'
+report={'subtitles':result['subtitles'],'duration':len(clean)/sr,'stemReconstructionMaxError':error,'vocalRms':float(np.sqrt(np.mean(v*v))),'backgroundRms':float(np.sqrt(np.mean(b*b)))}
+(root/'report.json').write_text(json.dumps(report,ensure_ascii=False,indent=2));print(json.dumps(report,ensure_ascii=False))

scripts/qa/video-canvas-smart-edit.cjs

@@ -50,7 +50,10 @@ const fs = require("node:fs"),
     ),
   );
   await esbuild.build({
-    entryPoints: ["lib/video-canvas/smart-edit/render.ts", "lib/video-canvas/smart-edit/local.ts"],
+    entryPoints: [
+      "lib/video-canvas/smart-edit/render.ts",
+      "lib/video-canvas/smart-edit/local.ts",
+    ],
     outdir: ".qa/smart-edit/build",
     bundle: true,
     platform: "node",
@@ -86,7 +89,15 @@ const fs = require("node:fs"),
     subtitles: false,
     warnings: [],
     clips: [
-      { id: "1", assetId: "a", sourceIn: 0, sourceOut: 2000000, speed: 1, gain: 1, label: "red" },
+      {
+        id: "1",
+        assetId: "a",
+        sourceIn: 0,
+        sourceOut: 2000000,
+        speed: 1,
+        gain: 1,
+        label: "red",
+      },
       {
         id: "2",
         assetId: "b",
@@ -128,6 +139,468 @@ const fs = require("node:fs"),
         probe.duration / 1e6,
     );
   }
+  for (const type of ["fade", "fadeblack", "fadewhite"]) {
+    const out = path.join(dir, "manual-" + type);
+    fs.mkdirSync(out, { recursive: true });
+    const mixed = {
+      ...plan,
+      subtitles: true,
+      clips: Array.from({ length: 4 }, (_, i) => ({
+        ...plan.clips[0],
+        id: "mixed-" + i,
+        sourceIn: (i % 3) * 1000000,
+        sourceOut: ((i % 3) + 1) * 1000000,
+        transition: { type, duration: i === 2 ? 0 : 0.3 },
+      })),
+      captions: [
+        {
+          id: "manual",
+          start: 0,
+          end: 2000000,
+          text: "手动字幕 测试",
+          style: {
+            fontSize: 36,
+            color: "#ffff00",
+            outlineColor: "#000000",
+            outlineWidth: 2,
+            opacity: 1,
+            spacing: 2,
+          },
+        },
+      ],
+    };
+    const result = await renderEditPlan(mixed, files, undefined, out);
+    const probe = await probeMedia(result.file);
+    assert.ok(Math.abs(probe.duration / 1e6 - 3.4) < 0.15, JSON.stringify(probe));
+    assert.match(
+      fs.readFileSync(path.join(out, "subtitles.ass"), "utf8"),
+      /Noto Sans CJK SC,36/,
+    );
+    assert.match(fs.readFileSync(result.srt, "utf8"), /手动字幕/);
+    const pixel = execFileSync("ffmpeg", [
+      "-v",
+      "error",
+      "-ss",
+      "0.833333",
+      "-i",
+      result.file,
+      "-frames:v",
+      "1",
+      "-vf",
+      "crop=2:2:400:200",
+      "-f",
+      "rawvideo",
+      "-pix_fmt",
+      "rgb24",
+      "-",
+    ]);
+    if (type === "fadeblack")
+      assert.ok(pixel[0] < 70, "black transition must darken the midpoint");
+    if (type === "fadewhite")
+      assert.ok(pixel[1] > 180, "white transition must brighten the midpoint");
+    if (type === "fade")
+      assert.ok(pixel[0] > 220 && pixel[1] < 20, "crossfade must preserve equal red sources");
+    console.log(
+      "PASS: real " +
+        type +
+        " → hard cut → " +
+        type +
+        ", styled manual caption, duration " +
+        probe.duration / 1e6,
+    );
+  }
+  for (const ratio of ["21:9", "16:9", "9:16", "1:1", "4:3", "3:4"]) {
+    const out = path.join(dir, "ratio-" + ratio.replace(":", "-"));
+    fs.mkdirSync(out, { recursive: true });
+    const result = await renderEditPlan(
+      { ...plan, ratio, resolution: "480p", clips: [{ ...plan.clips[0], sourceOut: 1000000 }] },
+      files,
+      undefined,
+      out,
+    );
+    const probe = await probeMedia(result.file);
+    const [w, h] = ratio.split(":").map(Number);
+    assert.equal(Math.min(probe.width, probe.height), 480);
+    assert.ok(Math.abs(probe.width / probe.height - w / h) < 0.005, JSON.stringify(probe));
+    assert.equal(probe.width % 2, 0);
+    assert.equal(probe.height % 2, 0);
+    console.log("PASS: real 480P " + ratio + " " + probe.width + "×" + probe.height);
+  }
+  for (const type of ["fade", "fadeblack", "fadewhite"]) {
+    for (const duration of [2.2, 4]) {
+      const out = path.join(dir, `center-${type}-${duration}`);
+      fs.mkdirSync(out, { recursive: true });
+      const centered = {
+        ...plan,
+        resolution: "480p",
+        clips: [
+          { ...plan.clips[0], sourceOut: 2e6 },
+          {
+            ...plan.clips[0],
+            id: "incoming",
+            sourceIn: 2e6,
+            sourceOut: 4e6,
+            transition: { type, duration, alignment: "center" },
+          },
+        ],
+      };
+      const result = await renderEditPlan(centered, files, undefined, out),
+        probe = await probeMedia(result.file);
+      assert.ok(
+        Math.abs(probe.duration - 4e6) < 150000,
+        `centered transition must keep 4 seconds: ${JSON.stringify(probe)}`,
+      );
+      const pixel = execFileSync("ffmpeg", [
+        "-v",
+        "error",
+        "-ss",
+        "2",
+        "-i",
+        result.file,
+        "-frames:v",
+        "1",
+        "-vf",
+        "crop=2:2:400:200",
+        "-f",
+        "rawvideo",
+        "-pix_fmt",
+        "rgb24",
+        "-",
+      ]);
+      if (type === "fadeblack")
+        assert.ok(pixel[0] < 30 && pixel[2] < 30, `black midpoint ${pixel.slice(0, 3)}`);
+      if (type === "fadewhite")
+        assert.ok(
+          pixel[0] > 220 && pixel[1] > 220 && pixel[2] > 220,
+          `white midpoint ${pixel.slice(0, 3)}`,
+        );
+      if (type === "fade")
+        assert.ok(pixel[0] > 80 && pixel[2] > 80, `crossfade midpoint ${pixel.slice(0, 3)}`);
+      console.log(
+        `PASS: centered ${type} ${duration}s preserves 4s edit; midpoint pixel verified`,
+      );
+    }
+  }
+  // Verify edited geometry, detached audio, gain, fades and black tail in actual encoded media.
+  const sine = path.join(dir, "steady-tone.mp4");
+  execFileSync("ffmpeg", [
+    "-v",
+    "error",
+    "-y",
+    "-f",
+    "lavfi",
+    "-i",
+    "color=c=red:s=320x180:r=30:d=4",
+    "-f",
+    "lavfi",
+    "-i",
+    "sine=frequency=440:sample_rate=48000:duration=4",
+    "-c:v",
+    "libx264",
+    "-threads",
+    "2",
+    "-c:a",
+    "aac",
+    sine,
+  ]);
+  const tone = { ...b, id: "tone", subtitles: [], ...(await probeMedia(sine)) };
+  const extensionFiles = { tone: sine };
+  const base = {
+    ...plan,
+    resolution: "480p",
+    assets: [tone],
+    clips: [
+      {
+        ...plan.clips[0],
+        assetId: "tone",
+        sourceOut: 2e6,
+      },
+    ],
+  };
+  const actual = async (name, edited) => {
+    const out = path.join(dir, name);
+    fs.mkdirSync(out, { recursive: true });
+    return renderEditPlan(edited, extensionFiles, undefined, out);
+  };
+  const pixelAt = (file, time, x, y) =>
+    execFileSync("ffmpeg", [
+      "-v",
+      "error",
+      "-ss",
+      String(time),
+      "-i",
+      file,
+      "-frames:v",
+      "1",
+      "-vf",
+      `crop=2:2:${x}:${y}`,
+      "-f",
+      "rawvideo",
+      "-pix_fmt",
+      "rgb24",
+      "-",
+    ]);
+  const rms = (file, time, duration = 0.2) => {
+    const data = execFileSync("ffmpeg", [
+      "-v",
+      "error",
+      "-ss",
+      String(time),
+      "-i",
+      file,
+      "-t",
+      String(duration),
+      "-vn",
+      "-ac",
+      "1",
+      "-ar",
+      "48000",
+      "-f",
+      "f32le",
+      "-",
+    ]);
+    let sum = 0;
+    for (let i = 0; i < data.length; i += 4) sum += data.readFloatLE(i) ** 2;
+    return Math.sqrt(sum / (data.length / 4));
+  };
+  const baseline = await actual("audio-baseline", base);
+  const detached = await actual("audio-detached", {
+    ...base,
+    clips: [
+      { ...base.clips[0], audioDetached: true, transform: { scale: 0.5, x: 0.2, y: -0.1 } },
+    ],
+    audioClips: [
+      {
+        id: "track",
+        assetId: "tone",
+        start: 0,
+        sourceIn: 0,
+        sourceOut: 2e6,
+        speed: 1,
+        gain: 1,
+        fadeIn: 0,
+        fadeOut: 0,
+      },
+    ],
+  });
+  const ratio = rms(detached.file, 0.8) / rms(baseline.file, 0.8);
+  assert.ok(ratio > 0.9 && ratio < 1.1, `detached sound must not double: ${ratio}`);
+  assert.ok(
+    pixelAt(detached.file, 0.8, 600, 150)[0] > 220,
+    "shifted scaled video should be red",
+  );
+  assert.ok(
+    pixelAt(detached.file, 0.8, 150, 150)[0] < 10,
+    "outside transformed video should be black",
+  );
+  const delayedPlan = {
+    ...base,
+    clips: [{ ...base.clips[0], audioDetached: true }],
+    audioClips: [
+      {
+        id: "delayed",
+        assetId: "tone",
+        start: 1e6,
+        sourceIn: 0,
+        sourceOut: 4e6,
+        speed: 2,
+        gain: 0.5,
+        fadeIn: 0.5,
+        fadeOut: 0.5,
+      },
+    ],
+  };
+  const delayed = await actual("audio-delayed-tail", delayedPlan);
+  const delayedProbe = await probeMedia(delayed.file);
+  assert.ok(Math.abs(delayedProbe.duration - 3e6) < 150000, JSON.stringify(delayedProbe));
+  assert.ok(rms(delayed.file, 0.4) < 0.0001, "detached original must stay silent");
+  const middleRms = rms(delayed.file, 1.7);
+  assert.ok(
+    middleRms > rms(baseline.file, 0.8) * 0.4 && middleRms < rms(baseline.file, 0.8) * 0.6,
+    "independent gain should be applied once",
+  );
+  assert.ok(rms(delayed.file, 1.05, 0.1) < middleRms * 0.4, "audio fade in");
+  assert.ok(rms(delayed.file, 2.8, 0.1) < middleRms * 0.5, "audio fade out");
+  assert.ok(pixelAt(delayed.file, 2.6, 200, 200)[0] < 10, "audio beyond video uses black tail");
+  const withCenter = {
+    ...base,
+    clips: [
+      base.clips[0],
+      {
+        ...base.clips[0],
+        id: "center-audio",
+        sourceIn: 2e6,
+        sourceOut: 4e6,
+        transition: { type: "fadeblack", duration: 2.2, alignment: "center" },
+      },
+    ],
+  };
+  const centerSound = await actual("center-original-audio", withCenter);
+  const centerDetached = await actual("center-detached-audio", {
+    ...withCenter,
+    clips: [withCenter.clips[0], { ...withCenter.clips[1], audioDetached: true }],
+    audioClips: [
+      {
+        id: "center-detached",
+        assetId: "tone",
+        start: 2e6,
+        sourceIn: 2e6,
+        sourceOut: 4e6,
+        speed: 1,
+        gain: 1,
+        fadeIn: 1.1,
+        fadeOut: 0,
+        fadeInStart: 0.5,
+      },
+    ],
+  });
+  for (const time of [1.3, 2.1, 2.6, 3.5]) {
+    const ratio = rms(centerDetached.file, time, 0.1) / rms(centerSound.file, time, 0.1);
+    assert.ok(
+      ratio > 0.85 && ratio < 1.15,
+      `detaching centered-transition audio must preserve its envelope at ${time}s: ${ratio}`,
+    );
+  }
+  console.log("PASS: centered transition audio envelope preserved after detachment");
+  const captioned = await actual("caption-position", {
+    ...base,
+    subtitles: true,
+    captions: [
+      {
+        id: "pos",
+        text: "位置测试",
+        start: 0,
+        end: 2e6,
+        style: {
+          fontSize: 80,
+          color: "#ffffff",
+          outlineColor: "#000000",
+          outlineWidth: 8,
+          outlineOpacity: 0,
+          opacity: 1,
+          spacing: 0,
+          x: 0.25,
+          y: 0.5,
+        },
+      },
+    ],
+  });
+  const frame = execFileSync(
+    "ffmpeg",
+    [
+      "-v",
+      "error",
+      "-ss",
+      "0.8",
+      "-i",
+      captioned.file,
+      "-frames:v",
+      "1",
+      "-f",
+      "rawvideo",
+      "-pix_fmt",
+      "rgb24",
+      "-",
+    ],
+    { maxBuffer: 4 * 1024 * 1024 },
+  );
+  let whitePixels = 0,
+    blackPixels = 0;
+  for (let y = 170; y < 240; y++)
+    for (let x = 90; x < 340; x++) {
+      const offset = (y * 854 + x) * 3;
+      if (frame[offset] > 200 && frame[offset + 1] > 200 && frame[offset + 2] > 200)
+        whitePixels++;
+      if (frame[offset] < 30 && frame[offset + 1] < 30 && frame[offset + 2] < 30) blackPixels++;
+    }
+  assert.ok(whitePixels > 150, `caption must render at moved position: ${whitePixels}`);
+  assert.ok(
+    blackPixels < 10,
+    `zero outline opacity must remove only the black outline: ${blackPixels}`,
+  );
+  console.log(
+    "PASS: real transform pixels, caption position, detached audio without doubling, 2x audio, gain, fades, delay and black tail",
+  );
+  const boosted = await actual("positive-gain", {
+    ...base,
+    clips: [{ ...base.clips[0], gain: 2 }],
+  });
+  const boostRatio = rms(boosted.file, 0.8) / rms(baseline.file, 0.8);
+  assert.ok(
+    boostRatio > 1.85 && boostRatio < 2.15,
+    `+6 dB should double amplitude: ${boostRatio}`,
+  );
+  const muted = await actual("muted-gain", {
+    ...base,
+    clips: [{ ...base.clips[0], gain: 2, muted: true }],
+  });
+  assert.ok(
+    rms(muted.file, 0.8) < 0.0001,
+    "mute must silence original sound without overwriting gain",
+  );
+  const independentMuted = await actual("muted-independent-gain", {
+    ...delayedPlan,
+    audioClips: delayedPlan.audioClips.map((clip) => ({ ...clip, muted: true })),
+  });
+  assert.ok(rms(independentMuted.file, 1.7) < 0.0001, "mute must silence detached sound");
+  const fontHashes = new Set();
+  for (const fontFamily of ["source-han", "noto-sans-sc", "noto-serif-sc", "roboto", "inter", ...(process.env.SMART_EDIT_TEST_CATALOG_FONTS ? ["catalog-0006-font", "catalog-0007-abeezee", "catalog-0014-abril-fatface", "catalog-0365-cjgdgmh", "catalog-1021-molle", "catalog-1685-scjssht"] : [])]) {
+    const output = await actual("font-" + fontFamily, {
+      ...base,
+      subtitles: true,
+      captions: [
+        {
+          id: "font",
+          text: "字幕字体 AaGg 123",
+          start: 0,
+          end: 2e6,
+          style: {
+            fontFamily,
+            referenceHeight: 1080,
+            fontSize: 56,
+            color: "#ffffff",
+            outlineColor: "#ffffff",
+            outlineWidth: 0,
+            opacity: 1,
+            spacing: 0,
+            x: 0.5,
+            y: 0.5,
+          },
+        },
+      ],
+    });
+    const pixels = execFileSync(
+      "ffmpeg",
+      [
+        "-v",
+        "error",
+        "-ss",
+        "0.8",
+        "-i",
+        output.file,
+        "-frames:v",
+        "1",
+        "-f",
+        "rawvideo",
+        "-pix_fmt",
+        "rgb24",
+        "-",
+      ],
+      { maxBuffer: 4 * 1024 * 1024 },
+    );
+    let light = 0;
+    for (let i = 0; i < pixels.length; i += 3)
+      if (pixels[i] > 200 && pixels[i + 1] > 200 && pixels[i + 2] > 200) light++;
+    assert.ok(light > 150, `${fontFamily} must draw glyphs: ${light}`);
+    fontHashes.add(require("node:crypto").createHash("sha256").update(pixels).digest("hex"));
+  }
+  assert.ok(
+    fontHashes.size >= 4,
+    "selected fonts must change actual glyph pixels, not all fall back to one face",
+  );
+  console.log(
+    `PASS: positive gain amplitude, original/detached mute, five packaged fonts${process.env.SMART_EDIT_TEST_CATALOG_FONTS ? " and six catalog fonts" : ""}`,
+  );
   if (process.env.SMART_EDIT_PYTHON) {
     const prepared = await prepareAnalysis(files.b, "fixture", dir, true);
     const analysis = await pythonAnalysis(

scripts/qa/video-render-service.cjs

@@ -28,6 +28,25 @@ const sleep = (ms) => new Promise((r) => setTimeout(r, ms));
   }
   plan.transition = 0.5;
   plan.subtitles = true;
+  plan.clips[1].gain = 2;
+  plan.captions = [
+    {
+      id: "font-http",
+      start: 0,
+      end: 2e6,
+      text: "API 字体验证",
+      style: {
+        fontFamily: "noto-serif-sc",
+        referenceHeight: 1080,
+        fontSize: 56,
+        color: "#ffffff",
+        outlineColor: "#ffffff",
+        outlineWidth: 0,
+        opacity: 1,
+        spacing: 0,
+      },
+    },
+  ];
   let child;
   const start = async () => {
     child = spawn(process.execPath, ["services/video-render/dist/server.cjs"], {
@@ -45,7 +64,9 @@ const sleep = (ms) => new Promise((r) => setTimeout(r, ms));
     child.stderr.on("data", (b) => fs.appendFileSync(path.join(data, "log"), b));
     for (let i = 0; i < 60; i++) {
       try {
-        if ((await fetch(base + "/health", { headers: { authorization: "Bearer " + token } })).ok)
+        if (
+          (await fetch(base + "/health", { headers: { authorization: "Bearer " + token } })).ok
+        )
           return;
       } catch {}
       await sleep(100);
@@ -103,7 +124,15 @@ const sleep = (ms) => new Promise((r) => setTimeout(r, ms));
     const out = path.join(data, "download.mp4");
     fs.writeFileSync(out, Buffer.from(await result.arrayBuffer()));
     const probe = JSON.parse(
-      execFileSync("ffprobe", ["-v", "error", "-show_format", "-show_streams", "-of", "json", out]),
+      execFileSync("ffprobe", [
+        "-v",
+        "error",
+        "-show_format",
+        "-show_streams",
+        "-of",
+        "json",
+        out,
+      ]),
     );
     assert.ok(Math.abs(Number(probe.format.duration) - 3.5) < 0.15);
     assert.equal(probe.streams[0].width, 1280);

scripts/qa/video-upscale-service.cjs

@@ -0,0 +1,103 @@
+// Real authenticated GPU API smoke test. Inputs must be the documented synthetic one-second clips.
+const fs = require("node:fs"),
+  crypto = require("node:crypto"),
+  assert = require("node:assert/strict");
+const { execFileSync } = require("node:child_process");
+const base = process.env.VIDEO_UPSCALE_QA_URL,
+  token = process.env.VIDEO_UPSCALE_QA_TOKEN,
+  source = process.env.VIDEO_UPSCALE_QA_SOURCE;
+if (!base || !token || !source)
+  throw Error(
+    "Set VIDEO_UPSCALE_QA_URL, VIDEO_UPSCALE_QA_TOKEN and VIDEO_UPSCALE_QA_SOURCE"
+  );
+const sha = (s) => crypto.createHash("sha256").update(s).digest("hex"),
+  headers = {
+    authorization: "Bearer " + token,
+    "content-type": "application/json",
+  };
+const wait = (ms) => new Promise((resolve) => setTimeout(resolve, ms));
+(async () => {
+  const request = async (route, body) => {
+    const r = await fetch(base + route, {
+      method: body ? "POST" : "GET",
+      headers,
+      body: body && JSON.stringify(body),
+      signal: AbortSignal.timeout(15000),
+    });
+    return { status: r.status, body: await r.json() };
+  };
+  assert.equal((await fetch(base + "/health")).status, 401);
+  const input = {
+    id: sha("upscale-qa-" + crypto.randomUUID()),
+    tenant: sha("isolated-qa"),
+    plan: { url: source, resolution: "4K" },
+  };
+  assert.equal((await request("/v1/upscales", input)).status, 202);
+  assert.equal((await request("/v1/upscales", input)).status, 200);
+  assert.equal(
+    (
+      await request("/v1/upscales", {
+        ...input,
+        plan: { ...input.plan, resolution: "2K" },
+      })
+    ).status,
+    409
+  );
+  const host = process.env.VIDEO_UPSCALE_QA_SSH_HOST,
+    control = process.env.VIDEO_UPSCALE_QA_SSH_CONTROL;
+  if (host) {
+    for (let i = 0; i < 30; i++) {
+      if ((await request("/v1/upscales/" + input.id)).body.status === "running")
+        break;
+      await wait(100);
+    }
+    execFileSync(
+      "ssh",
+      [
+        ...(control ? ["-S", control] : []),
+        "-o",
+        "BatchMode=yes",
+        host,
+        "systemctl restart video-upscale",
+      ],
+      { timeout: 30000, stdio: "pipe" }
+    );
+    await wait(1500);
+    assert.equal((await request("/v1/upscales", input)).status, 200);
+  }
+  let result;
+  for (let i = 0; i < 240; i++) {
+    result = (await request("/v1/upscales/" + input.id)).body;
+    if (["succeeded", "failed"].includes(result.status)) break;
+    await wait(2000);
+  }
+  assert.equal(result.status, "succeeded", JSON.stringify(result));
+  assert.equal(result.result.width, 3840);
+  assert.equal(result.result.height, 2160);
+  assert.equal(result.result.frames, 2);
+  assert.equal(result.result.audio, true);
+  if (host) assert.equal(result.attempts, 2);
+  const response = await fetch(base + "/v1/upscales/" + input.id + "/video", {
+    headers,
+  });
+  assert(response.ok);
+  fs.mkdirSync(".qa/video-upscale-service", { recursive: true });
+  fs.writeFileSync(
+    ".qa/video-upscale-service/result.mp4",
+    Buffer.from(await response.arrayBuffer())
+  );
+  assert.equal(
+    (await fetch(base + "/v1/upscales/" + input.id + "/video")).status,
+    401
+  );
+  fs.writeFileSync(
+    ".qa/video-upscale-service/report.json",
+    JSON.stringify(result, null, 2)
+  );
+  console.log(
+    "PASS: GPU upscale, authenticated download, idempotency, conflicting parameters, restart recovery, dimensions/frame count/audio"
+  );
+})().catch((e) => {
+  console.error(e);
+  process.exitCode = 1;
+});

services/smart-edit/README.md

@@ -12,7 +12,7 @@
 
 诊断：`$SMART_EDIT_PYTHON services/smart-edit/analyze.py --check`。API `/api/canvas/smart-edit/capabilities` 仅对登录用户返回可用功能与模型列表，不返回密钥或本地路径。ASR 未准备好时页面关闭自动转写选项，仍可导入 SRT 和进行画面混剪。
 
-资源：整个智能剪辑作业通过 PostgreSQL advisory lock 限制为同一数据库下同时一个重处理任务，租约心跳沿用画布任务队列。Python/FFmpeg 限制线程数，单素材 ≤180 秒、≤125 MiB、至多6份，总素材≤10分钟，成片≤90秒、≤12片段；分析缓存按用户隔离，内容 hash 校验，7天过期、4GiB上限。低清代理仅用于分析，成片使用原素材。数据库锁繁忙时任务保存为可恢复状态，页面轮询或返回画布会继续运行。
+资源：整个智能剪辑作业通过 PostgreSQL advisory lock 限制为同一数据库下同时一个重处理任务，租约心跳沿用画布任务队列。Python/FFmpeg 限制线程数，单素材 ≤180 秒、≤125 MiB、至多12份，总原始素材≤10分钟，成片≤300秒、≤12片段；分析缓存按用户隔离，内容 hash 校验，7天过期、4GiB上限。低清代理仅用于分析，成片使用原素材。数据库锁繁忙时任务保存为可恢复状态，页面轮询或返回画布会继续运行。
 
 算法：PySceneDetect 检测镜头，librosa 检测节拍，FunASR 返回语句时间戳（采用 FunClip 的时间戳剪辑思路）。去静音为独立实现的 RMS 阈值算法，参考 Auto-Editor 的粗剪与余量机制；没有复制 Auto-Editor 或 FunClip 源码，也不依赖它们的 UI。来源与许可证记录在项目 docs/video-canvas-smart-edit-research-2026-09-17.md。PySceneDetect 为 BSD-3-Clause，librosa 为 ISC，FunASR 源码和各权重许可应按所安装版本一并保留。
 
@@ -22,3 +22,19 @@
 
 
 远程合成：设置 `SMART_EDIT_RENDER_URL` 与 `SMART_EDIT_RENDER_TOKEN` 后，render操作进入独立API，不占用本地主处理锁。独立服务部署见 `../video-render/README.md`，其千问仅提供可选文字能力。分析操作仍使用本机Python/代理文件。
+
+## 人声分离、转写和完整字体库（2026-09-18）
+
+`bash services/smart-edit/setup.sh --with-all` 在独立环境安装全部依赖。也可使用 `--with-separation` 或 `--with-asr`。
+
+- **人声分离**：`prepare_separation.py` 下载并校验官方 HTDemucs 权重。将输出的 `SMART_EDIT_SEPARATION_MODEL_DIR` 加入应用环境。请求阶段离线运行，CPU 2线程、7秒分块、每次最多180秒；支持仅保留人声、仅保留背景音、输出两个音轨，保持裁剪、变速、音量、静音及转场包络。视频原声自动关闭，独立音轨可继续分离；失败保留原内容。
+- **自动转写**：字幕列表直接提供入口。FunASR 同时开启 `pred_timestamp` 与 `sentence_timestamp`；时间戳限制在素材范围内，再映射至剪辑时间轴。保留手动字幕，重试只替换此前自动生成的字幕。任务存数据库，可退出后再打开；处理完成前锁住当前编辑，避免覆盖并行改动。
+- **字体库**：对应参考站2002项的目录见 `lib/video-canvas/smart-edit/font-catalog.json`（1997项）及 `fonts.ts`（原有5项）。已逐一下载、解析和校验，完整新增字体约757MiB。原始与处理后SHA-256均固定；保留字体内版权/许可字段。浏览器和FFmpeg使用同一处理后文件，独立字体族名避免粗细变体串字，变量字体固定为选定默认实例。两款参考字体的OS/2版本声明与实际长度不一致，按已有字段修复为v4，不改变字形轮廓。
+- **安装字体**：默认按选择懒加载，需Python的`requirements-fonts.txt`。部署前推荐执行 `$SMART_EDIT_PYTHON services/smart-edit/prepare_fonts.py` 预热全部目录，或者复制已校验的字体缓存。`SMART_EDIT_FONT_CACHE_DIR` 指定共享缓存目录；`SMART_EDIT_FONT_PYTHON` 可单独指定字体Python。原有5款字体仍随应用/合成服务打包。首次下载失败会明确报错，不静默替换成默认字体。
+- **远程合成字体**：独立合成服务同样安装字体依赖，设置同一缓存或预热目录；构建时已带上字体转换脚本及固定目录。`SMART_EDIT_FONT_SCRIPT` 可覆盖脚本位置。ASR/分离仍在主应用配置的处理器运行，合成远程执行。
+
+当前限制：视频素材最多12份，每份180秒；成片最长300秒，最多12视频片段、24独立音频片段。含派生分离轨最多36份资源、累计1800秒。实际模型耗时随硬件和素材变化，合成字体的抗锯齿仍受浏览器与libass渲染器影响。
+
+真实模型回归：提供含“智能剪辑”和“人声分离”的中文WAV，运行 `$SMART_EDIT_PYTHON scripts/qa/smart-edit-models.py input.wav .qa/model-check`。验证实际语音识别、逐句时间戳、真实混音分离及两轨重构误差，无付费模型调用。
+
+本地主站全链路：在隔离数据库与存储环境启动真实Next，准备测试账号JSON和中文视频后，执行 `SMART_EDIT_QA_DIR=.qa/isolated-main-chain node scripts/qa/smart-edit-main-chain.cjs`。脚本只允许本地地址，验证登录、上传、方案、分离三个分支、转写、实际字幕字体成片和SRT。完整记录见 `docs/smart-edit-completion-2026-09-18.md`。

services/smart-edit/analyze.py

@@ -14,9 +14,10 @@ for key in ('OMP_NUM_THREADS', 'OPENBLAS_NUM_THREADS', 'MKL_NUM_THREADS', 'NUMBA
 
 def capabilities():
     import shutil
-    return {'scenes': importlib.util.find_spec('scenedetect') is not None,
+    from separate import available
+    return {'separation': available(), 'scenes': importlib.util.find_spec('scenedetect') is not None,
             'beats': importlib.util.find_spec('librosa') is not None,
-            'asr': importlib.util.find_spec('funasr') is not None and all(Path(os.environ.get(k, '/nonexistent')).is_dir() for k in ('SMART_EDIT_ASR_MODEL','SMART_EDIT_VAD_MODEL','SMART_EDIT_PUNC_MODEL')),
+            'asr': importlib.util.find_spec('funasr') is not None and all(all((Path(os.environ.get(k, '/nonexistent'))/f).is_file() for f in ('model.pt','config.yaml')) for k in ('SMART_EDIT_ASR_MODEL','SMART_EDIT_VAD_MODEL','SMART_EDIT_PUNC_MODEL')),
             'ffmpeg': shutil.which('ffmpeg') is not None,
             'ffprobe': shutil.which('ffprobe') is not None}
 
@@ -41,6 +42,8 @@ def analyze(data):
     result={'scenes':[], 'active':[], 'subtitles':[], 'beats':[], 'warnings':[]}
     duration=data.get('duration',0)
     if data.get('video'):
+        import cv2
+        cv2.setNumThreads(2)
         from scenedetect import detect,AdaptiveDetector
         result['scenes']=[[round(a.get_seconds()*1e6),round(b.get_seconds()*1e6)] for a,b in detect(data['video'],AdaptiveDetector(min_scene_len=15),start_in_scene=True)] or [[0,duration]]
     if data.get('audio'):
@@ -61,7 +64,7 @@ def analyze(data):
             from funasr import AutoModel
             torch.set_num_threads(2)
             model=AutoModel(model=os.environ['SMART_EDIT_ASR_MODEL'],vad_model=os.environ['SMART_EDIT_VAD_MODEL'],punc_model=os.environ['SMART_EDIT_PUNC_MODEL'],device='cpu',disable_update=True,trust_remote_code=False)
-            outputs=model.generate(input=data['audio'],sentence_timestamp=True,batch_size_s=30)
+            outputs=model.generate(input=data['audio'],sentence_timestamp=True,pred_timestamp=True,batch_size_s=30)
             for output in outputs:
                 for sentence in output.get('sentence_info',[]):
                     start=sentence.get('start');end=sentence.get('end')
@@ -70,7 +73,10 @@ def analyze(data):
                         if times:start,end=times[0][0],times[-1][1]
                     text=str(sentence.get('text','')).strip()
                     if text and start is not None and end is not None and end>start:
-                        result['subtitles'].append({'start':round(start*1000),'end':round(end*1000),'text':text})
+                        bounded_start=max(0,round(start*1000))
+                        bounded_end=min(round(len(wave)/sr*1e6),round(end*1000))
+                        if bounded_end>bounded_start:
+                            result['subtitles'].append({'start':bounded_start,'end':bounded_end,'text':text})
             if not result['subtitles']:result['warnings'].append('语音识别没有返回有效语句，将按画面镜头处理；不会生成虚构字幕。')
     elif data.get('transcribe'):result['warnings'].append('素材没有音轨，跳过语音转写。')
     return result

services/smart-edit/font.py

@@ -0,0 +1,35 @@
+"""Freeze a font's default variation and give it a unique preview/export family name."""
+import sys
+from fontTools.ttLib import TTFont, newTable
+from fontTools.varLib.instancer import instantiateVariableFont
+
+def prepare(source, destination, family):
+    font = TTFont(source, recalcTimestamp=False)
+    if 'fvar' in font:
+        axes = {axis.axisTag: axis.defaultValue for axis in font['fvar'].axes}
+        for axis in font['fvar'].axes:
+            if axis.axisTag == 'wght':
+                axes['wght'] = max(axis.minValue, min(axis.maxValue, 400))
+        font = instantiateVariableFont(font, axes, inplace=True)
+    for record in font['name'].names:
+        if record.nameID in (1, 2, 3, 4, 6, 16, 17):
+            value = 'Regular' if record.nameID in (2, 17) else family
+            if record.nameID == 6:
+                value = family.replace(' ', '-')[:63]
+            record.string = value.encode(record.getEncoding(), errors='replace')
+    # Two reference fonts declare OS/2 v5 but only contain the 96-byte v4 table.
+    # Correct the version tag without inventing missing metrics or changing outlines.
+    metrics = font.reader['OS/2']
+    if len(metrics) == 96 and int.from_bytes(metrics[:2], 'big') == 5:
+        table = newTable('OS/2')
+        table.decompile(b'\x00\x04' + metrics[2:], font)
+        font['OS/2'] = table
+    # A unique family per selected face prevents fontconfig confusing weight variants.
+    font['OS/2'].fsSelection &= ~((1 << 0) | (1 << 5))
+    font['OS/2'].fsSelection |= 1 << 6
+    font['OS/2'].usWeightClass = 400
+    font['head'].macStyle &= ~3
+    font.save(destination)
+
+if __name__ == '__main__':
+    prepare(*sys.argv[1:4])

services/smart-edit/prepare_fonts.py

@@ -0,0 +1,23 @@
+"""Prewarm the audited font catalogue. Downloads and outputs are checksum verified."""
+import concurrent.futures,hashlib,json,os,tempfile,urllib.request
+from pathlib import Path
+from font import prepare
+root=Path(os.environ.get('SMART_EDIT_FONT_CACHE_DIR','.cache/smart-edit-fonts')).resolve();root.mkdir(parents=True,exist_ok=True)
+catalog=json.loads(Path('lib/video-canvas/smart-edit/font-catalog.json').read_text())
+def digest(file):return hashlib.sha256(file.read_bytes()).hexdigest()
+def install(font):
+    output=root/font['file']
+    if output.is_file() and digest(output)==font['preparedSha256']:return
+    with tempfile.TemporaryDirectory(dir=root) as work:
+        raw=Path(work)/'source';prepared=Path(work)/'prepared.ttf'
+        with urllib.request.urlopen(font['url'],timeout=90) as response:
+            data=response.read(40*1024*1024+1)
+        if len(data)>40*1024*1024:raise ValueError('字体超过大小上限')
+        raw.write_bytes(data)
+        if digest(raw)!=font['sha256']:raise ValueError('原始字体校验失败: '+font['name'])
+        prepare(str(raw),str(prepared),font['family'])
+        if digest(prepared)!=font['preparedSha256']:raise ValueError('字体处理版本不匹配: '+font['name'])
+        prepared.replace(output)
+with concurrent.futures.ThreadPoolExecutor(max_workers=3) as pool:
+    list(pool.map(install,catalog))
+print(json.dumps({'fonts':len(catalog)+5,'catalogueFonts':len(catalog),'cache':str(root)},ensure_ascii=False))

services/smart-edit/prepare_separation.py

@@ -0,0 +1,18 @@
+"""Install only the checksum-pinned official HTDemucs checkpoint, outside request handling."""
+import hashlib,json,os,urllib.request
+from pathlib import Path
+from separate import CHECKSUM
+root=Path(os.environ.get('SMART_EDIT_SEPARATION_MODEL_DIR','.smart-edit-models/demucs')).resolve()
+root.mkdir(parents=True,exist_ok=True)
+file=root/'955717e8-8726e21a.th'
+url='https://dl.fbaipublicfiles.com/demucs/hybrid_transformer/955717e8-8726e21a.th'
+if not file.exists() or hashlib.sha256(file.read_bytes()).hexdigest()!=CHECKSUM:
+    temp=file.with_suffix('.part')
+    try:
+        urllib.request.urlretrieve(url,temp)
+        if hashlib.sha256(temp.read_bytes()).hexdigest()!=CHECKSUM:
+            raise RuntimeError('HTDemucs 权重校验失败')
+        temp.replace(file)
+    finally:
+        temp.unlink(missing_ok=True)
+print(json.dumps({'SMART_EDIT_SEPARATION_MODEL_DIR':str(root)},ensure_ascii=False))

services/smart-edit/requirements-asr.txt

@@ -3,3 +3,5 @@ funasr==1.4.15
 torch==2.6.0
 torchaudio==2.6.0
 modelscope==1.30.0
+# Verified with Paraformer sentence/character timestamps.
+transformers==5.17.0

services/smart-edit/requirements-fonts.txt

@@ -0,0 +1,2 @@
+fonttools==4.59.1
+brotli==1.2.0

services/smart-edit/requirements-separation.txt

@@ -0,0 +1,4 @@
+-r requirements.txt
+torch==2.6.0
+torchaudio==2.6.0
+demucs==4.0.1

services/smart-edit/separate.py

@@ -0,0 +1,59 @@
+"""Offline, bounded HTDemucs inference. Model installation is a separate operator step."""
+import json
+import os
+import sys
+from pathlib import Path
+for key in ('OMP_NUM_THREADS', 'OPENBLAS_NUM_THREADS', 'MKL_NUM_THREADS'):
+    os.environ[key] = '2'
+CHECKSUM = "8726e21a993978c7ba086d3872e7608d7d5bfca646ca4aca459ffda844faa8b4"
+SIGNATURE = '955717e8'
+
+def available():
+    import importlib.util
+    root = Path(os.environ.get('SMART_EDIT_SEPARATION_MODEL_DIR', '/nonexistent'))
+    return importlib.util.find_spec('demucs') is not None and any(root.glob(SIGNATURE + '-*.th'))
+
+def separate(source, destination):
+    if not available():
+        raise RuntimeError('人声分离模型尚未配置，请安装 Demucs 依赖和官方 HTDemucs 权重。')
+    import numpy as np
+    import soundfile as sf
+    import torch
+    from demucs.states import load_model
+    import hashlib
+    from demucs.apply import apply_model
+    torch.set_num_threads(2)
+    torch.manual_seed(0)
+    audio, sr = sf.read(source, dtype='float32', always_2d=True)
+    if sr != 44100 or audio.shape[1] != 2 or not 0 < len(audio) <= sr * 180:
+        raise ValueError('分离输入必须为 180 秒以内的 44.1kHz 双声道音频')
+    if not np.isfinite(audio).all():
+        raise ValueError('音频包含无效采样')
+    checkpoint = next(Path(os.environ['SMART_EDIT_SEPARATION_MODEL_DIR']).glob(SIGNATURE + '-*.th'))
+    if hashlib.sha256(checkpoint.read_bytes()).hexdigest() != CHECKSUM:
+        raise ValueError('人声分离权重校验失败')
+    # Official checkpoint verified above; Demucs 4 uses a pickled model class.
+    model = load_model(torch.load(checkpoint, map_location='cpu', weights_only=False))
+    model.eval()
+    wave = torch.from_numpy(audio.T.copy())
+    reference = wave.mean(0)
+    mean, std = reference.mean(), reference.std()
+    if std < 1e-8:
+        vocals = np.zeros_like(audio)
+    else:
+        with torch.inference_mode():
+            output = apply_model(model, ((wave - mean) / std)[None], device='cpu',
+                                 shifts=1, split=True, overlap=.25, segment=7,
+                                 progress=False, num_workers=0)[0]
+            vocals = (output[model.sources.index('vocals')] * std + mean).T.numpy()
+    # Complementary stems reconstruct the original exactly, with no independent gain normalization.
+    destination = Path(destination)
+    destination.mkdir(parents=True, exist_ok=True)
+    sf.write(destination / 'vocals.wav', vocals, sr, subtype='FLOAT')
+    sf.write(destination / 'background.wav', audio - vocals, sr, subtype='FLOAT')
+
+if __name__ == '__main__':
+    if sys.argv[1:] == ['--check']:
+        print(json.dumps({'separation': available()}))
+    else:
+        separate(*sys.argv[1:3])

services/smart-edit/setup.sh

@@ -4,9 +4,13 @@ set -euo pipefail
 smart_env="${SMART_EDIT_ENV_DIR:-$PWD/.venv-smart-edit}"
 smart_python="${SMART_EDIT_SETUP_PYTHON:-python3.11}"
 "$smart_python" -m venv "$smart_env"
-"$smart_env/bin/python" -m pip install -r services/smart-edit/requirements.txt
+"$smart_env/bin/python" -m pip install -r services/smart-edit/requirements.txt -r services/smart-edit/requirements-fonts.txt
-if [[ "${1:-}" == "--with-asr" ]]; then
+if [[ " $* " == *" --with-asr "* || " $* " == *" --with-all "* ]]; then
   "$smart_env/bin/python" -m pip install -r services/smart-edit/requirements-asr.txt
   "$smart_env/bin/python" services/smart-edit/prepare_models.py
 fi
+if [[ " $* " == *" --with-separation "* || " $* " == *" --with-all "* ]]; then
+  "$smart_env/bin/python" -m pip install -r services/smart-edit/requirements-separation.txt
+  "$smart_env/bin/python" services/smart-edit/prepare_separation.py
+fi
 SMART_EDIT_PYTHON="$smart_env/bin/python" "$smart_env/bin/python" services/smart-edit/analyze.py --check

services/video-render/README.md

@@ -30,7 +30,7 @@
 
 任务 ID 由主项目持久任务号导出：重复 POST 返回同一记录，参数改变返回 409。写盘采用临时文件、fsync、rename 后再响应。单服务进程拥有此数据目录，不能启动多个副本共用目录。服务重启后运行中任务最多恢复两次；完成记录重启后仍可下载。主项目轮询超时保存编号继续查询，不受浏览器页面生命周期限制。
 
-上限：请求 1 MiB，至多 30 个等待/运行任务，至多 200 个保留任务；完成/失败记录和产物保留 7 天，小时清理；原素材缓存按用户隔离，4 GiB / 7 天清理；素材 ≤125 MiB/份、≤180秒/份、≤6份且总长≤10分钟，成片≤90秒/12片段。磁盘可用不足10GiB拒绝重处理。外部素材禁止内网地址，DNS验证后绑定地址下载，每次重定向重新验证；FFmpeg禁用网络输入协议。
+上限：请求 1 MiB，至多 30 个等待/运行任务，至多 200 个保留任务；完成/失败记录和产物保留 7 天，小时清理；原素材缓存按用户隔离，4 GiB / 7 天清理；素材 ≤125 MiB/份、≤180秒/份、≤12份且总长≤10分钟，成片≤300秒/12片段；6种比例（含21:9）、480P/720P/1080P，长边最高2520像素。磁盘可用不足10GiB拒绝重处理。外部素材禁止内网地址，DNS验证后绑定地址下载，每次重定向重新验证；FFmpeg禁用网络输入协议。
 
 ## 验证
 
@@ -39,3 +39,7 @@
 随后 `node services/video-render/build.cjs && PATH=<ffmpeg目录>:$PATH node scripts/qa/video-render-service.cjs` 验证真实 HTTP、鉴权、FFmpeg、进程强制中断/恢复、幂等、冲突拒绝和完成记录重启保留。用合成色块/正弦音，不调用付费模型。
 
 依赖许可：Zod MIT；FFmpeg 许可取决于本机编译选项，Ubuntu libx264 构建含 GPL 组件。服务只运行本机已安装 FFmpeg，不把二进制混入仓库。代码与发布状态记录在 `docs/video-render-service-2026-09-17.md`。
+
+## 字体包
+
+构建会将 `public/fonts/smart-edit` 的真实字体和许可证复制到 `dist/fonts`，发布时必须与 `server.cjs` 一起携带。服务默认读取同目录 `fonts`，可用 `SMART_EDIT_FONTS_DIR` 指定绝对路径；主项目本地合成默认读取 `public/fonts/smart-edit`。缺失所选字体时返回明确错误，不悄悄使用其他字体。五种字体来自 Adobe/Google 官方仓库，来源、摘要和各自 SIL OFL 许可保存在字体目录。

services/video-render/build.cjs

@@ -16,4 +16,8 @@ esbuild
     target: "node22",
     format: "cjs",
   })
+  .then(() => {
+    fs.cpSync("public/fonts/smart-edit", "services/video-render/dist/fonts", { recursive: true });
+    fs.copyFileSync("services/smart-edit/font.py", "services/video-render/dist/font.py");
+  })
   .catch(() => process.exit(1));

services/video-render/server.ts

@@ -23,6 +23,8 @@ import {
 import { renderJob } from "../../lib/video-canvas/smart-edit/render-job";
 import { pruneSmartEditCache } from "../../lib/video-canvas/smart-edit/local";
 
+process.env.SMART_EDIT_FONT_SCRIPT ||= path.join(__dirname, "font.py");
+process.env.SMART_EDIT_FONTS_DIR ||= path.join(__dirname, "fonts");
 const root = path.resolve(process.env.RENDER_DATA_DIR || "/var/lib/video-render");
 const token = process.env.RENDER_API_TOKEN || "";
 if (token.length < 32) throw Error("RENDER_API_TOKEN must have at least 32 characters");
@@ -113,11 +115,17 @@ async function work() {
       job.stage = stage;
       await persist(job);
     });
-    job.result = { width: result.width, height: result.height, durationSeconds: result.duration };
+    job.result = {
+      width: result.width,
+      height: result.height,
+      durationSeconds: result.duration,
+    };
     // Keep only the deliverables. Intermediate files must not accumulate per job.
     for (const name of await readdir(dir))
       if (!["result.mp4", "subtitles.srt"].includes(name))
-        await rm(path.join(dir, name), { force: true });
+        await rm(path.join(dir, name), { recursive: true, force: true }).catch(() => {
+          console.warn(JSON.stringify({ event: "render-cleanup-deferred", id: job.id }));
+        });
     job.status = "succeeded";
     job.stage = "合成完成";
     await persist(job);
@@ -260,7 +268,8 @@ async function main() {
             );
           if (
             jobs.size >= 200 ||
-            [...jobs.values()].filter((j) => ["queued", "running"].includes(j.status)).length >= 30
+            [...jobs.values()].filter((j) => ["queued", "running"].includes(j.status)).length >=
+              30
           )
             return json(res, 429, { error: "合成队列已满，请稍后重试" });
           const job: Job = {
@@ -283,7 +292,9 @@ async function main() {
           accepting = false;
         }
       }
-      const match = url.pathname.match(/^\/v1\/renders\/([a-f0-9]{64})(?:\/(video|subtitles))?$/);
+      const match = url.pathname.match(
+        /^\/v1\/renders\/([a-f0-9]{64})(?:\/(video|subtitles))?$/,
+      );
       if (req.method === "GET" && match) {
         const job = jobs.get(match[1]);
         if (!job) return json(res, 404, { error: "任务不存在或已过期" });
@@ -321,7 +332,10 @@ async function main() {
   });
   server.requestTimeout = 120000;
   server.headersTimeout = 10000;
-  server.listen(Number(process.env.RENDER_PORT || 18340), process.env.RENDER_HOST || "127.0.0.1");
+  server.listen(
+    Number(process.env.RENDER_PORT || 18340),
+    process.env.RENDER_HOST || "127.0.0.1",
+  );
   const fatal = () => {
     console.error("Render queue persistence failure");
     process.exit(1);

services/video-upscale/README.md

@@ -0,0 +1,46 @@
+# 独立视频高清服务
+
+主项目负责用户/项目鉴权、持久任务、归档；此服务只下载视频并在独立 GPU 上逐帧超分。没有 CPU 回退，不使用千问显卡。当前部署固定 `CUDA_VISIBLE_DEVICES=3`，GPU 2 有硬件故障历史，不分配给本服务。
+
+## 模型和环境
+
+使用 [Real-ESRGAN](https://github.com/xinntao/Real-ESRGAN) 的 `realesr-general-x4v3`，官方 SRVGGNetCompact 架构（64 通道、32 卷积层、PReLU、4 倍输出）；源码与 BSD-3-Clause 许可位于 `vendor/`，来源 commit 见 `vendor/SOURCE.md`。不是 Topaz 模型。模型文件不提交仓库。
+
+权重：[官方 v0.2.5.0 发布](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesr-general-x4v3.pth)，SHA-256 `8dc7edb9ac80ccdc30c3a5dca6616509367f05fbc184ad95b731f05bece96292`，推理前强校验。
+
+P40 环境实测：Python 3.12.3、PyTorch 2.5.1+cu121、NumPy 1.26.4、Pillow 10.4.0、系统 FFmpeg/ffprobe（libx264/AAC）。PyTorch 使用 CUDA 12.1 官方 wheel 源；必须固定此版本，不能未经 P40 验证直接升级 CUDA/PyTorch。FP32 推理，192px tile + 40px halo，整帧输出在 CPU 内存，避免全帧网络推理耗尽显存。每个服务同时处理一个任务。
+
+## 部署与 API
+
+构建：`node services/video-upscale/build.cjs`，将 `dist/` 内容放到 `/opt/video-upscale/app/`。环境位于 `/opt/video-upscale/venv/`，权重 `/opt/video-upscale/models/realesr-general-x4v3.pth`。参考 `video-upscale.service` 创建专用 `video-upscale` 用户和其拥有的 `/var/lib/video-upscale`。
+
+`/etc/video-upscale.env` 权限 600，包含 `UPSCALE_API_TOKEN`（至少 32 字符）、`UPSCALE_HOST=127.0.0.1`、`UPSCALE_PORT=18341`、`UPSCALE_DATA_DIR=/var/lib/video-upscale`。禁止将凭据放进 Git 或浏览器。
+
+所有 API 要求 `Authorization: Bearer <token>`：
+
+- `GET /health`：服务、队列和 GPU 配置。
+- `POST /v1/upscales`：`{id,tenant,plan:{url,resolution}}`，id/tenant 均为 SHA-256 标识，resolution 为 `1080P`/`2K`/`4K`。
+- `GET /v1/upscales/:id`：状态、阶段、帧数/帧率/音轨、尺寸与时长。
+- `GET /v1/upscales/:id/video`：鉴权读取结果。
+
+48 服务器使用 `video-upscale-tunnel.service`，复用已有专用隧道账号和固定 known_hosts，只扩展该公钥的 `permitlisten` 到 `127.0.0.1:18341`；原 18340 合成隧道不变。主项目接入变量为 `VIDEO_UPSCALE_URL=http://127.0.0.1:18341` 和 `VIDEO_UPSCALE_TOKEN`，部署时从 root-only `/etc/video-upscale-client.env` 注入，不能依赖本机临时 SSH 转发。
+
+任务落盘 fsync + rename；相同 ID/参数幂等，相同 ID/不同参数 409。服务重启保留 running 记录，最多恢复两次；`KillMode=mixed` 先通知 Node 保存关闭状态，再终止子进程，避免把正常重启误记为推理失败。主项目保存远程 ID/服务指纹，轮询超时、读取/归档失败继续查询同一任务。
+
+## 处理边界
+
+输入 ≤125 MiB、≤120 秒、≤60fps、≤4K 像素数；支持 MP4/MOV、Matroska/WebM、AVI、FLV、MPEG-TS。仅恒定帧率；VFR 明确拒绝，不能静默改变音画同步。手机 90/180/270 度旋转信息会归一化，保留帧数、帧率、音轨。音频重编码 AAC，视频 H.264/yuv420p + faststart。不做补帧。
+
+目标以 1920×1080 / 2560×1440 / 3840×2160 为横屏边界，竖屏交换宽高，等比例缩放、偶数尺寸。目标必须大于原图且至多 4 倍；目标已满足或要求超过 4 倍时拒绝，不拿插值假称模型支持更高倍率。网络输出可能产生细节变化，不能保证真实细节还原。
+
+单任务超时 2 小时；≤30 个在途任务，≤200 个保留任务，完成/失败保留 7 天；可用磁盘 <10 GiB 拒绝开始。输入禁止内网/DNS 重绑定/不安全重定向，FFmpeg 网络协议关闭；输出下载 ≤300 MiB。systemd 内存 8 GiB、CPU 4 核、任务/线程 256，滤镜线程固定 1，避免 128 核宿主机默认线程数导致编码失败。
+
+## 验收
+
+`scripts/qa/video-upscale-service.cjs` 验证真实 API、GPU、音轨/帧数/尺寸、重复提交、参数冲突、未鉴权下载及重启恢复。需设置 `VIDEO_UPSCALE_QA_URL`、`VIDEO_UPSCALE_QA_TOKEN`、`VIDEO_UPSCALE_QA_SOURCE`（960×540、2fps、1 秒、含音轨的合成测试片）；可选 `VIDEO_UPSCALE_QA_SSH_HOST` / `VIDEO_UPSCALE_QA_SSH_CONTROL` 执行独立服务重启。只重启视频高清服务，不重启千问或主项目。
+
+测试素材生成示例（应上传到获准访问的测试存储域名）：
+
+```sh
+ffmpeg -f lavfi -i 'testsrc2=size=960x540:rate=2:duration=1' -f lavfi -i 'sine=frequency=440:sample_rate=48000:duration=1' -c:v libx264 -pix_fmt yuv420p -c:a aac -shortest landscape.mp4
+```

services/video-upscale/build.cjs

@@ -0,0 +1,32 @@
+const fs = require("node:fs"),
+  path = require("node:path");
+const esbuild = require(
+  path.resolve(
+    "node_modules/.pnpm",
+    fs
+      .readdirSync("node_modules/.pnpm")
+      .find((n) => n.startsWith("esbuild@0.25.")),
+    "node_modules/esbuild"
+  )
+);
+esbuild
+  .build({
+    entryPoints: ["services/video-upscale/server.ts"],
+    outfile: "services/video-upscale/dist/server.cjs",
+    bundle: true,
+    platform: "node",
+    target: "node22",
+    format: "cjs",
+  })
+  .then(() => {
+    fs.copyFileSync(
+      "services/video-upscale/upscale.py",
+      "services/video-upscale/dist/upscale.py"
+    );
+    fs.cpSync(
+      "services/video-upscale/vendor",
+      "services/video-upscale/dist/vendor",
+      { recursive: true }
+    );
+  })
+  .catch(() => process.exit(1));

services/video-upscale/job.ts

@@ -0,0 +1,85 @@
+import { spawn } from "node:child_process";
+import path from "node:path";
+import { downloadCanvasMedia } from "../../lib/video-canvas/download-media";
+import {
+  videoUpscaleSchema,
+  type VideoUpscaleInput,
+} from "../../lib/video-canvas/video-upscale-contract";
+export async function upscaleJob(
+  value: VideoUpscaleInput,
+  _tenant: string,
+  dir: string,
+  progress: (stage: string) => Promise<void>
+) {
+  const input = videoUpscaleSchema.parse(value);
+  const file = path.join(dir, "input.bin");
+  await downloadCanvasMedia(input.url, file);
+  await progress("GPU 正在进行 AI 超分");
+  return new Promise<any>((resolve, reject) => {
+    const child = spawn(
+      process.env.UPSCALE_PYTHON || "/opt/video-upscale/venv/bin/python",
+      [
+        process.env.UPSCALE_SCRIPT || path.join(__dirname, "upscale.py"),
+        "--input",
+        file,
+        "--output",
+        path.join(dir, "result.mp4"),
+        "--resolution",
+        input.resolution,
+        "--model",
+        process.env.UPSCALE_MODEL ||
+          "/opt/video-upscale/models/realesr-general-x4v3.pth",
+      ],
+      { stdio: ["ignore", "pipe", "pipe"], detached: true }
+    );
+    let buffer = "",
+      result: any,
+      error = "",
+      stages = Promise.resolve();
+    const timer = setTimeout(() => {
+      try {
+        process.kill(-child.pid!, "SIGKILL");
+      } catch {}
+    }, 2 * 3600000);
+    child.stdout.on("data", (chunk) => {
+      buffer += chunk;
+      const lines = buffer.split("\n");
+      buffer = lines.pop()!;
+      for (const line of lines)
+        try {
+          const value = JSON.parse(line);
+          if (value.result) result = value.result;
+          if (value.stage)
+            stages = stages
+              .then(() => progress(value.stage))
+              .catch(() => {
+                try {
+                  process.kill(-child.pid!, "SIGKILL");
+                } catch {}
+              });
+        } catch {}
+    });
+    child.stderr.on("data", (chunk) => {
+      error = (error + chunk).slice(-4000);
+    });
+    child.on("error", (e) => {
+      clearTimeout(timer);
+      reject(e);
+    });
+    child.on("close", async (code) => {
+      clearTimeout(timer);
+      await stages.catch(() => {});
+      if (code !== 0 || !result) {
+        console.error(
+          "Upscale inference failed",
+          error.replace(/https?:\/\/\S+/g, "[url]")
+        );
+        reject(
+          Error(
+            "AI 超分失败，请检查视频是否为恒定帧率、120 秒以内，且放大倍率为 1–4 倍"
+          )
+        );
+      } else resolve({ ...result, duration: result.durationSeconds });
+    });
+  });
+}

services/video-upscale/server.ts

@@ -0,0 +1,328 @@
+import http from "node:http";
+import { createReadStream } from "node:fs";
+import {
+  mkdir,
+  readFile,
+  rename,
+  readdir,
+  rm,
+  stat,
+  statfs,
+  open,
+} from "node:fs/promises";
+import path from "node:path";
+import { createHash, timingSafeEqual } from "node:crypto";
+import { pipeline } from "node:stream/promises";
+import { z } from "zod";
+import {
+  videoUpscaleSchema,
+  type VideoUpscaleInput,
+} from "../../lib/video-canvas/video-upscale-contract";
+import { upscaleJob } from "./job";
+
+const root = path.resolve(
+  process.env.UPSCALE_DATA_DIR || "/var/lib/video-upscale"
+);
+const token = process.env.UPSCALE_API_TOKEN || "";
+if (token.length < 32)
+  throw Error("UPSCALE_API_TOKEN must have at least 32 characters");
+const jobs = new Map<string, Job>();
+type Job = {
+  id: string;
+  digest: string;
+  tenant: string;
+  plan: VideoUpscaleInput;
+  status: "queued" | "running" | "succeeded" | "failed";
+  stage: string;
+  createdAt: string;
+  updatedAt: string;
+  attempts: number;
+  error?: string;
+  result?: Record<string, any>;
+};
+const inputSchema = z.object({
+  id: z.string().regex(/^[a-f0-9]{64}$/),
+  tenant: z.string().regex(/^[a-f0-9]{64}$/),
+  plan: videoUpscaleSchema,
+});
+let busy = false,
+  stopping = false,
+  accepting = false;
+const file = (id: string) => path.join(root, id + ".json");
+async function persist(job: Job) {
+  job.updatedAt = new Date().toISOString();
+  const temp = file(job.id) + ".tmp";
+  const handle = await open(temp, "w", 0o600);
+  try {
+    await handle.writeFile(JSON.stringify(job));
+    await handle.sync();
+  } finally {
+    await handle.close();
+  }
+  await rename(temp, file(job.id));
+  const directory = await open(root, "r");
+  try {
+    await directory.sync();
+  } finally {
+    await directory.close();
+  }
+  jobs.set(job.id, job);
+}
+function publicJob(job: Job) {
+  return {
+    id: job.id,
+    status: job.status,
+    stage: job.stage,
+    attempts: job.attempts,
+    createdAt: job.createdAt,
+    updatedAt: job.updatedAt,
+    error: job.error,
+    result: job.result,
+  };
+}
+async function prune() {
+  for (const job of jobs.values())
+    if (
+      ["succeeded", "failed"].includes(job.status) &&
+      Date.now() - Date.parse(job.updatedAt) > 7 * 86400000
+    ) {
+      await rm(path.join(root, job.id), { recursive: true, force: true });
+      await rm(file(job.id), { force: true });
+      jobs.delete(job.id);
+    }
+}
+async function work() {
+  if (busy || stopping) return;
+  const job = [...jobs.values()].find((j) => j.status === "queued");
+  if (!job) return;
+  busy = true;
+  const dir = path.join(root, job.id);
+  try {
+    await prune();
+    const disk = await statfs(root);
+    if (disk.bavail * disk.bsize < 10 * 1024 ** 3)
+      throw Error("超分磁盘剩余空间不足 10 GiB");
+    job.status = "running";
+    job.attempts++;
+    job.stage = "准备素材";
+    await persist(job);
+    await rm(dir, { recursive: true, force: true });
+    await mkdir(dir, { mode: 0o700 });
+    const result = await upscaleJob(
+      job.plan,
+      job.tenant,
+      dir,
+      async (stage) => {
+        job.stage = stage;
+        await persist(job);
+      }
+    );
+    job.result = result;
+    // Keep only the deliverables. Intermediate files must not accumulate per job.
+    for (const name of await readdir(dir))
+      if (name !== "result.mp4")
+        await rm(path.join(dir, name), { recursive: true, force: true }).catch(
+          () => {
+            console.warn(
+              JSON.stringify({ event: "upscale-cleanup-deferred", id: job.id })
+            );
+          }
+        );
+    job.status = "succeeded";
+    job.stage = "超分完成";
+    await persist(job);
+  } catch (error) {
+    if (stopping) return; // Preserve running journal; restart will resume this same job.
+    job.status = "failed";
+    job.stage = "超分失败";
+    // Do not expose ffmpeg arguments, private cache paths or signed source URLs.
+    job.error =
+      error instanceof Error &&
+      !/Command failed|https?:|\/var\/|\/tmp\//.test(error.message)
+        ? error.message.slice(0, 300)
+        : "素材处理失败，请检查格式、参数或服务日志";
+    await rm(dir, { recursive: true, force: true }).catch(() => {});
+    await persist(job);
+    console.error(
+      JSON.stringify({ event: "upscale-failed", id: job.id, error: job.error })
+    );
+  } finally {
+    busy = false;
+  }
+  setImmediate(
+    () =>
+      void work().catch((e) => {
+        console.error("Render queue persistence failure");
+        process.exit(1);
+      })
+  );
+}
+function json(res: http.ServerResponse, status: number, value: unknown) {
+  res.writeHead(status, {
+    "content-type": "application/json; charset=utf-8",
+    "cache-control": "no-store",
+  });
+  res.end(JSON.stringify(value));
+}
+async function body(req: http.IncomingMessage) {
+  let size = 0;
+  const chunks: Buffer[] = [];
+  for await (const chunk of req) {
+    size += chunk.length;
+    if (size > 1024 * 1024) throw Error("请求超过 1 MiB");
+    chunks.push(chunk);
+  }
+  return JSON.parse(Buffer.concat(chunks).toString("utf8"));
+}
+function authenticated(req: http.IncomingMessage) {
+  const expected = Buffer.from("Bearer " + token),
+    actual = Buffer.from(req.headers.authorization || "");
+  return expected.length === actual.length && timingSafeEqual(expected, actual);
+}
+async function main() {
+  await mkdir(root, { recursive: true, mode: 0o700 });
+  // This service runs as a single systemd process. A restart requeues local work at most twice.
+  for (const name of await readdir(root))
+    if (/^[a-f0-9]{64}\.json$/.test(name)) {
+      const job: Job = JSON.parse(
+        await readFile(path.join(root, name), "utf8")
+      );
+      if (job.id + ".json" !== name) throw Error("Invalid upscale record");
+      if (job.status === "running") {
+        job.status = job.attempts < 3 ? "queued" : "failed";
+        job.stage =
+          job.status === "queued"
+            ? "服务重启，等待恢复超分"
+            : "多次中断，已停止";
+        if (job.status === "failed") job.error = job.stage;
+        await persist(job);
+      }
+      jobs.set(job.id, job);
+    }
+  await prune();
+  const server = http.createServer(async (req, res) => {
+    if (!authenticated(req)) return json(res, 401, { error: "Unauthorized" });
+    const url = new URL(req.url || "/", "http://localhost");
+    try {
+      if (req.method === "GET" && url.pathname === "/health")
+        return json(res, 200, {
+          status: "ok",
+          version: 1,
+          active: busy ? 1 : 0,
+          queued: [...jobs.values()].filter((j) => j.status === "queued")
+            .length,
+          render: "realesr-general-x4v3",
+          gpu: process.env.CUDA_VISIBLE_DEVICES,
+        });
+      if (req.method === "POST" && url.pathname === "/v1/upscales") {
+        const input = inputSchema.parse(await body(req));
+        // Serialize admission across the async filesystem write to preserve idempotency.
+        if (accepting) return json(res, 503, { error: "请重试提交" });
+        accepting = true;
+        try {
+          const digest = createHash("sha256")
+            .update(JSON.stringify({ tenant: input.tenant, plan: input.plan }))
+            .digest("hex");
+          const existing = jobs.get(input.id);
+          if (existing)
+            return json(
+              res,
+              existing.digest === digest ? 200 : 409,
+              existing.digest === digest
+                ? publicJob(existing)
+                : { error: "任务 ID 已用于不同参数" }
+            );
+          if (
+            jobs.size >= 200 ||
+            [...jobs.values()].filter((j) =>
+              ["queued", "running"].includes(j.status)
+            ).length >= 30
+          )
+            return json(res, 429, { error: "超分队列已满，请稍后重试" });
+          const job: Job = {
+            ...input,
+            digest,
+            status: "queued",
+            stage: "排队中",
+            attempts: 0,
+            createdAt: new Date().toISOString(),
+            updatedAt: new Date().toISOString(),
+          };
+          await persist(job);
+          json(res, 202, publicJob(job));
+          void work().catch(() => {
+            console.error("Render queue failed");
+            process.exit(1);
+          });
+          return;
+        } finally {
+          accepting = false;
+        }
+      }
+      const match = url.pathname.match(
+        /^\/v1\/upscales\/([a-f0-9]{64})(?:\/(video))?$/
+      );
+      if (req.method === "GET" && match) {
+        const job = jobs.get(match[1]);
+        if (!job) return json(res, 404, { error: "任务不存在或已过期" });
+        if (!match[2]) return json(res, 200, publicJob(job));
+        if (job.status !== "succeeded")
+          return json(res, 409, { error: "成片尚未就绪" });
+        const output = path.join(root, job.id, "result.mp4");
+        const info = await stat(output);
+        res.writeHead(200, {
+          "content-type": "video/mp4",
+          "content-length": info.size,
+          "cache-control": "no-store",
+        });
+        await pipeline(createReadStream(output), res);
+        return;
+      }
+      json(res, 404, { error: "Not found" });
+    } catch (error) {
+      if (res.headersSent) {
+        res.destroy();
+        return;
+      }
+      json(res, 400, {
+        error:
+          error instanceof z.ZodError
+            ? "超分参数无效"
+            : error instanceof Error && !/https?:|\/var\//.test(error.message)
+              ? error.message.slice(0, 300)
+              : "请求处理失败",
+      });
+    }
+  });
+  server.requestTimeout = 120000;
+  server.headersTimeout = 10000;
+  server.listen(
+    Number(process.env.UPSCALE_PORT || 18341),
+    process.env.UPSCALE_HOST || "127.0.0.1"
+  );
+  const fatal = () => {
+    console.error("Render queue persistence failure");
+    process.exit(1);
+  };
+  const timer = setInterval(() => {
+    void work().catch(fatal);
+  }, 3000);
+  const cleanupTimer = setInterval(() => {
+    if (!busy && !accepting) void prune().catch(fatal);
+  }, 3600000);
+  cleanupTimer.unref();
+  for (const signal of ["SIGTERM", "SIGINT"])
+    process.on(signal, () => {
+      stopping = true;
+      clearInterval(timer);
+      clearInterval(cleanupTimer);
+      server.close();
+      // systemd KillMode=mixed also terminates FFmpeg; persisted running jobs resume on restart.
+      setTimeout(() => process.exit(0), 1000).unref();
+    });
+  console.log("Video upscale service ready");
+}
+void main().catch(() => {
+  console.error("Video upscale service startup failed");
+  process.exitCode = 1;
+});

services/video-upscale/upscale.py

@@ -0,0 +1,79 @@
+"""Bounded, tiled Real-ESRGAN video inference. No CPU fallback or frame interpolation."""
+import argparse, json, subprocess, sys, os, hashlib
+from fractions import Fraction
+from pathlib import Path
+import numpy as np
+import torch
+from PIL import Image
+sys.path.insert(0, str(Path(__file__).parent / 'vendor'))
+from srvgg_arch import SRVGGNetCompact
+
+def probe(file):
+    return json.loads(subprocess.check_output(['ffprobe','-v','error','-protocol_whitelist','file,pipe','-show_streams','-show_format','-of','json',file], timeout=30))
+
+def main():
+    p=argparse.ArgumentParser()
+    for name in ['input','output','model','resolution']: p.add_argument('--'+name, required=True)
+    p.add_argument('--tile', type=int, default=192)
+    a=p.parse_args()
+    if not torch.cuda.is_available(): raise RuntimeError('CUDA unavailable; refusing CPU fallback')
+    torch.set_num_threads(4)
+    info=probe(a.input); v=next(s for s in info['streams'] if s['codec_type']=='video')
+    if not any(name in info['format']['format_name'].split(',') for name in ['mov','mp4','matroska','webm','avi','flv','mpegts']): raise ValueError('Unsupported video container')
+    rotation=next((int(side['rotation'])%360 for side in v.get('side_data_list',[]) if 'rotation' in side),0)
+    if rotation%90: raise ValueError('Unsupported rotation angle')
+    w,h=int(v['width']),int(v['height'])
+    if rotation in [90,270]: w,h=h,w
+    fps=Fraction(v['avg_frame_rate'])
+    duration=float(info['format']['duration'])
+    if not 0<duration<=120 or not 0<float(fps)<=60 or w*h>3840*2160: raise ValueError('Input exceeds 120 s / 60 fps / 4K limits')
+    # Reject variable frame rate rather than silently altering timing/audio sync.
+    frame_times=subprocess.check_output(['ffprobe','-v','error','-protocol_whitelist','file,pipe','-select_streams','v:0','-show_entries','frame=best_effort_timestamp_time','-of','csv=p=0',a.input],timeout=60).decode().splitlines()
+    timestamps=[float(s.split(',')[0]) for s in frame_times if s.split(',')[0].strip()]
+    if any(abs((b-c)-float(1/fps))>max(.002,float(1/fps)*.05) for c,b in zip(timestamps,timestamps[1:])): raise ValueError('Variable frame rate video must be converted to constant frame rate first')
+    long,short={'1080P':(1920,1080),'2K':(2560,1440),'4K':(3840,2160)}[a.resolution]
+    scale=min((long if w>=h else short)/w,(short if w>=h else long)/h)
+    if not 1<scale<=4: raise ValueError('Target must exceed source resolution by no more than 4x')
+    ow,oh=int(w*scale)//2*2,int(h*scale)//2*2
+    model=SRVGGNetCompact(num_in_ch=3,num_out_ch=3,num_feat=64,num_conv=32,upscale=4,act_type='prelu')
+    if hashlib.sha256(Path(a.model).read_bytes()).hexdigest() != '8dc7edb9ac80ccdc30c3a5dca6616509367f05fbc184ad95b731f05bece96292': raise RuntimeError('Model checksum mismatch')
+    weights=torch.load(a.model,map_location='cpu',weights_only=True)
+    model.load_state_dict(weights.get('params_ema',weights.get('params',weights)),strict=True)
+    model.eval().to('cuda')
+    temp=a.output+'.silent.mp4'
+    decode=subprocess.Popen(['ffmpeg','-v','error','-filter_threads','1','-filter_complex_threads','1','-threads','2','-protocol_whitelist','file,pipe','-i',a.input,'-map','0:v:0','-f','rawvideo','-pix_fmt','rgb24','-vsync','0','pipe:1'],stdout=subprocess.PIPE)
+    encode=subprocess.Popen(['ffmpeg','-v','error','-filter_threads','1','-filter_complex_threads','1','-y','-f','rawvideo','-pix_fmt','rgb24','-s',f'{ow}x{oh}','-r',str(fps),'-i','pipe:0','-an','-c:v','libx264','-threads','4','-preset','fast','-crf','18','-pix_fmt','yuv420p',temp],stdin=subprocess.PIPE)
+    count=0
+    try:
+        with torch.inference_mode():
+            while True:
+                raw=decode.stdout.read(w*h*3)
+                if not raw: break
+                if len(raw)!=w*h*3: raise RuntimeError('Truncated video frame')
+                source=np.frombuffer(raw,dtype=np.uint8).reshape(h,w,3)
+                # CPU output buffer, bounded GPU tiles with enough halo for 35 convolutions.
+                output=np.empty((h*4,w*4,3),dtype=np.uint8)
+                for y in range(0,h,a.tile):
+                    for x in range(0,w,a.tile):
+                        x0,y0=max(0,x-40),max(0,y-40); x1,y1=min(w,x+a.tile+40),min(h,y+a.tile+40)
+                        tile=torch.from_numpy(source[y0:y1,x0:x1].copy()).permute(2,0,1).float().unsqueeze(0).to('cuda')/255
+                        enhanced=model(tile).clamp_(0,1).squeeze(0).permute(1,2,0).mul_(255).round_().byte().cpu().numpy()
+                        ex,ey=min(w,x+a.tile),min(h,y+a.tile)
+                        output[y*4:ey*4,x*4:ex*4]=enhanced[(y-y0)*4:(ey-y0)*4,(x-x0)*4:(ex-x0)*4]
+                frame=Image.fromarray(output)
+                if (ow,oh)!=(w*4,h*4): frame=frame.resize((ow,oh),Image.Resampling.LANCZOS)
+                encode.stdin.write(frame.tobytes()); count+=1
+                if count%5==0: print(json.dumps({'stage':f'AI 超分 {count}/{len(timestamps)} 帧'}),flush=True)
+        encode.stdin.close()
+        if decode.wait(timeout=30)!=0 or encode.wait(timeout=120)!=0: raise RuntimeError('Video codec failed')
+    finally:
+        for process in [decode,encode]:
+            if process.poll() is None: process.kill(); process.wait()
+    if count!=len(timestamps) or count==0: raise RuntimeError('Output frame count mismatch')
+    subprocess.run(['ffmpeg','-v','error','-filter_threads','1','-filter_complex_threads','1','-y','-protocol_whitelist','file,pipe','-i',temp,'-protocol_whitelist','file,pipe','-i',a.input,'-map','0:v:0','-map','1:a?','-map_metadata','1','-metadata:s:v:0','rotate=0','-c:v','copy','-c:a','aac','-b:a','192k','-movflags','+faststart',a.output],check=True,timeout=120)
+    actual=probe(a.output); stream=next(s for s in actual['streams'] if s['codec_type']=='video')
+    if int(stream['nb_frames'])!=count or abs(float(actual['format']['duration'])-duration)>.2: raise RuntimeError('Output timing mismatch')
+    if any(s['codec_type']=='audio' for s in actual['streams']) != any(s['codec_type']=='audio' for s in info['streams']): raise RuntimeError('Output audio stream missing')
+    os.unlink(temp)
+    print(json.dumps({'result':{'width':ow,'height':oh,'durationSeconds':float(actual['format']['duration']),'fps':float(fps),'frames':count,'audio':any(s['codec_type']=='audio' for s in actual['streams']),'engine':'Real-ESRGAN','model':'realesr-general-x4v3'}}),flush=True)
+if __name__=='__main__': main()

services/video-upscale/vendor/LICENSE

@@ -0,0 +1,29 @@
+BSD 3-Clause License
+
+Copyright (c) 2021, Xintao Wang
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+
+3. Neither the name of the copyright holder nor the names of its
+   contributors may be used to endorse or promote products derived from
+   this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.