feat: 补齐前端主流程并收紧工作台布局

统一工作台紧凑头部与形态图库间距，修复主体身份布局，并接入正式严格视频质量流水线。
2026-09-08 09:25:14 +08:00
parent 0f8b2074e4
commit 4edb0ff6cc
25 changed files with 355 additions and 233 deletions
@@ -2,10 +2,11 @@

 短剧 Agent 的前端工作台。支持剧本创作、拆解、视觉风格、角色选角、主体形态、分镜设计、镜头首帧与视频生产，使用真实后端 API，不包含演示数据或浏览器端模型调用。

-## 最新同步（2026-09-04）
+## 最新同步（2026-09-07）

- 对齐后端 `dev@05bb9e8b7e0a0e85303b16eb3a2ac634e01dd7eb`：图片／视频模型由后端统一配置，前端不再发送固定 Provider。
- 新增首帧视觉校验、有限自动修复、当前集小批量质量生成、视频抽样帧质量检查、视频修复候选与复检晋升、只读模型能力查询。默认不调用模型，付费操作需确认，结果可定位和导出。
+- 对齐后端 `dev@697cf864e1e8d6f66e3ad7beeebdfd0ae0d6c16c`：图片／视频模型由后端统一配置，前端不再发送固定 Provider。
+- 视频主入口使用严格质量流水线：候选生成完成后自动校验、有限修复，通过后才晋升主视频；首帧质量生成和手工抽样检查仍保留。
+- 分镜只使用正式项目级生成接口，不把 `generate-test`、`start-test` 或 `stream-test` 暴露为生产入口。
 - 保留主要布局和业务能力，导出／执行记录／提示词等次要入口收进菜单，长介绍和高级规则按需展开；关键错误、过期、未保存与费用提示继续直显。
 - 全部业务抽屉统一挂载到 `body`，侧栏及遮罩覆盖完整视口，不受 `main/content` 裁切；保持原有宽度、正文内部滚动和关闭后配置草稿。

@@ -153,10 +154,8 @@ Oxfmt 不负责代码质量，Oxlint 不负责 Vue 的完整类型推导；Vue S
 | POST | `/projects/:id/breakdown/resume-shots` | 复用已成功剧集镜头，生成缺失剧集 |
 | POST | `/projects/:id/breakdown/resume-storyboard` | 修复已有镜头的 SubjectRef / Form，再持久化 |
 | GET | `/projects/:id/storyboard-directions?episodeNo=N` | 正式 Shot ID、导演设计、单集覆盖率 |
-| POST | `/projects/:id/storyboard-directions/generate-test` | 单集导演设计；明确发送 `persist: true` |
 | POST | `/projects/:id/storyboard-directions/generate` | 项目批量；`concurrency`、`force` |
 | GET | `/projects/:id/storyboard-visual-states?episodeNo=N` | 已持久化的环境、主体状态、单集覆盖率 |
-| POST | `/projects/:id/storyboard-visual-states/generate-test` | 单集即时状态；`persist: true`、`maxRepairAttempts` |
 | POST | `/projects/:id/storyboard-visual-states/generate` | 项目批量；增加 `maxRepairAttempts` |
 | GET | `/storyboard-shots/:shotId/references` | 主体形态参考图、缺失原因 |
 | GET | `/storyboard-shots/:shotId/generation-spec` | 后端确定性编译结果，不调用模型 |
@@ -188,10 +187,11 @@ Oxfmt 不负责代码质量，Oxlint 不负责 Vue 的完整类型推导；Vue S
 | GET / POST | `/projects/:id/keyframes/readiness`、`/keyframes/generate` | 首帧就绪检查与批量 后端图片模型 生图 |
 | GET / POST | `/storyboard-shots/:shotId/keyframes`、`/keyframe` | 单镜头首帧历史与生图 |
 | PUT | `/storyboard-shots/:shotId/keyframes/:keyframeId/primary` | 切换主首帧 |
-| GET / POST | `/projects/:id/videos/readiness`、`/videos/generate` | 视频前置检查与批量 后端视频模型 任务 |
+| GET / POST | `/projects/:id/videos/readiness`、`/videos/generate-quality` | 视频前置检查与批量严格质量任务 |
 | GET | `/projects/:id/videos/status` | 项目最近视频任务状态 |
 | POST | `/projects/:id/videos/retry-failed` | 重试最近失败的视频任务 |
-| GET / POST | `/storyboard-shots/:shotId/videos` | 单镜头视频历史与任务创建 |
+| GET | `/storyboard-shots/:shotId/videos` | 单镜头视频历史 |
+| POST | `/storyboard-shots/:shotId/videos/generate-quality` | 单镜头严格质量候选；后台自动校验与有限修复 |
 | PUT | `/storyboard-shots/:shotId/videos/:videoId/primary` | 切换主视频 |
 | POST | `/projects/:id/production/start` | 高级串联入口；重复预检、费用确认，仅传两种 Provider，不生成首帧、不等待成片 |

@@ -213,9 +213,8 @@ API 层也提供 `state` 与 `breakdown/latest` 方法。页面通过共享 chec
 1. 入口为 `/projects/:projectId/storyboard`。先完成 Breakdown 的主体、形态、镜头与持久化，再生成 Direction，最后生成 VisualState。
 2. 当前 Direction 服务读取**最新** Breakdown 的 `state.breakdownResult.storyboardEpisodeShots`；VisualState 读取顶层 `state.storyboardEpisodeShots`、`subjectCandidates`、`subjectForms`。页面不使用历史可恢复快照代替这些生成前置条件。历史快照只补充镜头标题和剧情描述。
 3. 镜头 ID 来自正式数据库 GET，按 `shotId` 关联 Direction 和 VisualState。`shotNo` 仅在 Beat 内唯一，不能用它拼装接口 ID。VisualState 查询响应的环境字段是扁平结构，与单集生成返回的领域结构不同。
-4. 单集入口虽然名为 `generate-test`，也是后端目前唯一的单集生成接口。页面发送 `persist: true`，通过校验才保存；**单集始终重生成**，不受批量的覆盖开关影响。
-5. 批量默认 `force: false`，跳过完整剧集。重新点击“补齐”用于重试未完成项，没有虚构独立 retry 接口。`force: true` 才覆盖全部。VisualState 可设置非负整数修复次数，`0` 表示不修复。
-6. VisualState 单集生成需要该集所有正式镜头都有 Direction；批量接口由后端逐集检查，缺少前置条件的集数会单独失败。HTTP 200 也可能包含校验失败或部分失败，页面展示逐集诊断和修复次数，不视为全成功。
+4. 页面只使用正式项目批量接口；`generate-test` 暂不作为前端入口。批量默认 `force: false`，跳过完整剧集；`force: true` 才覆盖全部。VisualState 可设置非负整数修复次数，`0` 表示不修复。
+5. VisualState 批量接口由后端逐集检查；缺少完整 Direction 等前置条件的集数会单独失败。HTTP 200 也可能包含校验失败或部分失败，页面展示逐集诊断和修复次数，不视为全成功。
 7. 上游重生成**不会自动使下游旧数据失效**。覆盖 Direction 后请重生成 VisualState，修改设计／状态／参考图后按需覆盖 Prompt。覆盖率只说明字段已保存，不能证明下游与新版本一致。
 8. 分镜页用刷新按钮查询当前剧集的两类正式结果；这不是全项目实时进度。后端没有独立 Storyboard running / executionId 查询。回执只保留当前浏览器会话，切换页面不丢失，刷新浏览器会丢失。
 9. 参考图和 GenerationSpec 按需 GET，切换镜头或设计变化时取消旧查询。GenerationSpec 需要有效的 Direction、VisualState 和形态关联，不生成图片或视频。
@@ -270,7 +269,7 @@ API 层也提供 `state` 与 `breakdown/latest` 方法。页面通过共享 chec
 - 精简的进度接口及按 graph 分页的 checkpoint；
 - 将 Breakdown 改为 202 异步任务提交，再以轮询或 SSE 读取进度。

-前端未模拟这些能力，不调用 `stream-test` 来伪造进度。Storyboard 的单集 `generate-test` 仅在用户明确确认生成后调用。
+前端未模拟这些能力，也不调用 `stream-test`、`start-test` 或 Storyboard `generate-test` 来伪造生产功能。

 ## 部署

@@ -0,0 +1,64 @@
+# Design QA
+
+- Source visual truth:
+    - `/workspace/scratch/6e52996114b8/upload/83e0987d-5b6e-46b1-b39b-0674bf374f43.png`
+    - `/workspace/scratch/6e52996114b8/upload/60e23391-685b-4538-95a2-46f6334f5f6d.png`
+    - `/workspace/scratch/6e52996114b8/upload/430bec04-93fd-429f-a820-1edefdafc792.png`
+- Source pixel dimensions: `2048 × 819`, `447 × 203`, `238 × 198`.
+- Intended implementation viewport: desktop dark theme; the primary reference is `2048 × 819`.
+- Implementation screenshot: unavailable.
+- CSS size and density normalization: unavailable because the implementation could not be opened in the cloud browser.
+- State: 主体身份详情页已有身份母版；形态图库默认筛选和布局切换状态。
+
+## Full-view comparison evidence
+
+The source images were opened and inspected. The first source shows the generic `.asset-image` width expanding the identity-anchor summary image across the detail panel and squeezing its explanatory copy into a narrow column on the far right. The second source shows the gallery summary/actions and filter panel consuming too much vertical space. The third source establishes the compact dark workspace density used by the surrounding pages.
+
+The implementation passed lint, formatting, TypeScript, component/unit tests, and the production build. A local Sites preview started successfully at the required port, but the cloud browser rejected `terminal.local` with `net::ERR_BLOCKED_BY_CLIENT` on both the existing tab and one fresh tab. Therefore no browser-rendered implementation evidence is available.
+
+## Focused region comparison evidence
+
+- Identity anchor: `.identity-anchor > .asset-image` now has a component-local `80 × 80` square constraint with higher selector specificity than the generic AssetImage width rule. The adjacent copy has `min-width: 0`.
+- Identity detail container: the detail scrollbar, its container/content, and the panel now explicitly use `width: 100%`, `min-width: 0`, and `max-width: 100%` to contain long content.
+- Gallery density: the summary/actions moved into the `WorkspacePage` header and the filter region's vertical padding changed from `14px` to `5px`.
+- Page headers: every `WorkspacePage` caller now uses compact mode.
+
+These are code-level checks only and do not replace a rendered focused-region comparison.
+
+## Required fidelity surfaces
+
+- Fonts and typography: existing project font stack, sizes, weights, and wrapping were preserved; browser comparison is blocked.
+- Spacing and layout rhythm: the requested header compaction and `5px` gallery filter padding are implemented; browser measurement is blocked.
+- Colors and visual tokens: existing dark/light CSS variables were preserved; browser comparison is blocked.
+- Image quality and asset fidelity: no source assets were replaced or generated; existing backend images remain in use. Browser crop verification is blocked.
+- Copy and content: existing product copy is retained except where video generation now describes the formal strict-quality workflow.
+
+## Findings
+
+- [P1] Browser-rendered comparison unavailable
+    - Location: all three target states.
+    - Evidence: the preview process is healthy, but the required cloud browser returns `net::ERR_BLOCKED_BY_CLIENT` for `http://terminal.local:4173/`.
+    - Impact: exact visual fidelity, responsive behavior, interactions, and console state cannot be certified from the browser.
+    - Fix: open the same preview in an available Work Mode cloud browser, reproduce the target data/theme, capture the three states, and rerun this QA.
+
+## Primary interactions tested
+
+Automated component tests cover filters, layout switching, workspace drawers, project-level formal storyboard actions, video quality API parameters, disabled states, receipts, and navigation. Browser interactions were not available.
+
+## Console errors checked
+
+Not checked because the page could not be opened in the cloud browser.
+
+## Comparison history
+
+- Iteration 1: identified the P1 implementation-capture blocker after the local preview started. No visual iteration could be completed.
+
+## Implementation checklist
+
+- [x] Constrain the identity anchor thumbnail and detail width.
+- [x] Move gallery summary/actions into the compact header.
+- [x] Change gallery filter vertical padding from 14px to 5px.
+- [x] Apply compact mode to all workspace pages.
+- [ ] Capture and compare the rendered identity and gallery states in the cloud browser.
+
+final result: blocked
@@ -1,6 +1,6 @@
 # 后端功能覆盖说明

-核对日期：2026-09-03。后端：`qianlanse/short-drama-agent`，`dev@05bb9e8b7e0a0e85303b16eb3a2ac634e01dd7eb`。以 API 路由和实际 service／workflow 实现为依据，不只根据接口名称推断功能。本轮不修改后端。
+核对日期：2026-09-07。后端：`qianlanse/short-drama-agent`，`dev@697cf864e1e8d6f66e3ad7beeebdfd0ae0d6c16c`。以 API 路由和实际 service／workflow 实现为依据，不只根据接口名称推断功能。本轮不修改后端。

 ## 页面与接口能力

@@ -18,16 +18,16 @@
 | 身份参考图 | 主体身份图库 | `identity/images`、图片 `anchor` | 保留完整预览、历史、辅助视角与母版切换；修正道具继承说明 |
 | 正式形态提示词 | 形态卡片 → 提示词与生成；批量工具 | POST `subject-forms/:id/generation-prompt`、项目 `subject-forms/generation-prompts` | 新增单个生成／重生成、原始素材与正式文本对照、导出、批量补齐／覆盖及失败回执 |
 | 形态图片 | 形态图片及图库 | 项目 `subject-forms`、形态 `images`、主图 PUT、项目 `subject-images/generate` | 增加本批数量上限；支持后端自动补正式 Prompt；人物／场景／道具已锁定母版继承；过期刷新只生成候选 |
-| 导演设计与即时状态 | 分镜设计 | directions／visual-states 查询、单集 generate-test、项目 generate | 保留单集持久化（`persist: true`）、项目批量、覆盖、校验修复次数、逐集回执 |
+| 导演设计与即时状态 | 分镜设计 | directions／visual-states 查询、项目 `generate` | 只保留正式项目批量、覆盖、校验修复次数和逐集回执；不暴露 `generate-test` |
 | 参考图与规格 | 分镜参考图／生产单镜头 | `references`、`generation-spec`、`keyframe-spec`、`video-generation-spec` | 保留按需只读检查与缺失原因，查询不调用生成模型 |
 | 视频提示词 | 分镜设计／镜头生产 | 单镜头 `video-prompt`、项目 `video-prompts/readiness` 与 `generate` | 保留单镜头／项目生成，重写需确认 |
 | 镜头首帧 | 镜头生产 | 项目 `keyframes/readiness` 与 `generate`、单镜头 keyframe／keyframes、主图 PUT | 新增批量 `limit` 与成对尺寸；保留当前集／过期项／覆盖范围；显示因上限未执行数 |
-| 视频任务与主视频 | 镜头生产 | 项目 videos readiness／generate／status／retry-failed，逐镜 videos，单任务 status，主视频 PUT | 保留任务提交、轮询、播放、失败重试与主视频切换；不把 HTTP 完成当成片完成 |
+| 视频任务与主视频 | 镜头生产 | 项目 videos readiness／generate-quality／status／retry-failed，逐镜 generate-quality／videos，单任务 status，主视频 PUT | 主入口改为严格质量候选：后台自动校验、有限修复，通过后晋升；保留状态刷新、播放和失败重试 |
 | 高级串联生产 | 镜头生产批量面板底部（默认折叠） | POST `/projects/:id/production/start` | 新增预检、再次校验、费用确认、分阶段回执与诊断导出；明确不是全链自动制片 |

 ## 新增质量检查与修复

-已接入：首帧 `validate`、`repair`，项目 `keyframes/generate-quality`，视频 `validate`／`repair`，公开图片模型能力及逐镜能力 GET。生成请求不再传固定 Provider。质量接口的详细参数、默认一镜范围、费用确认、主图晋升、三点视频抽样和会话回执限制见 [前端简化与质量功能同步](frontend-simplification.md)。首帧校验历史尚无 GET，不虚构读取接口；视频已有结果从 `rawJson.videoValidation` 读取。
+已接入：首帧 `validate`、`repair`，项目 `keyframes/generate-quality`，视频 `validate`／`repair`，以及单镜头和项目级 `videos/generate-quality`。严格视频候选由后端 Poller 自动执行 Validator，失败时按上限创建 Repair Candidate，只有 PASS 才晋升 Primary；前端识别 `rawJson.qualityCandidate`，禁止未通过候选绕过校验手动设为主视频。生成请求不再传固定 Provider。质量接口的详细参数、默认一镜范围、费用确认、主图晋升、三点视频抽样和会话回执限制见 [前端简化与质量功能同步](frontend-simplification.md)。首帧校验历史尚无 GET，不虚构读取接口；视频已有结果从 `rawJson.videoValidation` 读取。

 ## 高级串联生产边界

@@ -63,7 +63,7 @@
 | `timeline` | 新诊断从 `timeline/grouped` 展平并按 index 排序，保持同一份数据 |
 | 图片／首帧／视频的 GET primary | 从后端完整列表的 primary 标志选取，无需再次发相同内容请求 |
 | POST `/projects/:id/resume` | 实际只返回恢复信息，并不启动恢复工作流；不显示虚假的“继续执行”按钮 |
-| `projects/stream-test`、`breakdown/start-test` | 测试专用，不暴露生产入口；Storyboard 的 generate-test 是现有正式单集生成入口，仍传 `persist: true` |
+| `projects/stream-test`、`breakdown/start-test`、Storyboard `generate-test` | 测试或试运行路由，不暴露生产入口；分镜只使用正式项目级 `generate` |
 | 手工保存形态 generationPrompt、取消运行任务、删除项目、合并最终视频 | 后端没有对应正式路由，本轮不虚构按钮、成功提示或客户端模拟实现 |

 ## 回归验证
@@ -35,13 +35,17 @@
 | 已完成首帧 → 质量检查 → 仅校验 | POST `/storyboard-shots/:shotId/keyframes/:keyframeId/validate` | 调用视觉模型，不重生成、不切换主首帧 |
 | 首帧 → 质量检查 → 校验并修复 | POST 同路径 `/repair` | 有限重生成；新候选通过才晋升主首帧；原图直接通过不会自动晋升 |
 | 生产工具栏 → 首帧质量生成 | POST `/projects/:projectId/keyframes/generate-quality` | 当前集、默认最多一镜、并发 1、最多修复 1 次；每镜候选通过后才设主图 |
+| 单镜头 → 提交视频质量任务 | POST `/storyboard-shots/:shotId/videos/generate-quality` | 创建严格质量候选；生成完成后由后端自动校验，失败时最多自动修复 2 次，通过后才晋升主视频 |
+| 批量生产 → 视频成片 | POST `/projects/:projectId/videos/generate-quality` | 显式限制本批镜头数和每镜修复次数；只提交 ready 镜头，后续 Validator／Repair 由 Poller 推进 |
 | 已完成视频 → 质量检查 | POST `/storyboard-shots/:shotId/videos/:videoId/validate` | 三点抽帧视觉检查，不检测音频／口型／完整连续运动；普通候选不切换主视频，修复候选通过后自动晋升 |

 批量质量配置与普通批量面板独立。后端非 force 先根据全项目是否有过期项决定筛选，再筛当前集；因此其他集存在过期项而当前集无过期项时，本集不会先补缺失首帧。前端明确提示切换剧集，不悄悄扩大范围。

 允许文字按行输入并清理空白／重复；未填写不自行授权文字。宽高同时留空或填写正整数，修复次数可为 0。确认区显示本次最多生图／视觉校验调用次数，修改参数或目标后撤销费用确认。

-### 视频修复与复检晋升（本轮后端追加更新）
+### 视频自动质量链与手工复检
+
+生产页的单镜和批量视频主入口使用 `generate-quality`。HTTP 回执只表示初始 Candidate 已创建；后台 Poller 会在 Provider 完成后自动执行 Validator，失败时按 `maxRepairAttempts` 创建下一代 Repair Candidate，只有通过校验的版本才会晋升 Primary。前端读取 `rawJson.qualityCandidate` 和 `qualityPipeline` 标记质量候选，并禁止未通过候选使用普通“设为主视频”绕过校验。

 视频质量面板新增“生成修复候选”，使用 POST `/storyboard-shots/:shotId/videos/:videoId/repair`，只接受 `maxRepairAttempts`（默认 2，正整数）。该参数是修复链的上限，不是本次自动生成次数。本次只提交一个视频候选，不自动等待、复检或循环重试；修复约束和允许文字沿用已保存的失败校验。

@@ -69,7 +69,7 @@ function exportScript() {
 </script>

 <template>
-    <WorkspacePage split class="script-workspace-page"
+    <WorkspacePage split compact class="script-workspace-page"
        ><template #header
            ><div class="stage-strip">
                <!-- 以下是创作阶段进度 -->
@@ -50,7 +50,10 @@ const {
    keyframeLimit,
    keyframeWidth,
    keyframeHeight,
+    videoLimit,
+    videoRepairAttempts,
    keyframeValid,
+    videoQualityValid,
    batchValid,
    blocked,
    run
@@ -100,15 +103,15 @@ const stages = computed(() => {
            key: 'videos' as const,
            code: '03',
            title: '视频成片',
-            description: '以主首帧作为第一帧，连同提示词和主体参考图提交视频模型。',
+            description: '生成质量候选，完成后自动校验；失败时有限修复，通过后才晋升主视频。',
            icon: Film,
            data: data?.videos,
-            done: data?.videoStatus.completed ?? 0,
+            done: data?.videos.skipped ?? 0,
            action: force.value
-                ? '为全部就绪镜头新增视频'
+                ? '新增项目质量视频候选'
                : staleVideos.value > 0
                  ? `更新 ${staleVideos.value} 个过期主视频`
-                  : '提交就绪视频任务'
+                  : '提交视频质量任务'
        }
    ]
 })
@@ -237,8 +240,7 @@ watch(
                            </div>
                            <DetailDisclosure title="处理范围与覆盖规则" class="mt-3"
                                ><p class="text-xs leading-6 text-muted">
-                                    提示词与视频批量操作面向全项目；首帧默认补当前剧集，存在过期主首帧时优先更新全项目过期项，覆盖模式处理全项目；过期首帧成功后会自动接替旧主图，其余新增候选。图片模型
-                                    与视频模型均可能产生费用。
+                                    提示词与视频就绪检查面向全项目；视频实际提交受本批镜头上限控制，并使用严格质量流水线。首帧默认补当前剧集，存在过期主首帧时优先更新全项目过期项，覆盖模式处理全项目；过期首帧成功后会自动接替旧主图，其余新增候选。图片模型、视频模型与视觉校验均可能产生费用。
                                </p></DetailDisclosure
                            >
                        </section>
@@ -375,14 +377,15 @@ watch(
                                                        blocked ||
                                                        !batchValid ||
                                                        (stage.key === 'keyframes' && !keyframeValid) ||
+                                                        (stage.key === 'videos' && !videoQualityValid) ||
                                                        !(stage.data?.ready ?? 0)
                                                    "
                                                    acknowledgement
                                                    :description="
                                                        stage.key === 'videos'
                                                            ? staleVideos && !force
-                                                                ? '当前存在过期主视频，非覆盖模式只会为这些镜头创建新的视频模型任务；新视频完成后会自动接替过期主视频。'
-                                                                : '只为当前就绪镜头创建视频模型异步任务。创建成功不代表视频已经完成，任务会在后台继续运行。'
+                                                                ? `当前存在过期主视频，非覆盖模式只会为其中至多 ${videoLimit || '—'} 个镜头启动严格质量流水线；候选自动校验并有限修复，通过后才接替旧主视频。`
+                                                                : `只为至多 ${videoLimit || '—'} 个就绪镜头启动严格质量流水线。任务创建后由后端自动校验，最多修复 ${videoRepairAttempts || '—'} 次，通过后才晋升主视频。`
                                                            : stage.key === 'keyframes'
                                                              ? `处理缺失或过期主首帧。过期首帧成功后自动接替旧主图（覆盖模式也如此），其余已有主图只新增候选。本批上限：${keyframeLimit || '不限制'}；尺寸：${keyframeWidth && keyframeHeight ? `${keyframeWidth} × ${keyframeHeight}` : '后端默认'}。`
                                                              : '只处理通过视频提示词就绪检查的镜头；覆盖模式会重写已有提示词。'
@@ -424,6 +427,29 @@ watch(
                                                        上限须为正整数；宽高须同时留空或填写正整数。
                                                    </p>
                                                </div>
+                                                <div v-if="stage.key === 'videos'" class="mt-4 grid gap-3">
+                                                    <label
+                                                        ><span class="field-label">本批镜头上限</span
+                                                        ><NInputNumber
+                                                            v-model:value="videoLimit"
+                                                            :min="1"
+                                                            :precision="0"
+                                                            :disabled="operation.pending"
+                                                            :input-props="{ 'aria-label': '视频质量任务镜头上限' }"
+                                                    /></label>
+                                                    <label
+                                                        ><span class="field-label">每镜最多自动修复</span
+                                                        ><NInputNumber
+                                                            v-model:value="videoRepairAttempts"
+                                                            :min="1"
+                                                            :precision="0"
+                                                            :disabled="operation.pending"
+                                                            :input-props="{ 'aria-label': '视频自动修复次数' }"
+                                                    /></label>
+                                                    <p v-if="!videoQualityValid" class="text-xs text-danger">
+                                                        本批上限和自动修复次数都必须是正整数。
+                                                    </p>
+                                                </div>
                                            </article>
                                        </div>
                                        <div v-if="query.data.value" class="panel mt-4 p-5">
@@ -3,7 +3,7 @@ import type { PromptBatchResult, StoryboardBatchInput } from '../storyboard/type
 import type {
    GenerateKeyframeInput,
    GenerateKeyframesInput,
-    GenerateVideosInput,
+    GenerateQualityVideosInput,
    KeyframeBatchResult,
    KeyframeReadiness,
    KeyframeSpec,
@@ -13,7 +13,9 @@ import type {
    RetryVideosResult,
    ShotKeyframe,
    ShotVideo,
-    VideoBatchResult,
+    ShotVideoQualityResult,
+    VideoQualityBatchResult,
+    VideoQualityConfig,
    VideoGenerationSpec,
    VideoReadiness
 } from './types'
@@ -54,8 +56,9 @@ export const productionApi = {
        }),
    videoReadiness: (projectId: string, force: boolean, signal?: AbortSignal) =>
        request<VideoReadiness>(`${projectPath(projectId)}/videos/readiness?force=${force}`, { signal }),
-    generateVideos: (projectId: string, input: GenerateVideosInput) =>
-        request<VideoBatchResult>(`${projectPath(projectId)}/videos/generate`, {
+    /** 启动项目级严格视频质量流水线；创建后由后端自动完成校验、有限修复与晋升。 */
+    generateQualityVideos: (projectId: string, input: GenerateQualityVideosInput) =>
+        request<VideoQualityBatchResult>(`${projectPath(projectId)}/videos/generate-quality`, {
            method: 'POST',
            body: input,
            timeoutMs: 0
@@ -86,10 +89,11 @@ export const productionApi = {
        request<VideoGenerationSpec>(`${shotPath(shotId)}/video-generation-spec`, { signal }),
    listVideos: (shotId: string, signal?: AbortSignal) =>
        request<ShotVideo[]>(`${shotPath(shotId)}/videos`, { signal }),
-    generateVideo: (shotId: string) =>
-        request<ShotVideo | null>(`${shotPath(shotId)}/videos`, {
+    /** 启动单镜头严格视频质量流水线；返回初始候选，最终主视频以后端状态为准。 */
+    generateQualityVideo: (shotId: string, input: VideoQualityConfig) =>
+        request<ShotVideoQualityResult>(`${shotPath(shotId)}/videos/generate-quality`, {
            method: 'POST',
-            body: {},
+            body: input,
            timeoutMs: 0
        }),
    refreshVideo: (videoId: string, signal?: AbortSignal) =>
@@ -33,8 +33,11 @@ function exportReceipt() {
            >
        </p>
        <p v-else-if="receipt.kind === 'videos'" class="mt-3 text-xs text-muted">
-            全项目 {{ receipt.result.total }} 镜 · 本次目标 {{ receipt.result.targetCount }} · 已创建任务
-            {{ receipt.result.created }} · 跳过 {{ receipt.result.skipped }} · 创建失败 {{ receipt.result.failed }}
+            全项目 {{ receipt.result.total }} 镜 · 就绪 {{ receipt.result.readyTargetCount }} · 本次目标
+            {{ receipt.result.targetCount }} · 已创建质量候选 {{ receipt.result.created }} · 待后续提交
+            {{ receipt.result.remainingReady }} · 跳过 {{ receipt.result.skipped }} · 进行中
+            {{ receipt.result.inProgress }} · 阻塞 {{ receipt.result.blocked }} · 创建失败
+            {{ receipt.result.failed }}
        </p>
        <p v-else class="mt-3 text-xs text-muted">
            检测到 {{ receipt.result.totalFailed }} 个失败镜头 · 已重新提交 {{ receipt.result.retried }} · 提交失败
@@ -47,7 +50,7 @@ function exportReceipt() {
            :show-icon="false"
            class="mt-3"
        >
-            视频任务已经创建，但尚未代表成片完成。后台会继续查询 视频模型，以下方任务状态和视频资产为准。
+            严格质量候选已经创建，但尚未代表成片完成。后台会继续查询视频模型、执行视觉校验，并在失败时有限修复；只有校验通过的候选才会晋升主视频。
        </NAlert>
        <NAlert v-if="receipt.result.failed" role="alert" type="error" :show-icon="false" class="mt-3">
            本次存在部分失败，成功提交或生成的结果不会回滚。请查看逐镜错误后再重试。
@@ -27,7 +27,7 @@ import ActionMenu from '../../../components/ui/ActionMenu.vue'
 import { AppDialog } from '../../../components/ui'
 import ModelCapabilities from './ModelCapabilities.vue'
 import type { QualityTarget } from '../quality.types'
-import { savedVideoValidation, videoRepairInfo } from '../quality'
+import { savedVideoValidation, videoQualityInfo, videoRepairInfo } from '../quality'

 /** 单镜头生产面板只负责正式资产，不修改上游导演设计与即时状态。 */
 const props = defineProps<{
@@ -162,14 +162,17 @@ async function setPrimaryKeyframe() {
    await refreshAll()
 }

-/** 创建单镜头视频模型视频任务；任务完成由后台轮询确认。 */
+/** 创建单镜头严格质量候选；后端完成生成后自动校验、有限修复并按结果晋升。 */
 async function generateVideo() {
    if (!canGenerateVideo.value) return
    const shotId = props.shot.shotId
-    await runOperation(props.projectId, `提交镜头 ${props.shot.shotNo} 视频任务`, async () => {
-        const result = await productionApi.generateVideo(shotId)
-        if (!result || result.shotId !== shotId || !isActiveVideo(result))
-            throw new Error('接口未返回有效的视频生成任务。')
+    await runOperation(props.projectId, `提交镜头 ${props.shot.shotNo} 视频质量任务`, async () => {
+        const result = await productionApi.generateQualityVideo(shotId, {
+            maxRepairAttempts: 2,
+            allowedTexts: []
+        })
+        if (!result || result.shotId !== shotId || !result.candidate || !isActiveVideo(result.candidate))
+            throw new Error('接口未返回有效的视频质量候选任务。')
    })
    await refreshAll()
 }
@@ -190,7 +193,12 @@ async function setPrimaryVideo() {
    const videoId = confirmingVideo.value
    if (assetBlocked.value || !videoId) return
    const video = videos.value.find(item => item.id === videoId)
-    if (!video || (videoRepairInfo(video.rawJson) && savedVideoValidation(video.rawJson)?.passed !== true)) return
+    if (
+        !video ||
+        ((videoRepairInfo(video.rawJson) || videoQualityInfo(video.rawJson)) &&
+            savedVideoValidation(video.rawJson)?.passed !== true)
+    )
+        return
    const shotId = props.shot.shotId
    confirmingVideo.value = ''
    await runOperation(props.projectId, `设置镜头 ${props.shot.shotNo} 主视频`, async () => {
@@ -396,6 +404,19 @@ async function inspectSpecs() {
                            <NTag v-if="videoRepairInfo(item.rawJson) && !item.isPrimary" size="small" :bordered="false"
                                >修复候选 · 待复检</NTag
                            >
+                            <NTag
+                                v-else-if="videoQualityInfo(item.rawJson) && !item.isPrimary"
+                                size="small"
+                                :bordered="false"
+                                >质量候选 ·
+                                {{
+                                    savedVideoValidation(item.rawJson)
+                                        ? savedVideoValidation(item.rawJson)?.passed
+                                            ? '已通过'
+                                            : '未通过'
+                                        : '自动校验中'
+                                }}</NTag
+                            >
                            <NTag v-if="item.isPrimary && videoStale" size="small" :bordered="false">已过期</NTag>
                            <span class="text-[10px] text-muted"
                                >{{ item.durationSeconds || shot.durationSeconds || '—' }}s</span
@@ -430,7 +451,7 @@ async function inspectSpecs() {
                                    item.status === 'completed' &&
                                    item.videoUrl &&
                                    !item.isPrimary &&
-                                    (!videoRepairInfo(item.rawJson) ||
+                                    (!(videoRepairInfo(item.rawJson) || videoQualityInfo(item.rawJson)) ||
                                        savedVideoValidation(item.rawJson)?.passed === true)
                                "
                                :disabled="assetBlocked"
@@ -454,15 +475,17 @@ async function inspectSpecs() {
            <p v-else class="mt-4 text-xs text-muted">尚无视频任务记录。</p>
            <div class="mt-4 flex flex-wrap gap-3">
                <ConfirmAction
-                    :label="videoStale ? '重新生成主视频' : currentVideo ? '再生成一个视频候选' : '提交本镜视频任务'"
+                    :label="
+                        videoStale ? '重新生成质量视频' : currentVideo ? '再生成一个质量候选' : '提交本镜视频质量任务'
+                    "
                    :disabled="!canGenerateVideo"
                    acknowledgement
                    :description="
                        videoStale
-                            ? '使用当前视频提示词、主首帧和主体参考图重新创建视频模型任务；完成后会自动接替当前过期主视频。'
+                            ? '使用当前视频提示词、主首帧和主体参考图启动严格质量流水线；最多自动修复 2 次，通过后才接替当前过期主视频。'
                            : currentVideo
-                              ? '使用当前视频提示词、主首帧和主体参考图新增一个视频模型视频候选，不自动替换主视频。'
-                              : '使用当前视频提示词、主首帧和主体参考图创建视频模型任务；任务会在后台继续运行。'
+                              ? '新增一个严格质量候选；后端会自动校验并最多修复 2 次，通过后才替换主视频。'
+                              : '启动严格视频质量流水线；任务在后台生成、校验并有限修复，通过后才晋升主视频。'
                    "
                    primary
                    @confirm="generateVideo"
@@ -104,3 +104,29 @@ export function videoRepairInfo(rawJson?: string | null): { attempt: number } |
        return null
    }
 }
+
+/** 识别严格质量候选及其自动修复配置，防止未通过候选被手动绕过校验设为主视频。 */
+export function videoQualityInfo(
+    rawJson?: string | null
+): { allowedTexts: string[]; maxRepairAttempts: number } | null {
+    try {
+        const snapshot: unknown = JSON.parse(rawJson || '{}')
+        if (!snapshot || typeof snapshot !== 'object' || Array.isArray(snapshot)) return null
+        const record = snapshot as Record<string, unknown>
+        if (record.qualityCandidate !== true) return null
+        const pipeline = record.qualityPipeline
+        if (!pipeline || typeof pipeline !== 'object' || Array.isArray(pipeline)) return null
+        const config = pipeline as Record<string, unknown>
+        const maxRepairAttempts = config.maxRepairAttempts
+        if (typeof maxRepairAttempts !== 'number' || !Number.isSafeInteger(maxRepairAttempts) || maxRepairAttempts <= 0)
+            return null
+        return {
+            allowedTexts: Array.isArray(config.allowedTexts)
+                ? config.allowedTexts.filter((item): item is string => typeof item === 'string')
+                : [],
+            maxRepairAttempts
+        }
+    } catch {
+        return null
+    }
+}
@@ -208,21 +208,45 @@ export interface KeyframeBatchResult {
    failures: { shotId: string; error: string }[]
 }

-/** 项目视频批量生成参数。 */
-export interface GenerateVideosInput {
-    concurrency: number
-    force: boolean
+/** 严格视频质量流水线的共用参数；候选必须校验通过后才会晋升主视频。 */
+export interface VideoQualityConfig {
+    allowedTexts?: string[]
+    maxRepairAttempts: number
 }

-/** 视频批量任务创建回执；created 只表示已提交任务，不表示视频已完成。 */
-export interface VideoBatchResult {
+/** 项目级严格视频质量生成参数；limit 限制本次实际提交的镜头数。 */
+export interface GenerateQualityVideosInput extends VideoQualityConfig {
+    concurrency: number
+    force: boolean
+    limit?: number
+}
+
+/** 单镜头严格视频质量任务回执；candidate 只是异步候选，不代表已经通过校验。 */
+export interface ShotVideoQualityResult {
+    shotId: string
+    qualityPipeline: Required<VideoQualityConfig>
+    candidate: ShotVideo | null
+}
+
+/** 项目级严格视频质量任务回执；后续校验与修复由后端 Poller 自动推进。 */
+export interface VideoQualityBatchResult {
+    projectId: string
+    concurrency: number
+    limit: number | null
+    force: boolean
+    qualityPipeline: Required<VideoQualityConfig>
    total: number
+    readyTargetCount: number
    targetCount: number
+    remainingReady: number
    created: number
    skipped: number
-    readiness: Omit<VideoReadiness, 'total' | 'items'>
+    inProgress: number
+    blocked: number
    failed: number
+    readiness: Omit<VideoReadiness, 'total' | 'items'>
    failures: { shotId: string; error: string }[]
+    candidateIds: string[]
 }

 /** 项目最近一次视频任务的逐镜状态。 */
@@ -260,7 +284,7 @@ export interface RetryVideosResult {
 export type ProductionReceipt =
    | { kind: 'prompts'; title: string; result: import('../storyboard/types').PromptBatchResult }
    | { kind: 'keyframes'; title: string; result: KeyframeBatchResult }
-    | { kind: 'videos'; title: string; result: VideoBatchResult }
+    | { kind: 'videos'; title: string; result: VideoQualityBatchResult }
    | { kind: 'retry'; title: string; result: RetryVideosResult }

 /** 项目生产页的浏览器会话数据。 */
@@ -22,6 +22,8 @@ export function useProduction() {
    const keyframeLimit = ref<number | ''>('')
    const keyframeWidth = ref<number | ''>('')
    const keyframeHeight = ref<number | ''>('')
+    const videoLimit = ref<number | null>(1)
+    const videoRepairAttempts = ref<number | null>(2)
    const snapshot = computed(() => breakdownSnapshot(context.checkpoints.value))
    const sourceEpisodes = computed(
        () => snapshot.value?.storyboardEpisodeShots ?? snapshot.value?.breakdownResult?.storyboardEpisodeShots ?? []
@@ -80,6 +82,14 @@ export function useProduction() {
            (keyframeLimit.value === '' || (Number.isSafeInteger(keyframeLimit.value) && keyframeLimit.value > 0)) &&
            validOptionalSize(keyframeWidth.value, keyframeHeight.value)
    )
+    /** 视频质量流水线要求有限的提交范围和至少一次自动修复额度。 */
+    const videoQualityValid = computed(
+        () =>
+            Number.isSafeInteger(videoLimit.value) &&
+            (videoLimit.value ?? 0) > 0 &&
+            Number.isSafeInteger(videoRepairAttempts.value) &&
+            (videoRepairAttempts.value ?? 0) > 0
+    )
    const blocked = computed(
        () =>
            operation.value.pending ||
@@ -91,7 +101,13 @@ export function useProduction() {

    /** 批量操作只向就绪镜头提交；最终资产状态由轮询查询确认。 */
    async function run(command: ProductionCommand) {
-        if (blocked.value || !batchValid.value || (command === 'keyframes' && !keyframeValid.value)) return
+        if (
+            blocked.value ||
+            !batchValid.value ||
+            (command === 'keyframes' && !keyframeValid.value) ||
+            (command === 'videos' && !videoQualityValid.value)
+        )
+            return
        const projectId = id.value
        const input = { concurrency: concurrency.value, force: force.value }
        const target = getProductionSession(projectId)
@@ -106,10 +122,10 @@ export function useProduction() {
                  ? `更新 ${staleKeyframes} 个过期主首帧`
                  : `补齐第 ${episodeNo.value} 集主首帧`,
            videos: force.value
-                ? '新增项目视频候选'
+                ? '新增项目质量视频候选'
                : staleVideos > 0
                  ? `更新 ${staleVideos} 个过期主视频`
-                  : '提交就绪视频任务',
+                  : '提交视频质量任务',
            'retry-videos': '重试失败视频任务'
        }
        await runOperation(projectId, labels[command], async () => {
@@ -137,7 +153,11 @@ export function useProduction() {
                target.receipt = {
                    kind: 'videos',
                    title: labels[command],
-                    result: await productionApi.generateVideos(projectId, input)
+                    result: await productionApi.generateQualityVideos(projectId, {
+                        ...input,
+                        limit: videoLimit.value ?? undefined,
+                        maxRepairAttempts: videoRepairAttempts.value ?? 0
+                    })
                }
            } else {
                target.receipt = {
@@ -170,7 +190,10 @@ export function useProduction() {
        keyframeLimit,
        keyframeWidth,
        keyframeHeight,
+        videoLimit,
+        videoRepairAttempts,
        keyframeValid,
+        videoQualityValid,
        batchValid,
        blocked,
        run
@@ -65,7 +65,7 @@ function clearFilters() {
 }
 </script>
 <template>
-    <WorkspacePage split>
+    <WorkspacePage split compact>
        <template #header>
            <div class="project-index-chrome">
                <!-- 以下是标题与新建 -->
@@ -224,17 +224,11 @@ watch(
                                                    >导演摄影设计
                                                </h3>
                                                <p class="mt-2 text-xs leading-6 text-muted">
-                                                    景别、机位、运镜与构图。单集操作会重新生成并覆盖；批量默认跳过完整剧集。
+                                                    景别、机位、运镜与构图。正式项目生成默认跳过已完整剧集，覆盖模式会重新生成全部剧集。
                                                </p>
                                            </div>
                                            <div class="flex flex-wrap gap-2">
                                                <ConfirmAction
-                                                    label="生成本集导演设计"
-                                                    :disabled="!allowed['direction-one']"
-                                                    acknowledgement
-                                                    description="调用模型生成当前整集的导演设计，校验通过后保存。如果此集已有设计，将覆盖原结果。"
-                                                    @confirm="run('direction-one')"
-                                                /><ConfirmAction
                                                    :label="force ? '重生成全部导演设计' : '补齐项目导演设计'"
                                                    :disabled="!allowed['direction-all']"
                                                    acknowledgement
@@ -259,12 +253,6 @@ watch(
                                            </div>
                                            <div class="flex flex-wrap gap-2">
                                                <ConfirmAction
-                                                    label="生成本集即时状态"
-                                                    :disabled="!allowed['visual-one']"
-                                                    acknowledgement
-                                                    description="依据本集导演设计和主体形态生成即时视觉状态，按配置尝试自动修复。校验通过才保存，已有状态将被覆盖。"
-                                                    @confirm="run('visual-one')"
-                                                /><ConfirmAction
                                                    :label="force ? '重生成全部即时状态' : '补齐项目即时状态'"
                                                    :disabled="!allowed['visual-all']"
                                                    acknowledgement
@@ -1,6 +1,5 @@
 import { request } from '../../lib/http'
 import type {
-    DirectionGeneration,
    EpisodeDirections,
    EpisodeVisualStates,
    PromptBatchResult,
@@ -8,8 +7,7 @@ import type {
    ShotReferences,
    StoryboardBatchInput,
    StoryboardBatchResult,
-    VisualStateBatchInput,
-    VisualStateGeneration
+    VisualStateBatchInput
 } from './types'
 import type { ShotGenerationSpec } from './generation-spec.types'

@@ -23,30 +21,18 @@ function shotPath(id: string) {
    return `/storyboard-shots/${encodeURIComponent(id)}`
 }

-/** Storyboard 路由契约；单集入口虽名为 generate-test，但 persist=true 才适用于工作台。 */
+/** Storyboard 正式路由契约；测试路由不暴露为前端生产入口。 */
 export const storyboardApi = {
    directions: (id: string, episodeNo: number, signal?: AbortSignal) =>
        request<EpisodeDirections>(`${projectPath(id)}/storyboard-directions?episodeNo=${episodeNo}`, { signal }),
    visualStates: (id: string, episodeNo: number, signal?: AbortSignal) =>
        request<EpisodeVisualStates>(`${projectPath(id)}/storyboard-visual-states?episodeNo=${episodeNo}`, { signal }),
-    generateDirection: (id: string, episodeNo: number) =>
-        request<DirectionGeneration>(`${projectPath(id)}/storyboard-directions/generate-test`, {
-            method: 'POST',
-            body: { episodeNo, persist: true },
-            timeoutMs: 0
-        }),
    generateDirections: (id: string, input: StoryboardBatchInput) =>
        request<StoryboardBatchResult>(`${projectPath(id)}/storyboard-directions/generate`, {
            method: 'POST',
            body: input,
            timeoutMs: 0
        }),
-    generateVisualState: (id: string, episodeNo: number, maxRepairAttempts: number) =>
-        request<VisualStateGeneration>(`${projectPath(id)}/storyboard-visual-states/generate-test`, {
-            method: 'POST',
-            body: { episodeNo, persist: true, maxRepairAttempts },
-            timeoutMs: 0
-        }),
    generateVisualStates: (id: string, input: VisualStateBatchInput) =>
        request<StoryboardBatchResult>(`${projectPath(id)}/storyboard-visual-states/generate`, {
            method: 'POST',
@@ -3,24 +3,13 @@ import { NScrollbar, NAlert, NButton, NTable } from 'naive-ui'
 import { computed } from 'vue'
 import { Download } from '@lucide/vue'
 import { downloadText } from '../../../lib/format'
-import type { StoryboardIssue, StoryboardReceipt } from '../types'
+import type { StoryboardReceipt } from '../types'

 /** 独立展示提交结果，不能将 HTTP 成功或局部成功显示为整批完成。 */
 const props = defineProps<{ receipt: StoryboardReceipt }>()
-const issues = computed<StoryboardIssue[]>(() => {
-    if (props.receipt.kind === 'direction') return props.receipt.result.validation.issues
-    if (props.receipt.kind === 'visual-state')
-        return props.receipt.result.remainingIssues.length
-            ? props.receipt.result.remainingIssues
-            : props.receipt.result.validation.issues
-    return []
-})
-const failed = computed(() => {
-    if (props.receipt.kind === 'batch') return props.receipt.result.failedEpisodes > 0
-    if (props.receipt.kind === 'prompts') return props.receipt.result.failed > 0
-    if (props.receipt.kind === 'visual-state' && !props.receipt.result.persisted) return true
-    return !props.receipt.result.validation.valid
-})
+const failed = computed(() =>
+    props.receipt.kind === 'batch' ? props.receipt.result.failedEpisodes > 0 : props.receipt.result.failed > 0
+)

 /** 导出原始回执，保留后端提供的完整失败诊断。 */
 function exportReceipt() {
@@ -91,24 +80,6 @@ function exportReceipt() {
                当前批量接口不返回成功镜头的 Prompt 正文；选择镜头后点击“读取／生成提示词”，已有内容会直接复用。
            </p>
        </template>
-        <template v-else>
-            <p class="mt-4 text-sm" :class="failed ? 'text-danger' : 'text-success'">
-                {{ receipt.result.validation.valid ? '生成校验通过' : '校验失败，未保存本次生成结果'
-                }}<span v-if="receipt.kind === 'visual-state'">
-                    · 自动修复 {{ receipt.result.repairAttempts }} 次 ·
-                    {{ receipt.result.persisted ? '已持久化' : '未持久化' }}</span
-                >
-            </p>
-            <p v-if="receipt.kind === 'direction' && receipt.result.validation.valid" class="mt-2 text-xs text-muted">
-                已请求持久化，数据库实际覆盖率以下方刷新后的数据为准。
-            </p>
-            <ul class="mt-3 space-y-2 text-xs text-danger">
-                <li v-for="(issue, index) in issues" :key="index">
-                    第 {{ issue.episodeNo }} 集 / Beat {{ issue.beatNo ?? '—' }} / Shot {{ issue.shotNo ?? '—' }}
-                    {{ issue.subjectRef }}：{{ issue.message }}
-                </li>
-            </ul>
-        </template>
        <p class="mt-4 text-[11px] text-muted">回执仅保留在当前浏览器会话中。数据库结果是覆盖率和镜头详情的依据。</p>
    </section>
 </template>
@@ -23,7 +23,7 @@ export interface SubjectVisualState {
    transientState: string
 }

-/** GET 返回扁平化的环境字段，与 generate-test 的嵌套 environment 不同。 */
+/** GET 返回已经持久化的扁平化环境字段。 */
 export interface ShotVisualState {
    timeOfDay: string
    weather: string
@@ -62,33 +62,6 @@ export interface StoryboardIssue {
    message: string
 }

-/** 单集生成结果中的校验信息，HTTP 200 不代表 valid。 */
-export interface StoryboardValidation {
-    valid: boolean
-    issues: StoryboardIssue[]
-}
-
-/** 单集导演设计返回领域结构；是否落库应再查 GET 覆盖率确认。 */
-export interface DirectionGeneration {
-    episodeNo: number
-    direction: {
-        episodeNo: number
-        beatDirections: { beatNo: number; shots: (ShotDirection & { shotNo: number })[] }[]
-    }
-    validation: StoryboardValidation
-}
-
-/** 单集 VisualState 的自动修复与落库回执。 */
-export interface VisualStateGeneration {
-    episodeNo: number
-    visualState: unknown
-    validation: StoryboardValidation
-    repaired: boolean
-    repairAttempts: number
-    remainingIssues: StoryboardIssue[]
-    persisted: boolean
-}
-
 /** 项目批量生成中的单集结果；保留失败原因和修复诊断。 */
 export interface EpisodeGenerationResult {
    episodeNo: number
@@ -178,8 +151,6 @@ export interface DesignedShot {

 /** 生成回执按操作区分，切换剧集时仍保留原始目标信息。 */
 export type StoryboardReceipt =
-    | { kind: 'direction'; title: string; episodeNo: number; result: DirectionGeneration }
-    | { kind: 'visual-state'; title: string; episodeNo: number; result: VisualStateGeneration }
    | { kind: 'batch'; title: string; result: StoryboardBatchResult }
    | { kind: 'prompts'; title: string; result: PromptBatchResult }

@@ -6,8 +6,8 @@ import { getOperation, runOperation } from '../workflows/operations'
 import { storyboardApi } from './api'
 import { getStoryboardSession, mergeDesignedShots, storyboardPrerequisites } from './model'

-/** 分镜页允许的生成命令，保持单集与批量参数边界。 */
-export type StoryboardCommand = 'direction-one' | 'direction-all' | 'visual-one' | 'visual-all' | 'prompts'
+/** 分镜页只提供正式项目级生成命令，不暴露后端测试路由。 */
+export type StoryboardCommand = 'direction-all' | 'visual-all' | 'prompts'

 /** 分镜查询、依赖判断与操作入口；不修改后端 graph。 */
 export function useStoryboard() {
@@ -87,10 +87,7 @@ export function useStoryboard() {
            data.value.directions.coverage === 1
    )
    const allowed = computed<Record<StoryboardCommand, boolean>>(() => ({
-        'direction-one': !blocked.value && prerequisites.value.directionEpisode && !!data.value?.directions.shotCount,
        'direction-all': !blocked.value && prerequisites.value.directionProject && batchValid.value,
-        'visual-one':
-            !blocked.value && prerequisites.value.visualEpisode && directionComplete.value && repairsValid.value,
        'visual-all': !blocked.value && prerequisites.value.visualProject && batchValid.value && repairsValid.value,
        prompts: !blocked.value && shots.value.length > 0 && batchValid.value
    }))
@@ -99,28 +96,17 @@ export function useStoryboard() {
    async function run(command: StoryboardCommand) {
        if (!allowed.value[command]) return
        const projectId = id.value
-        const number = episodeNo.value
        const batch = { concurrency: concurrency.value, force: force.value }
        const repairs = maxRepairAttempts.value
        const target = getStoryboardSession(projectId)
        const titles: Record<StoryboardCommand, string> = {
-            'direction-one': `第 ${number} 集导演设计`,
            'direction-all': '项目导演设计',
-            'visual-one': `第 ${number} 集即时状态`,
            'visual-all': '项目即时状态',
            prompts: '项目视频提示词'
        }
        target.receipt = null
        await runOperation(projectId, `生成${titles[command]}`, async () => {
            switch (command) {
-                case 'direction-one':
-                    target.receipt = {
-                        kind: 'direction',
-                        title: titles[command],
-                        episodeNo: number,
-                        result: await storyboardApi.generateDirection(projectId, number)
-                    }
-                    break
                case 'direction-all':
                    target.receipt = {
                        kind: 'batch',
@@ -128,14 +114,6 @@ export function useStoryboard() {
                        result: await storyboardApi.generateDirections(projectId, batch)
                    }
                    break
-                case 'visual-one':
-                    target.receipt = {
-                        kind: 'visual-state',
-                        title: titles[command],
-                        episodeNo: number,
-                        result: await storyboardApi.generateVisualState(projectId, number, repairs)
-                    }
-                    break
                case 'visual-all':
                    target.receipt = {
                        kind: 'batch',
@@ -545,7 +545,7 @@ watch(
                            @save="save"
                            @dirty="dirty = $event"
                        />
-                        <div v-if="anchor" class="mt-5 flex items-center gap-4 pt-5">
+                        <div v-if="anchor" class="identity-anchor mt-5 flex items-center gap-4 pt-5">
                            <AssetImage
                                :src="anchor.imageUrl"
                                :alt="`${subject.name}当前身份母版`"
@@ -677,6 +677,25 @@ watch(
 .identity-workspace > div {
    @apply min-h-0 overflow-hidden overscroll-contain;
 }
+.identity-detail,
+.identity-detail > .n-scrollbar-container,
+.identity-detail .n-scrollbar-content,
+.identity-detail .panel {
+    @apply w-full min-w-0 max-w-full;
+}
+.identity-detail .n-scrollbar-content {
+    @apply wrap-anywhere;
+}
+.identity-detail .panel {
+    @apply box-border;
+}
+/* AssetImage 的通用 width:100% 不得撑满母版摘要行，否则会把说明文字挤到最右侧。 */
+.identity-anchor > .asset-image {
+    @apply h-20 w-20 shrink-0 aspect-square;
+}
+.identity-anchor > div {
+    @apply min-w-0;
+}
 .identity-subject-list {
    @apply max-h-none min-h-0 flex-1;
 }
@@ -305,8 +305,8 @@ async function showStaleForms() {

 <template>
    <WorkspacePage compact class="gallery-workspace-page"
-        ><template #default>
-            <!-- 筛选、关联说明和图片共用正文滚动区，避免镜头切换时两块滚动区域高度不同步。 -->
+        ><template #header>
+            <!-- 图库统计和固定操作使用与其他工作台一致的紧凑头部。 -->
            <div class="workspace-toolbar">
                <span class="toolbar-summary">主图 {{ completeCount }} / {{ forms.length }}</span>
                <NButton v-if="staleForms.length" text type="error" size="small" @click="toolsOpen = true"
@@ -476,7 +476,9 @@ async function showStaleForms() {
                        </section>
                    </WorkspaceTools>
                </div>
-            </div>
+            </div></template
+        ><template #default>
+            <!-- 关联说明、筛选和图片共用正文滚动区，避免镜头切换时两块滚动区域高度不同步。 -->
            <NAlert v-if="staleForms.length" type="error" :show-icon="false" class="mt-3" title="形态主图需要更新">
                {{ staleForms.length }}
                个形态主图与当前已锁定身份母版不一致。请生成候选、验图并切换主图，再重建相关首帧。
@@ -851,7 +853,7 @@ async function showStaleForms() {
    @apply [overflow-anchor:none];
 }
 .form-image-filter-region {
-    @apply my-4 py-3.5 px-4 bg-(--app-subtle);
+    @apply my-4 py-[5px] px-4 bg-(--app-subtle);
    container-type: inline-size;
 }
 .form-image-toolbar {
@@ -162,7 +162,7 @@ describe('全站平面主题', () => {
        )
        expect(css).toMatch(/\.asset-impact-links-scroll\.n-scrollbar\s*\{[^}]*height:\s*auto;[^}]*max-width:\s*100%;/)
        expect(readFileSync('src/features/subject-images/SubjectImagesPage.vue', 'utf8')).toMatch(
-            /<WorkspacePage\b[^>]*>\s*<template #default>/
+            /<WorkspacePage\b[^>]*>\s*<template #header>/
        )
    })

@@ -184,7 +184,7 @@ describe('全站平面主题', () => {
        // DOM 环境不计算几何；保护容器宽度断点和居中契约，真实坐标仍需浏览器验收。
        const css = readAllStyles()
        expect(css).toMatch(
-            /\.form-image-filter-region\s*\{[^}]*padding(?:-block)?:\s*14px;[^}]*background:\s*var\(--app-subtle\);/
+            /\.form-image-filter-region\s*\{[^}]*padding(?:-block)?:\s*5px;[^}]*background:\s*var\(--app-subtle\);/
        )
        expect(css).toMatch(/\.form-image-toolbar\s*\{[^}]*align-items:\s*center;[^}]*gap:\s*12px 16px;/)
        expect(css).toMatch(
@@ -99,12 +99,19 @@ describe('镜头生产数据契约', () => {
        })
        vi.stubGlobal('fetch', fetcher)
        await productionApi.videoReadiness('project/1', true)
-        await productionApi.generateVideos('project/1', { concurrency: 3, force: true })
+        await productionApi.generateQualityVideos('project/1', {
+            concurrency: 3,
+            force: true,
+            limit: 2,
+            maxRepairAttempts: 2
+        })
        expect(fetcher.mock.calls[0]?.[0]).toBe('/api/projects/project%2F1/videos/readiness?force=true')
-        expect(fetcher.mock.calls[1]?.[0]).toBe('/api/projects/project%2F1/videos/generate')
+        expect(fetcher.mock.calls[1]?.[0]).toBe('/api/projects/project%2F1/videos/generate-quality')
        expect(JSON.parse(String(fetcher.mock.calls[1]?.[1]?.body))).toEqual({
            concurrency: 3,
-            force: true
+            force: true,
+            limit: 2,
+            maxRepairAttempts: 2
        })
    })

@@ -181,7 +181,7 @@ describe('过期首帧到具体素材定位', () => {

    it('切换镜头时筛选、关联说明和图片共用稳定纵向容器，长关联列表单独横向滚动', async () => {
        const { keyframes, videos, references, fetcher } = await gallery('?sourceShotId=shot-target')
-        expect(wrapper!.find('.workspace-heading-scroll').exists()).toBe(false)
+        expect(wrapper!.find('.workspace-heading-scroll').exists()).toBe(true)
        const scroll = wrapper!.get<HTMLElement>('.workspace-scroll > .n-scrollbar-container').element
        const content = wrapper!.get('.workspace-scroll-content').element
        const sticky = wrapper!.get('.gallery-sticky-controls').element
@@ -1111,7 +1111,7 @@ describe('工作台页面交互', () => {
        vi.stubGlobal('fetch', fetcher)
        await mountStoryboard()
        await flushPromises()
-        await confirmGeneration('生成本集导演设计')
+        await confirmGeneration('补齐项目导演设计')
        expect(button('补齐项目导演设计').disabled).toBe(true)
        await drawerPanel().get('[aria-label="关闭操作面板"]').trigger('click')
        await flushPromises()
@@ -1123,61 +1123,69 @@ describe('工作台页面交互', () => {
        wrapper = undefined
        finish(
            new Response(
-                '{"data":{"episodeNo":1,"direction":{"episodeNo":1,"beatDirections":[]},"validation":{"valid":true,"issues":[]}}}'
+                `{"data":{"projectId":"${fixture.id}","episodeCount":2,"completedEpisodes":1,"skippedEpisodes":1,"failedEpisodes":0,"shotCount":4,"directionCount":4,"coverage":1,"episodes":[]}}`
            )
        )
        await flushPromises()
        expect(getStoryboardSession(fixture.id).receipt).toMatchObject({
-            kind: 'direction',
-            episodeNo: 1,
-            title: '第 1 集导演设计'
+            kind: 'batch',
+            title: '项目导演设计'
        })
        expect(getOperation(fixture.id).pending).toBe(false)
        const posts = fetcher.mock.calls.filter(([, init]) => init?.method === 'POST')
        expect(posts).toHaveLength(1)
-        expect(JSON.parse(posts[0]![1]!.body as string)).toEqual({ episodeNo: 1, persist: true })
+        expect(JSON.parse(posts[0]![1]!.body as string)).toEqual({ concurrency: 2, force: false })
    })

-    it('分镜必须先补齐导演设计，校验失败显示未保存，修复次数 0 不被默认值覆盖', async () => {
-        let complete = false
+    it('分镜只提供正式项目生成，逐集失败回执保留零次修复参数', async () => {
        const fetcher = vi.fn<typeof fetch>().mockImplementation(async (url, init) => {
            if (init?.method === 'POST')
                return new Response(
                    JSON.stringify({
                        data: {
+                            projectId: fixture.id,
+                            episodeCount: 1,
+                            completedEpisodes: 0,
+                            skippedEpisodes: 0,
+                            failedEpisodes: 1,
+                            shotCount: 2,
+                            visualStateCount: 0,
+                            coverage: 0,
+                            episodes: [
+                                {
                                    episodeNo: 1,
-                            visualState: {},
-                            validation: {
-                                valid: false,
-                                issues: [{ episodeNo: 1, beatNo: 1, shotNo: 1, message: '缺少主体状态' }]
-                            },
-                            repaired: false,
+                                    status: 'failed',
+                                    shotCount: 2,
+                                    visualStateCount: 0,
                                    repairAttempts: 0,
-                            remainingIssues: [],
-                            persisted: false
+                                    error: '缺少主体状态'
+                                }
+                            ]
                        }
                    })
                )
            const result = String(url).includes('storyboard-directions')
-                ? directionsResult(fixture.id, 1, complete)
+                ? directionsResult(fixture.id, 1, true)
                : visualStatesResult(fixture.id)
            return new Response(JSON.stringify({ data: result }))
        })
        vi.stubGlobal('fetch', fetcher)
        await mountStoryboard()
        await flushPromises()
-        expect(button('生成本集导演设计').disabled).toBe(false)
-        expect(button('生成本集即时状态').disabled).toBe(true)
+        expect(wrapper!.text()).not.toContain('生成本集导演设计')
+        expect(wrapper!.text()).not.toContain('生成本集即时状态')
+        expect(button('补齐项目导演设计').disabled).toBe(false)
+        expect(button('补齐项目即时状态').disabled).toBe(false)
        expect(fetcher.mock.calls.every(([, init]) => init?.method !== 'POST')).toBe(true)
-        complete = true
-        button('刷新分镜').click()
-        await flushPromises()
-        expect(button('生成本集即时状态').disabled).toBe(false)
        await drawerPanel().get('#storyboard-repairs').setValue('0')
-        await confirmGeneration('生成本集即时状态')
+        await confirmGeneration('补齐项目即时状态')
        const post = fetcher.mock.calls.find(([, init]) => init?.method === 'POST')!
-        expect(JSON.parse(post[1]!.body as string)).toEqual({ episodeNo: 1, persist: true, maxRepairAttempts: 0 })
-        expect(drawerPanel().text()).toContain('校验失败，未保存本次生成结果')
+        expect(JSON.parse(post[1]!.body as string)).toEqual({
+            concurrency: 2,
+            force: false,
+            maxRepairAttempts: 0
+        })
+        expect(drawerPanel().text()).toContain('本次存在失败')
        expect(drawerPanel().text()).toContain('缺少主体状态')
    })

@@ -1291,11 +1299,11 @@ describe('工作台页面交互', () => {
        await flushPromises()
        expect(wrapper!.text()).toContain('shot-db-1-1')
        expect(wrapper!.text()).toContain('数据库暂不可用')
-        expect(button('生成本集导演设计').disabled).toBe(true)
+        expect(button('补齐项目导演设计').disabled).toBe(true)
        fail = false
        button('重试查询').click()
        await flushPromises()
-        expect(button('生成本集导演设计').disabled).toBe(false)
+        expect(button('补齐项目导演设计').disabled).toBe(false)
    })

    it('镜头工具按需 GET，切换镜头会废弃旧参考图，未完成设计不能编译规格', async () => {
@@ -43,21 +43,13 @@ describe('后端 API 契约', () => {
        })
        expect(fetcher.mock.calls[4]![1]!.body).toBeUndefined()
    })
-    it('分镜单集显式持久化，批量保持 force 和零次修复参数', async () => {
+    it('分镜只调用正式项目接口，并保持 force 和零次修复参数', async () => {
        const fetcher = vi.fn<typeof fetch>().mockImplementation(async () => new Response('{"data":{}}'))
        vi.stubGlobal('fetch', fetcher)
-        await storyboardApi.generateDirection('a/b', 2)
-        await storyboardApi.generateDirections('p', { concurrency: 2, force: false })
-        await storyboardApi.generateVisualState('p', 2, 0)
+        await storyboardApi.generateDirections('a/b', { concurrency: 2, force: false })
        await storyboardApi.generateVisualStates('p', { concurrency: 3, force: true, maxRepairAttempts: 0 })
        expect(fetcher.mock.calls.map(([url, init]) => [url, init?.method, JSON.parse(init!.body as string)])).toEqual([
-            ['/api/projects/a%2Fb/storyboard-directions/generate-test', 'POST', { episodeNo: 2, persist: true }],
-            ['/api/projects/p/storyboard-directions/generate', 'POST', { concurrency: 2, force: false }],
-            [
-                '/api/projects/p/storyboard-visual-states/generate-test',
-                'POST',
-                { episodeNo: 2, persist: true, maxRepairAttempts: 0 }
-            ],
+            ['/api/projects/a%2Fb/storyboard-directions/generate', 'POST', { concurrency: 2, force: false }],
            [
                '/api/projects/p/storyboard-visual-states/generate',
                'POST',
@@ -86,7 +78,7 @@ describe('后端 API 契约', () => {
        expect(JSON.parse(fetcher.mock.calls[4]![1]!.body as string)).toEqual({ force: false })
    })

-    it('分镜生成保留 200 中的校验失败和持久化标志，长请求不自动超时或重试', async () => {
+    it('分镜正式批量生成保留部分失败回执，长请求不自动超时或重试', async () => {
        vi.useFakeTimers()
        let finish!: (response: Response) => void
        const fetcher = vi.fn<typeof fetch>().mockImplementation(
@@ -96,12 +88,16 @@ describe('后端 API 契约', () => {
                })
        )
        vi.stubGlobal('fetch', fetcher)
-        const pending = storyboardApi.generateVisualState('p', 1, 2)
+        const pending = storyboardApi.generateVisualStates('p', {
+            concurrency: 1,
+            force: false,
+            maxRepairAttempts: 2
+        })
        await vi.advanceTimersByTimeAsync(120_000)
        expect(fetcher).toHaveBeenCalledOnce()
        expect(fetcher.mock.calls[0]?.[1]?.signal?.aborted).toBe(false)
-        finish(new Response('{"data":{"validation":{"valid":false,"issues":[]},"persisted":false}}'))
-        await expect(pending).resolves.toMatchObject({ validation: { valid: false }, persisted: false })
+        finish(new Response('{"data":{"failedEpisodes":1,"episodes":[{"episodeNo":1,"status":"failed"}]}}'))
+        await expect(pending).resolves.toMatchObject({ failedEpisodes: 1, episodes: [{ status: 'failed' }] })
    })
    it('解包 data，同时保留创建项目的顶层 202 结构', async () => {
        const fetcher = vi