docs: restore English portfolio and improve platform showcase

Author: jiapengLi11

README.en.md

@@ -2,7 +2,7 @@
 
 > 中文优先的工程作品集：[文档总入口](docs/README.zh-CN.md) | [15 分钟公开演示](docs/REPRODUCIBLE_DEMO.zh-CN.md) | [核心代码导读](docs/CODE_WALKTHROUGH.zh-CN.md) | [项目证据与边界](docs/PROJECT_EVIDENCE.zh-CN.md)
 
-[简体中文](README.md) | [English](README.en.md) | [完整中文长版](README.zh-CN.md)
+[English](README.md) | [简体中文](README.zh-CN.md) | [中文文档总入口](docs/README.zh-CN.md)
 
 [![CI](https://github.com/jiapengLi11/yolo-label-recovery/actions/workflows/ci.yml/badge.svg)](https://github.com/jiapengLi11/yolo-label-recovery/actions/workflows/ci.yml)
 [![Release](https://img.shields.io/github/v/release/jiapengLi11/yolo-label-recovery)](https://github.com/jiapengLi11/yolo-label-recovery/releases)
@@ -27,7 +27,7 @@
 
 项目的核心不是“用模型自动生成更多框”，而是建立一条风险可控的证据链：模型只提出候选，规则解释关系，人工授权决策，数据库保证协作一致性，安全写回生成可回滚的数据版本，固定测试集最终判断模型是否真的变好。
 
-![真实多人审核平台](docs/assets/platform-review-productivity.png)
+![真实五候选联合场景多人审核平台](docs/assets/platform-review-multibox.png)
 
 ## 一张图看完整链路
 
@@ -118,14 +118,11 @@ flowchart LR
 
 ![真实多人审核进度看板](docs/assets/platform-dashboard.png)
 
-<p>
-  <img src="docs/assets/platform-review.png" width="49%" alt="真实联合场景 Web 审核工作台">
-  <img src="docs/assets/platform-admin.png" width="49%" alt="审核账号与项目分配面板">
-</p>
+![审核账号与项目分配面板](docs/assets/platform-admin.png)
 
-![真实一键审核、租约状态与最近决定纠错流程](docs/assets/platform-review-productivity.png)
+以上均为 Vue + Spring Boot 应用的真实运行截图，不是设计示意图。首页主视觉特意选用一张包含 `3 person + 1 vest + 1 tractor` 共五个候选的真实地下场景，因此左侧候选列表、完整图片画布、租约状态、置信度、受约束决策、快捷键和释放任务入口可在同一屏完整展示。登录页展示角色入口，管理员界面展示账号创建与项目成员分配能力。
 
-以上均为 Vue + Spring Boot 应用的真实运行截图，不是设计示意图。登录页展示审核人员实际使用的角色入口；工作台同时展示已领取任务、真实审核图、类别置信度、受约束决策按钮和项目实时进度；管理员界面展示账号创建与项目成员分配能力。
+早期 Tk 桌面审核器与 Web 平台并非重复实现：桌面版适合离线交付和单人审核，Web 版进一步解决多人身份、原子领单、租约续期、项目隔离、版本冲突和不可变审计问题。
 
 ```mermaid
 flowchart LR

README.md

@@ -15,11 +15,11 @@ The original Tk reviewer remains the best zero-dependency choice for one reviewe
   <img src="assets/platform-admin.png" width="49%" alt="Admin account and membership management">
 </p>
 
-![Real joint-scene review workspace / 真实联合场景审核工作台](assets/platform-review.png)
+![Real five-candidate joint-scene review workspace / 真实五候选联合场景审核工作台](assets/platform-review-multibox.png)
 
-The screenshots above were captured from the actual Vue and Spring Boot services against a real, user-approved review package. They demonstrate the real login flow, project progress, atomic task claim, read-only review visual delivery, constrained decisions, account creation and member assignment. They are not UI mockups and do not represent model-accuracy evidence.
+The screenshots above were captured from the actual Vue and Spring Boot services against a real, user-approved review package. The primary workspace contains five grouped candidates from one underground image (`3 person + 1 vest + 1 tractor`) and exposes the complete candidate rail, image canvas, lease state, decision panel, shortcuts and release path. They demonstrate the real login flow, project progress, atomic task claim, read-only review visual delivery, constrained decisions, account creation and member assignment. They are not UI mockups and do not represent model-accuracy evidence.
 
-以上截图由真实运行的 Vue 与 Spring Boot 服务生成，并加载了经许可使用的真实审核包。截图覆盖登录、项目进度、原子领取任务、只读审核图加载、受约束决策、账号创建和成员分配，不是界面示意图，也不作为模型精度证据。
+以上截图由真实运行的 Vue 与 Spring Boot 服务生成，并加载了经许可使用的真实审核包。主工作台在一张地下场景原图中按图聚合展示 `3 person + 1 vest + 1 tractor` 共五个候选，并完整保留候选列表、图片画布、租约、决策区、快捷键和释放入口。截图覆盖登录、项目进度、原子领取任务、只读审核图加载、受约束决策、账号创建和成员分配，不是界面示意图，也不作为模型精度证据。
 
 ## 3. Architecture / 架构
 

README.zh-CN.md

@@ -65,9 +65,9 @@ OOM 时 Batch 32 → 16 → 8 动态回退
 
 最初的“选择决定 → 再确认 → 手动下一条”看起来更安全，但每条任务增加两次操作，审核量一大就严重拖慢效率。最终工作台采用一键提交、自动前进，并以“最近审核可修订”补偿误操作。
 
-![真实多人审核工作台](assets/platform-review.png)
+![真实五候选多人审核工作台](assets/platform-review-multibox.png)
 
-![审核生产力设计](assets/platform-review-productivity.png)
+该工作台在同一真实地下场景中聚合展示五个候选框，并完整呈现租约、决策区、快捷键和释放任务入口。
 
 同一图片的候选框在左侧聚合显示，审核员可以在联合场景中判断 person、helmet、smoking 是否属于合理嵌套，而不是在割裂的截图里逐框猜测。图片支持适应窗口、原始尺寸和缩放，但默认不铺满整个画布，避免工具栏和上下文被挤出屏幕。
 

docs/COLLABORATION_PLATFORM.md

@@ -39,7 +39,7 @@ The difficult part was operational reliability on a long `K x N` scan. I added t
 7. Open `examples/near_duplicates/output/near_duplicate_report.html` to show review compression and split leakage.
 8. Open `examples/prioritization/output/prioritization_report.html` to show a limited-budget six-class review queue.
 9. Run `yolo-label-recovery doctor` to show environment diagnostics.
-10. Open `docs/assets/platform-dashboard.png` and `docs/assets/platform-review.png` to explain the validated two-account web workflow without requiring a live server.
+10. Open `docs/assets/platform-dashboard.png` and `docs/assets/platform-review-multibox.png` to explain the validated two-account web workflow and five-candidate grouped review without requiring a live server.
 
 This demonstration works without a GPU or private model weights. A full teacher scan remains an optional second demonstration when suitable public weights and data are available.
 

docs/MINING_SAFETY_AI_ENGINEERING_BLOG.zh-CN.md

@@ -40,7 +40,7 @@ The roadmap separates implemented behavior from ideas that still need evidence.
 
 - Installable CLI, dependency-light smoke tests, pytest suite, wheel build and GitHub Actions CI.
 - Public synthetic fixtures and pre-generated reports for audit, calibration, consensus, near duplicates, prioritization and human review.
-- Chinese-first GitHub landing page with a complete English mirror.
+- English-first GitHub landing page with a complete Chinese engineering guide.
 - Documentation hub, reproducible demo guide, core-code walkthrough, evidence matrix, mining-system design and interview Q&A.
 - Repository-local Markdown link validation and structured issue/pull-request templates.
 - Illustrated project blog and a generated development/interview handbook.
@@ -52,7 +52,7 @@ The roadmap separates implemented behavior from ideas that still need evidence.
 - Export of reviewer throughput, disagreement and correction metrics with privacy-safe aggregation.
 - MySQL Testcontainers concurrency tests for claim, lease expiry and stale-version conflicts.
 - Versioned review-decision export manifest and writeback change summary.
-- English versions of the new Chinese-first code, demo and evidence guides.
+- English versions of the detailed Chinese code, demo and evidence guides.
 
 ## Full mining-system evolution
 

docs/PORTFOLIO_GUIDE.md

@@ -1148,7 +1148,7 @@ def build_document(output_path: Path = OUT):
     chapter(doc, "16", "Vue 审核工作台：状态、心跳与对象 URL", "能解释前端如何保持任务租约、加载受保护图片并约束决策。")
     add_logic_bridge(doc, "后端已提供安全、事务化的任务 API", "把异步请求、租约和图片资源组合成不易出错的用户流程", "将前后端一起打包部署并验证网络路径")
     add_picture(doc, ROOT / "docs/assets/platform-dashboard.png", 6.45, "图 7  真实多人审核进度看板", "真实运行平台的候选总量、待审核、协作占用、疑难升级和完成率")
-    add_picture(doc, ROOT / "docs/assets/platform-review.png", 6.45, "图 8  真实联合场景 Web 审核工作台", "真实审核图、候选状态、置信度、建议动作、租约和受约束决策按钮")
+    add_picture(doc, ROOT / "docs/assets/platform-review-multibox.png", 6.45, "图 8  真实五候选联合场景 Web 审核工作台", "同一地下场景聚合展示三个人员、一个背心和一个拖拉机候选，并完整呈现租约、置信度、快捷键和受约束决策")
     add_picture(doc, ROOT / "docs/assets/platform-admin.png", 6.45, "图 9  管理员账号与项目成员分配", "管理员创建独立审核账号并把账号分配到指定项目")
     add_code(doc, "platform/frontend/src/api.ts", 108, 137, "request 包装器集中处理 localStorage JWT、Authorization 头、204 和错误响应，并提供最近审核等强类型接口。")
     add_code(doc, "platform/frontend/src/App.vue", 63, 75, "decisionOptions 根据后端推荐动作动态生成按钮，但它只是体验层，不能替代服务端校验。")
@@ -1540,7 +1540,7 @@ def build_document(output_path: Path = OUT):
     add_callout(doc, "自测标准", "不看手册，用 15 分钟把 mine_000123.jpg 从源标签讲到新版模型评测，并在每个阶段说出一个失败分支和一个验证证据。能做到，前后逻辑才真正连起来。", "tip")
 
     chapter(doc, "26", "用户反馈驱动的审核生产力迭代", "从真实审核速度、误操作恢复和前端可观测性出发，把功能堆叠改造成可持续使用的生产工作台。")
-    add_picture(doc, ROOT / "docs" / "assets" / "platform-review-productivity.png", 6.6, "图 13  一键审核、租约状态与最近决定纠错", "真实运行平台：当前任务、最近审核、连接状态和按图聚合候选同时可见")
+    add_picture(doc, ROOT / "docs" / "assets" / "platform-review-multibox.png", 6.6, "图 13  五候选按图聚合审核与租约状态", "真实运行平台：候选列表、完整图片、连接状态、受约束决策和快捷键同时可见")
     add_callout(doc, "本轮最重要的产品结论", "审核员每天处理的是大量重复判断。主流程每增加一次确认或一次手动跳转，都会按候选数量线性放大成本。安全设计必须保护数据，但不能无证据地牺牲吞吐。", "info")
 
     doc.add_heading("26.1 为什么‘先选择、再确认、再下一条’被回退", level=2)
@@ -1737,15 +1737,15 @@ def build_document(output_path: Path = OUT):
     ])
     add_callout(doc, "STAR 收束", "Situation：多源六类数据联合场景漏标。Task：降低人工成本且不污染原标签。Action：六 Teacher、几何分流、流式推理、多人审核、事务与审计。Result：形成可运行、可追溯、可继续评测的数据闭环；模型收益仍以固定测试集和现场专项集为准。", "tip")
 
-    chapter(doc, "36", "把 GitHub 变成中文优先的面试作品集", "让面试官在 30 秒、5 分钟、15 分钟和 30 分钟四种停留时间内都能看到完整证据。")
-    add_paragraph(doc, "大厂面试官通常不会从仓库第一行读到最后一行。首页必须先回答：解决什么真实问题、最难的工程矛盾是什么、有哪些真实运行证据、如何在没有私有数据和 GPU 的情况下复现、哪些能力仍是下一阶段设计。当前仓库因此使用中文默认 README，英文作为完整镜像，并增加文档总入口。")
+    chapter(doc, "36", "把 GitHub 变成面向国际面试官的工程作品集", "让面试官在 30 秒、5 分钟、15 分钟和 30 分钟四种停留时间内都能看到完整证据。")
+    add_paragraph(doc, "大厂面试官通常不会从仓库第一行读到最后一行。首页必须先回答：解决什么真实问题、最难的工程矛盾是什么、有哪些真实运行证据、如何在没有私有数据和 GPU 的情况下复现、哪些能力仍是下一阶段设计。当前仓库使用英文作为 GitHub 默认首页，降低国际面试官的阅读门槛；中文长版、代码导读和学习手册则负责承载更深入的原理、实验与面试复盘。")
     add_table(doc, ["停留时间", "首页应提供的证据", "面试官形成的判断"], [
         ["30 秒", "一句话价值、技术栈、真实审核界面", "不是孤立脚本，而是完整工程作品"],
         ["5 分钟", "证据矩阵、完成度边界、生产规模数字", "知道你没有把规划冒充落地"],
         ["15 分钟", "无 GPU 演示、预期输出和失败验收", "第三方能够复现关键机制"],
         ["30 分钟", "核心代码导读、事务时序和数据不变量", "能够接受后端、算法和系统设计深挖"],
     ], [1500, 3900, 3960])
-    add_picture(doc, ROOT / "docs" / "assets" / "platform-review-productivity.png", 6.65, "图 16  GitHub 首页使用的真实多人审核工作台", "真实审核界面、同图候选、一键决策、租约和最近审核")
+    add_picture(doc, ROOT / "docs" / "assets" / "platform-review-multibox.png", 6.65, "图 16  GitHub 首页使用的真实多人审核工作台", "真实地下场景、五个同图候选、一键决策、租约、快捷键和释放任务入口")
     add_paragraph(doc, "文档中心按目标而非文件名组织：快速理解、公开复现、核心代码、协作平台、证据边界、完整矿区系统和面试问答。这样首页负责导航，专题文档负责深度，学习手册负责把知识、代码和实验连接起来。")
     add_table(doc, ["文档", "解决的问题", "面试用途"], [
         ["README.md", "项目是什么、为什么值得继续看", "开场与作品展示"],
@@ -1946,7 +1946,7 @@ def build_document(output_path: Path = OUT):
 
     chapter(doc, "附录 E", "现场演示与面试速查", "在联调、答辩或面试前十分钟快速确认环境、入口、证据和边界。")
     add_table(doc, ["场景", "先打开/执行", "必须确认"], [
-        ["GitHub 展示", "README.md -> 文档总入口", "中文首页、截图、CI 和相对链接正常"],
+        ["GitHub 展示", "README.md -> README.zh-CN.md -> 文档总入口", "英文默认首页、中文深度手册、真实截图、CI 和相对链接正常"],
         ["无 GPU 演示", "create_synthetic_dataset + audit", "预期 FAIL 且问题数量符合 fixture"],
         ["审核演示", "review-build + review-ui", "同图候选、允许动作、自动保存和恢复"],
         ["多人平台", "登录 -> 项目 -> 领取 -> 决策 -> 最近审核", "两个账号不重复领取、纠错有审计"],

docs/ROADMAP.md

@@ -8,10 +8,9 @@ This directory turns the offline review queue into a multi-user, auditable web w
 
 ![Real review dashboard](../docs/assets/platform-dashboard.png)
 
-<p>
-  <img src="../docs/assets/platform-review.png" width="49%" alt="Real review task">
-  <img src="../docs/assets/platform-admin.png" width="49%" alt="Account administration">
-</p>
+![Real five-candidate review workspace](../docs/assets/platform-review-multibox.png)
+
+![Account administration](../docs/assets/platform-admin.png)
 
 ## Components / 目录
 
@@ -38,9 +37,7 @@ The UI-library decision is documented in [ADR 0004](../docs/adr/0004-custom-revi
 
 ## Docker deployment / Docker 部署
 
-![Running review workspace](../docs/assets/platform-review-workspace.png)
-
-The image above is a capture of the running application with a real review visual and seeded collaboration records, not a mockup. Additional login and administration screenshots are available in [COLLABORATION_PLATFORM.md](../docs/COLLABORATION_PLATFORM.md).
+The workspace above is a capture of the running application with one real underground image and five grouped candidates, not a mockup. Additional login, dashboard and administration screenshots are available in [COLLABORATION_PLATFORM.md](../docs/COLLABORATION_PLATFORM.md).
 
 ```powershell
 cd platform