diff --git a/.dockerignore b/.dockerignore
index 12ecee6..d6cd567 100644
--- a/.dockerignore
+++ b/.dockerignore
@@ -3,6 +3,7 @@ target
 data
 work
 outputs
+ci-output
 .env
 *.db
 *.db-shm
diff --git a/.gitea/workflows/README.md b/.gitea/workflows/README.md
index 05a4fb8..c2cf76c 100644
--- a/.gitea/workflows/README.md
+++ b/.gitea/workflows/README.md
@@ -2,7 +2,9 @@
 
 [ci-deploy.yml](ci-deploy.yml) 是当前唯一工作流：main push 或 workflow_dispatch 触发，运行于 ces-control，生产并发组串行且不取消正在运行的发布，任务超时 60 分钟。
 
-顺序为检出 → ci-ces-control.sh 验证 → 构建 SHA 生产镜像 → main 分支执行 deploy-ces-control.sh。设置 DOCKER_BUILDKIT=0 使用当前 builder 路径，检出不保留认证信息；工作流不更改共享 Runner。PostgreSQL 保留为预置资源，基础信息由文件保存；发布只为旧基础信息做一次只读导入，不再写入数据库。
+顺序为检出 → ci-ces-control.sh 验证及执行 Endor 示例 → 展示报告并上传 artifact → 构建 SHA 生产镜像 → main 分支执行 deploy-ces-control.sh。报告展示和上传使用 always()，验证失败也保留诊断。Artifact 名为 endor-execution-output-<sha>，包含 endor-examples.md 和 endor-examples.json，保留 30 天；从对应 Actions run 页面下载。使用 actions/upload-artifact@v3 连接当前 Gitea 的 artifact 服务，接口参考 [action 定义](https://github.com/actions/upload-artifact/blob/v3/action.yml)。
+
+设置 DOCKER_BUILDKIT=0 使用当前 builder 路径，检出不保留认证信息；工作流不更改共享 Runner。PostgreSQL 保留为预置资源，基础信息由文件保存；发布只为旧基础信息做一次只读导入，不再写入数据库。
 
 脚本职责见 [scripts](../../scripts/README.md)，完整发布、保留数据和回退规则见 [部署](../../docs/operations/deployment.md)。main 推送会进入现有发布链，不能把推送成功写成 CI 或部署成功；按用户约定不主动跟踪状态。
 
diff --git a/.gitea/workflows/ci-deploy.yml b/.gitea/workflows/ci-deploy.yml
index 5a17e13..f323878 100644
--- a/.gitea/workflows/ci-deploy.yml
+++ b/.gitea/workflows/ci-deploy.yml
@@ -25,6 +25,21 @@ jobs:
       - name: Verify SQLite, PostgreSQL and host behavior
         run: sh scripts/ci-ces-control.sh
 
+      - name: Show Endor example execution output
+        if: always()
+        run: cat ci-output/endor-examples.md
+
+      - name: Upload Endor example execution output
+        if: always()
+        uses: actions/upload-artifact@v3
+        with:
+          name: endor-execution-output-${{ gitea.sha }}
+          path: |
+            ci-output/endor-examples.md
+            ci-output/endor-examples.json
+          if-no-files-found: error
+          retention-days: 30
+
       - name: Build production image
         run: docker build --target runtime -t "endorphin-aster-vm:$ASTER_RELEASE_SHA" .
 
diff --git a/.gitignore b/.gitignore
index 4c5bf6b..7cbc573 100644
--- a/.gitignore
+++ b/.gitignore
@@ -2,6 +2,7 @@
 /data/
 /work/
 /outputs/
+/ci-output/
 .env
 *.db
 *.db-shm
diff --git a/README.md b/README.md
index bbe4624..ab96f21 100644
--- a/README.md
+++ b/README.md
@@ -44,4 +44,4 @@ Endor 采用平台解释执行模型；AST 到字节码的生成属于解释器
 
 本地不编译、不测试、不下载依赖；可做文档和静态文件检查。运行验证交给 ces-control；按用户约定不主动跟踪 CI／部署结果。命令和使用条件见 [验证](docs/operations/validation.md)。
 
-文档随代码版本更新：就地解释放各目录 README，跨目录契约放 docs，对外阅读副本以仓库文档为准。忽略的 work、target、data 和 Git 内部目录不属于需要维护 README 的源码目录。
+文档随代码版本更新：就地解释放各目录 README，跨目录契约放 docs，对外阅读副本以仓库文档为准。忽略的 work、target、data、ci-output 和 Git 内部目录不属于需要维护 README 的源码目录。Endor 示例实际输出由 CI 保存为 artifact，查看方式见 [验证报告](docs/operations/validation.md)。
diff --git a/crates/compiler/examples/README.md b/crates/compiler/examples/README.md
index d0d77c5..50fa90f 100644
--- a/crates/compiler/examples/README.md
+++ b/crates/compiler/examples/README.md
@@ -4,6 +4,7 @@
 
 | 文件 | 输入模式与预期结果 |
 | --- | --- |
+| [cases.json](cases.json) | CI 实际输出报告的用例清单；覆盖本目录每个 .endor 文件，指定 run／eval 和精确预期 stdout |
 | [hello.endor](hello.endor) | run；main 返回“你好，Endor 🌱” |
 | [calls.endor](calls.endor) | run；前向函数引用、参数和 let，返回“函数调用已接通” |
 | [shadowing.endor](shadowing.endor) | run；参数遮蔽全局函数名并传递函数值，返回“局部名称优先” |
@@ -19,3 +20,5 @@ cli eval crates/compiler/examples/literal.endor
 ```
 
 run 默认调用无参数 main，可用 --entry name 选择另一个无参数入口；eval 解析一个完整表达式，二者都支持 - 读取 stdin 到 EOF。类型名称暂只保留语法，Callable 不要求预先注册语言类。例子不依赖 print 或其他默认标准库。编译器尚无算术运行语义；前端 basic.endor 含算术，只是解析示例，不能混作可执行示例。规则见 [编译契约](../../../docs/runtime/compiler.md)，返回 [compiler](../README.md)。
+
+CI 使用 [endor_examples.py](../../../scripts/endor_examples.py) 再次执行真实 CLI 并保存实际结果，不只显示 Rust 测试名称。对应 Actions run 的 endor-execution-output-<sha> artifact 包含源码、命令、预期／实际 stdout、stderr、退出码和状态；位置与失败规则见 [验证报告](../../../docs/operations/validation.md)。新增 .endor 示例时同步 cases.json，未覆盖文件会让报告步骤失败。
diff --git a/crates/compiler/examples/cases.json b/crates/compiler/examples/cases.json
new file mode 100644
index 0000000..c5303ce
--- /dev/null
+++ b/crates/compiler/examples/cases.json
@@ -0,0 +1,35 @@
+{
+  "format_version": 1,
+  "cases": [
+    {
+      "name": "hello",
+      "mode": "run",
+      "source": "hello.endor",
+      "expected_stdout": "你好，Endor 🌱\n"
+    },
+    {
+      "name": "calls",
+      "mode": "run",
+      "source": "calls.endor",
+      "expected_stdout": "函数调用已接通\n"
+    },
+    {
+      "name": "shadowing",
+      "mode": "run",
+      "source": "shadowing.endor",
+      "expected_stdout": "局部名称优先\n"
+    },
+    {
+      "name": "number",
+      "mode": "run",
+      "source": "number.endor",
+      "expected_stdout": "-123456789012345678901234567890.00100\n"
+    },
+    {
+      "name": "literal",
+      "mode": "eval",
+      "source": "literal.endor",
+      "expected_stdout": "表达式 → 字节码 → VM\n"
+    }
+  ]
+}
diff --git a/docs/README.md b/docs/README.md
index 7582762..d9f0d52 100644
--- a/docs/README.md
+++ b/docs/README.md
@@ -22,7 +22,7 @@
 | LC_CTYPE、UTF-8、显式数据库配置和执行限额 | [配置](operations/configuration.md) | [core](../crates/core/README.md)、[storage](../crates/storage/README.md)、[server](../crates/server/README.md) |
 | CLI／HTTP 怎么运行，数据库在哪里使用 | [运行](operations/README.md) | [cli](../crates/cli/README.md)、[server](../crates/server/README.md) |
 | Docker、PostgreSQL、ces-control、回退与数据目录 | [部署](operations/deployment.md) | [scripts](../scripts/README.md)、[工作流](../.gitea/workflows/README.md) |
-| 检查覆盖什么，结果如何确认 | [验证](operations/validation.md) | [objects/tests](../crates/objects/tests/README.md)、[core/tests](../crates/core/tests/README.md)、[storage/tests](../crates/storage/tests/README.md) |
+| 检查覆盖什么，CI 中哪里查看 Endor 实际输出与 artifact | [验证](operations/validation.md) | [示例清单](../crates/compiler/examples/cases.json)、[报告脚本](../scripts/endor_examples.py)、[core/tests](../crates/core/tests/README.md)、[storage/tests](../crates/storage/tests/README.md) |
 | 改代码后文档写在哪里、怎样避免失效 | [维护规则](contributing/README.md) | 各目录 README 与本索引 |
 
 ## 多层目录
diff --git a/docs/contributing/handoff.md b/docs/contributing/handoff.md
index cd39539..e15067b 100644
--- a/docs/contributing/handoff.md
+++ b/docs/contributing/handoff.md
@@ -22,6 +22,7 @@ Objects 零依赖，统一拥有核心对象、Number、代码对象／指令数
 - Host 文件信息独立于数据库：host.version、deployment.release 与 metadata put/get 使用 UTF-8 JSON。旧 PostgreSQL 基础信息只读导入缺失文件，不删除原库，见 [文件信息](../operations/host-information.md)。
 - Storage 只提供显式 DatabaseConfig／DatabaseConnection；没有默认数据库、自动应用建表或基础信息 KV。特殊类的名称／接口尚未定义，没有将连接接入 VM。
 - CLI 提供引导、解析、源码执行与文件信息，Server HTTP 存活／只读 VM 信息；生产 PostgreSQL 作为未来特殊类的预置资源保留。
+- CI 使用 examples/cases.json 执行全部 Endor 示例，保存实际 stdout／stderr、退出码及状态；对应 run 页面展示 Markdown，并上传 endor-execution-output-<sha> artifact（Markdown／JSON，保留 30 天），失败也保留报告。查看方式见 [验证](../operations/validation.md)，本地不运行报告脚本。
 
 ## 资料位置
 
diff --git a/docs/operations/deployment.md b/docs/operations/deployment.md
index 5635e54..5c773c0 100644
--- a/docs/operations/deployment.md
+++ b/docs/operations/deployment.md
@@ -30,7 +30,7 @@ docker compose -f docker-compose.postgres.yml exec aster cli bootstrap
 
 [工作流](../../.gitea/workflows/README.md) 响应 main push 和 workflow_dispatch，在 ces-control 上串行运行：
 
-1. 检出代码，调用 ci-ces-control.sh；在 Docker 中检查 workspace（含 SQLite 适配）、独立 HTTP 冒烟及专用真实 PostgreSQL。
+1. 检出代码，调用 ci-ces-control.sh；在 Docker 中检查 workspace（含 SQLite 适配）、Endor 示例实际输出、独立 HTTP 冒烟及专用真实 PostgreSQL。清理检查容器前复制报告到宿主 ci-output，工作流无论验证成功或失败均展示并上传 endor-execution-output-<sha> artifact；字段与位置见 [验证报告](validation.md)。
 2. 检查成功后构建 endorphin-aster-vm:<sha> 生产镜像。
 3. main 上调用 deploy-ces-control.sh 更新 VM 宿主并保留预置 PostgreSQL。
 
diff --git a/docs/operations/validation.md b/docs/operations/validation.md
index 93cbbee..a318a80 100644
--- a/docs/operations/validation.md
+++ b/docs/operations/validation.md
@@ -16,10 +16,21 @@ cargo test -p storage --test postgres --locked -- --ignored
 python3 scripts/smoke.py
 ```
 
-check.sh 先运行 generate_ast.py --check 验证 ASDL 产物同步，以及仅远端执行的 test_generate_ast.py 生成器回归检查，再执行 cargo fmt --all --check、独立词法／语法文件的 rustfmt --edition 2024 --check crates/frontend/src/lexer.peg crates/frontend/src/grammar.peg、cargo test --workspace --locked 和 cargo clippy --workspace --all-targets --locked -- -D warnings。workspace 测试也包含 rustdoc 示例；显式忽略的 PostgreSQL 测试需要单独执行。
+check.sh 先运行 generate_ast.py --check 验证 ASDL 产物同步，以及仅远端执行的 test_generate_ast.py 生成器回归检查、test_endor_examples.py 报告失败路径检查，再执行 cargo fmt --all --check、独立词法／语法文件的 rustfmt --edition 2024 --check crates/frontend/src/lexer.peg crates/frontend/src/grammar.peg、cargo test --workspace --locked 和 cargo clippy --workspace --all-targets --locked -- -D warnings。workspace 测试也包含 rustdoc 示例；显式忽略的 PostgreSQL 测试需要单独执行。
 
 ci-ces-control.sh 负责准备全新专用临时 PostgreSQL 和测试 URL；成功路径使用真实数据库，不用模拟 socket 代替。测试库与生产数据不共享挂载。没有 ASTER_TEST_POSTGRES_URL 时，显式执行 PostgreSQL 测试会失败，不标记为通过。
 
+## Endor 示例实际输出
+
+在对应 Gitea 仓库的 Actions → 本次提交的 run 页面，打开 Show Endor example execution output 步骤可看实际运行报告；下载 endor-execution-output-<sha> artifact 可保存两种格式：
+
+- endor-examples.md：逐例展示源码、执行命令、预期值、实际 stdout／stderr、退出码和 PASS／FAIL／NOT_RUN。
+- endor-examples.json：保留精确输出（包括换行）、提交 SHA、UTC 时间、构建／定位命令结果和逐例状态，供工具读取。
+
+[endor_examples.py](../../scripts/endor_examples.py) 构建真实 CLI，通过 Cargo metadata 定位二进制并执行 [cases.json](../../crates/compiler/examples/cases.json) 中的全部示例；不以预期值冒充实际值。stdout 必须精确匹配、stderr 为空且退出码为 0 才通过。单例超时 30 秒；构建失败不运行示例并标记 NOT_RUN；报告本身的异常标记 FAIL。一个案例失败仍收集后续案例。示例报告不代表 workspace、数据库、宿主和部署全部通过，整体结果仍看工作流。
+
+ci-ces-control.sh 在检查容器退出后、清理前将两份文件复制到宿主 ci-output；验证步骤提前失败时已有 NOT_RUN 占位说明。报告运行不掩盖前面检查的失败；后续数据库或宿主检查失败也不丢报告。工作流用 always() 展示并上传，artifact 保留 30 天，不写进 Git 或生产数据目录。普通 Actions 日志仍保留完整测试／构建信息；报告上传失败会使工作流失败。
+
 ## 覆盖范围与文件位置
 
 | 范围 | 当前检查内容 | 入口 |
diff --git a/scripts/README.md b/scripts/README.md
index b54a413..c09425c 100644
--- a/scripts/README.md
+++ b/scripts/README.md
@@ -6,9 +6,11 @@
 | --- | --- |
 | [generate_ast.py](generate_ast.py) | 从 Endor.asdl 生成并格式化 Rust／TypeScript／JSON Schema；--write 更新产物，--check 只比较，无依赖下载或语言执行 |
 | [test_generate_ast.py](test_generate_ast.py) | 远端 unittest：schema 扩展同步、递归装箱／遍历、非法定义拒绝 |
-| [check.sh](check.sh) | 先检查 ASDL 产物同步和生成器行为，再运行 Cargo fmt 与独立 lexer.peg／grammar.peg 的 rustfmt 检查、workspace 测试、严格 Clippy；会使用 Cargo 依赖和构建产物 |
+| [endor_examples.py](endor_examples.py) | 仅远端：构建 CLI，按 examples/cases.json 执行每个真实 Endor 文件，严格比较 stdout／stderr／退出码；保存源码、命令、实际输出、PASS／FAIL／NOT_RUN 的 Markdown 和 JSON，失败也写报告 |
+| [test_endor_examples.py](test_endor_examples.py) | 仅远端：报告清单覆盖、精确输出、不匹配、构建失败、超时部分输出与异常时仍保存双格式报告；模拟子进程，不启动 VM |
+| [check.sh](check.sh) | 先检查 ASDL 产物同步、生成器与输出报告行为，再运行 Cargo fmt 与独立 lexer.peg／grammar.peg 的 rustfmt 检查、workspace 测试、严格 Clippy；会使用 Cargo 依赖和构建产物 |
 | [smoke.py](smoke.py) | 标准库 Python 脚本；构建后用 Cargo metadata 定位 cli／server 的 debug 二进制，支持 CARGO_TARGET_DIR；文件 metadata 跨进程与日文值；失效数据库变量下独立启动、不创建 SQLite 文件、HTTP 与 SIGTERM，不改生产配置 |
-| [ci-ces-control.sh](ci-ces-control.sh) | 需要 ASTER_RELEASE_SHA；构建 ci 镜像、创建临时网络／PostgreSQL／检查容器，运行全部检查；退出清理临时容器和网络，保留专属远端 Cargo 缓存卷 |
+| [ci-ces-control.sh](ci-ces-control.sh) | 需要 ASTER_RELEASE_SHA；构建 ci 镜像、创建临时网络／PostgreSQL／检查容器；检查后独立执行 Endor 示例，失败不掩盖原检查状态；清理容器前复制两份报告到宿主 ci-output，提前失败也保留 NOT_RUN 说明；退出清理临时容器和网络，保留专属远端 Cargo 缓存卷 |
 | [deploy-ces-control.sh](deploy-ces-control.sh) | ces-control root 执行；校验已构建镜像、锁定发布、保留密码和数据、验证 HTTP 与无数据库 CLI 引导后切换 current，失败尝试恢复旧版；保留镜像；应用服务 aster 内调用 cli bootstrap，基础信息改写文件；一次性只读导入旧基础信息，不覆盖新文件或删除原库 |
 
 CI／部署使用的镜像与路径仅属于本项目，禁止改动其他项目和共享 Runner。smoke.py 临时选择端口后再启动宿主，存在小的端口竞争窗口，指定环境遇到冲突可重试。
diff --git a/scripts/check.sh b/scripts/check.sh
index 2a5e793..32ac8a7 100644
--- a/scripts/check.sh
+++ b/scripts/check.sh
@@ -2,6 +2,7 @@
 set -eu
 python3 scripts/generate_ast.py --check
 python3 scripts/test_generate_ast.py
+python3 scripts/test_endor_examples.py
 cargo fmt --all --check
 rustfmt --edition 2024 --check crates/frontend/src/lexer.peg crates/frontend/src/grammar.peg
 cargo test --workspace --locked
diff --git a/scripts/ci-ces-control.sh b/scripts/ci-ces-control.sh
index 9df0f3d..eeb5273 100644
--- a/scripts/ci-ces-control.sh
+++ b/scripts/ci-ces-control.sh
@@ -12,6 +12,12 @@ network="endorphin-aster-vm-ci-$ci_id"
 postgres="endorphin-aster-vm-ci-pg-$ci_id"
 checks="endorphin-aster-vm-ci-checks-$ci_id"
 image="endorphin-aster-vm-ci:$release_sha"
+report_dir="$PWD/ci-output"
+mkdir -p "$report_dir"
+# Reset both outputs before any build, so even an early failure has an honest
+# NOT_RUN artifact rather than a previous attempt's example results.
+printf '# Endor 示例实际运行结果\n\n提交：%s\n\n状态：NOT_RUN\n\nCI 尚未完成示例执行；请查看验证步骤日志。\n' "$release_sha" > "$report_dir/endor-examples.md"
+printf '{"format_version":1,"commit":"%s","status":"NOT_RUN","error":"CI did not finish example execution; inspect the verification log.","cases":[]}\n' "$release_sha" > "$report_dir/endor-examples.json"
 cleanup() {
     docker rm -f "$checks" "$postgres" >/dev/null 2>&1 || true
     docker network rm "$network" >/dev/null 2>&1 || true
@@ -41,12 +47,29 @@ if [ "$ready" != true ]; then
     exit 1
 fi
 
+check_status=0
 docker run --name "$checks" --init --network "$network" \
     --mount type=volume,src=endorphin-aster-vm-ci-cargo,dst=/usr/local/cargo/registry \
     --mount type=volume,src=endorphin-aster-vm-ci-target,dst=/build/target \
     --env ASTER_TEST_POSTGRES_URL='postgres://aster:aster-ci-only@postgres:5432/aster?sslmode=disable' \
+    --env ASTER_RELEASE_SHA="$release_sha" \
     "$image" sh -eu -c '
-        sh scripts/check.sh
+        checks=0
+        sh scripts/check.sh || checks=$?
+        examples=0
+        python3 scripts/endor_examples.py --output-dir /build/ci-output || examples=$?
+        test "$checks" -eq 0
+        test "$examples" -eq 0
         cargo test -p storage --test postgres --locked -- --ignored
         python3 scripts/smoke.py
-    '
+    ' || check_status=$?
+
+# Copy reports before the EXIT trap removes the stopped checks container.
+# Keep the original validation failure status if report extraction also fails.
+copy_status=0
+docker cp "$checks:/build/ci-output/endor-examples.md" "$report_dir/endor-examples.md" || copy_status=$?
+docker cp "$checks:/build/ci-output/endor-examples.json" "$report_dir/endor-examples.json" || copy_status=$?
+if [ "$check_status" -ne 0 ]; then
+    exit "$check_status"
+fi
+exit "$copy_status"
diff --git a/scripts/endor_examples.py b/scripts/endor_examples.py
new file mode 100644
index 0000000..cea42e6
--- /dev/null
+++ b/scripts/endor_examples.py
@@ -0,0 +1,145 @@
+#!/usr/bin/env python3
+"""Run Endor example files and retain actual output on the designated CI host.
+
+Never run locally: this builds the CLI and executes language code.
+"""
+
+import argparse
+from datetime import datetime, timezone
+import json
+import os
+from pathlib import Path
+import re
+import shlex
+import subprocess
+
+
+ROOT = Path(__file__).resolve().parents[1]
+EXAMPLES = ROOT / "crates/compiler/examples"
+
+
+def decoded(value):
+    return value.decode("utf-8", "replace") if isinstance(value, bytes) else value or ""
+
+
+def invoke(command, timeout):
+    """Keep stdout, stderr and failure state even for launch errors/timeouts."""
+    record = {"command": command, "exit_code": None, "stdout": "", "stderr": "", "error": None}
+    try:
+        result = subprocess.run(command, cwd=ROOT, capture_output=True, text=True,
+                                encoding="utf-8", errors="replace", timeout=timeout)
+        record.update(exit_code=result.returncode, stdout=result.stdout, stderr=result.stderr)
+    except subprocess.TimeoutExpired as error:
+        record.update(stdout=decoded(error.stdout), stderr=decoded(error.stderr),
+                      error=f"timed out after {timeout} seconds")
+    except OSError as error:
+        record["error"] = str(error)
+    return record
+
+
+def load_cases():
+    manifest = json.loads((EXAMPLES / "cases.json").read_text(encoding="utf-8"))
+    if manifest.get("format_version") != 1:
+        raise ValueError("unsupported example manifest version")
+    cases = manifest["cases"]
+    names, sources = set(), set()
+    for case in cases:
+        if not re.fullmatch(r"[a-z][a-z0-9_-]*", case["name"]) or case["name"] in names:
+            raise ValueError("invalid or duplicate example name")
+        source = EXAMPLES / case["source"]
+        if source.resolve().parent != EXAMPLES.resolve() or source.suffix != ".endor":
+            raise ValueError("example source must be a direct .endor file")
+        if case["mode"] not in ("run", "eval") or not isinstance(case["expected_stdout"], str):
+            raise ValueError("invalid example mode or expected output")
+        if case["source"] in sources:
+            raise ValueError("duplicate example source")
+        case["source_text"] = source.read_text(encoding="utf-8")
+        names.add(case["name"])
+        sources.add(case["source"])
+    if not cases or sources != {path.name for path in EXAMPLES.glob("*.endor")}:
+        raise ValueError("manifest must cover every .endor example exactly once")
+    return cases
+
+
+def fence(text, language="text"):
+    # Keep source/output literal even if future examples print Markdown fences.
+    longest = max((len(match[0]) for match in re.finditer(r"`+", text)), default=0)
+    delimiter = "`" * max(3, longest + 1)
+    return f"{delimiter}{language}\n{text}" + ("" if text.endswith("\n") else "\n") + delimiter
+
+
+def markdown(report):
+    lines = ["# Endor 示例实际运行结果", "", f"提交：`{report['commit']}`",
+             f"状态：**{report['status']}**", f"生成时间（UTC）：{report['generated_at']}",
+             "", "本报告仅覆盖 Endor 示例；整个 CI/CD 的状态以工作流结果为准。"]
+    if report.get("error"):
+        lines += ["", "报告阶段错误：", fence(report["error"])]
+    for stage in report["preparation"]:
+        lines += ["", "## 准备：" + shlex.join(stage["command"]),
+                  f"退出码：{stage['exit_code']}；错误：{stage['error'] or '无'}"]
+        if stage["exit_code"] != 0 or stage["error"]:
+            lines += ["", "stdout：", fence(stage["stdout"]), "", "stderr：", fence(stage["stderr"])]
+    for case in report["cases"]:
+        lines += ["", f"## {case['name']} — {case['status']}", "",
+                  "命令：", fence(shlex.join(case["command"]), "sh"), "", "源码：",
+                  fence(case["source_text"], "endor"), "", "预期 stdout：",
+                  fence(case["expected_stdout"]), "", "实际 stdout：", fence(case["stdout"]),
+                  "", "实际 stderr：", fence(case["stderr"]), "",
+                  f"预期退出码：0；实际退出码：{case['exit_code']}",
+                  f"错误：{case['error'] or '无'}"]
+    return "\n".join(lines) + "\n"
+
+
+def collect(report):
+    for case in load_cases():
+        report["cases"].append({**case, "status": "NOT_RUN", "exit_code": None,
+                                "stdout": "", "stderr": "", "error": None,
+                                "expected_stderr": "", "expected_exit_code": 0,
+                                "command": ["cli", case["mode"], "crates/compiler/examples/" + case["source"]]})
+    build = invoke(["cargo", "build", "-p", "cli", "--locked"], timeout=600)
+    report["preparation"].append(build)
+    if build["exit_code"] != 0 or build["error"]:
+        report.update(status="NOT_RUN", error="CLI build failed; examples were not executed.")
+        return
+    metadata = invoke(["cargo", "metadata", "--locked", "--no-deps", "--format-version", "1"], timeout=60)
+    report["preparation"].append(metadata)
+    if metadata["exit_code"] != 0 or metadata["error"]:
+        report.update(status="NOT_RUN", error="Cannot locate the compiled CLI.")
+        return
+    binary = Path(json.loads(metadata["stdout"])["target_directory"]) / "debug" / "cli"
+    for case in report["cases"]:
+        command = [str(binary), *case["command"][1:]]
+        actual = invoke(command, timeout=30)
+        case.update({key: actual[key] for key in ["exit_code", "stdout", "stderr", "error"]})
+        case["binary"] = str(binary)
+        case["status"] = "PASS" if (
+            case["exit_code"] == case["expected_exit_code"] and not case["error"]
+            and case["stdout"] == case["expected_stdout"]
+            and case["stderr"] == case["expected_stderr"]
+        ) else "FAIL"
+    report["status"] = "PASS" if all(case["status"] == "PASS" for case in report["cases"]) else "FAIL"
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--output-dir", type=Path, required=True)
+    args = parser.parse_args()
+    args.output_dir.mkdir(parents=True, exist_ok=True)
+    report = {"format_version": 1, "commit": os.environ.get("ASTER_RELEASE_SHA", "unknown"),
+              "generated_at": datetime.now(timezone.utc).isoformat(), "status": "NOT_RUN",
+              "error": None, "preparation": [], "cases": []}
+    try:
+        collect(report)
+    except Exception as error:
+        report.update(status="FAIL", error=str(error))
+    # Always persist diagnostics before returning a failing CI exit code.
+    args.output_dir.joinpath("endor-examples.json").write_text(
+        json.dumps(report, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
+    rendered = markdown(report)
+    args.output_dir.joinpath("endor-examples.md").write_text(rendered, encoding="utf-8")
+    print(rendered)
+    return 0 if report["status"] == "PASS" else 1
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/scripts/test_endor_examples.py b/scripts/test_endor_examples.py
new file mode 100644
index 0000000..7b08bc3
--- /dev/null
+++ b/scripts/test_endor_examples.py
@@ -0,0 +1,80 @@
+"""Remote-only regression checks for example output capture and failure reports."""
+
+import contextlib
+import io
+import json
+from pathlib import Path
+import subprocess
+import tempfile
+import unittest
+from unittest.mock import patch
+
+import endor_examples as examples
+
+
+def result(stdout="", code=0):
+    return {"command": ["mock"], "exit_code": code, "stdout": stdout, "stderr": "", "error": None}
+
+
+def report():
+    return {"format_version": 1, "commit": "test", "generated_at": "test", "status": "NOT_RUN",
+            "preparation": [], "cases": [], "error": None}
+
+
+class ExampleReportTests(unittest.TestCase):
+    def test_manifest_runs_every_source_and_captures_exact_output(self):
+        cases = examples.load_cases()
+        responses = [result(), result(json.dumps({"target_directory": "/tmp/ci-target"}))]
+        responses += [result(case["expected_stdout"]) for case in cases]
+        collected = report()
+        with patch.object(examples, "invoke", side_effect=responses) as invoke:
+            examples.collect(collected)
+        self.assertEqual(collected["status"], "PASS")
+        self.assertEqual(invoke.call_count, len(cases) + 2)
+        for actual, expected in zip(collected["cases"], cases):
+            self.assertEqual(actual["stdout"], expected["expected_stdout"])
+            self.assertEqual(actual["status"], "PASS")
+            self.assertEqual(actual["binary"], "/tmp/ci-target/debug/cli")
+
+    def test_output_mismatch_fails_and_build_failure_never_claims_examples_ran(self):
+        cases = examples.load_cases()
+        responses = [result(), result(json.dumps({"target_directory": "/tmp/target"}))]
+        responses += [result("wrong\n")] + [result(case["expected_stdout"]) for case in cases[1:]]
+        collected = report()
+        with patch.object(examples, "invoke", side_effect=responses):
+            examples.collect(collected)
+        self.assertEqual(collected["status"], "FAIL")
+        self.assertEqual(collected["cases"][0]["stdout"], "wrong\n")
+        self.assertEqual(collected["cases"][0]["status"], "FAIL")
+        self.assertIn("wrong", examples.markdown(collected))
+        collected = report()
+        with patch.object(examples, "invoke", return_value=result(code=1)) as invoke:
+            examples.collect(collected)
+        self.assertEqual(invoke.call_count, 1)
+        self.assertEqual(collected["status"], "NOT_RUN")
+        self.assertTrue(all(case["status"] == "NOT_RUN" for case in collected["cases"]))
+
+    def test_timeout_keeps_partial_utf8_stdout_and_stderr(self):
+        failure = subprocess.TimeoutExpired(["cli"], 1, output="日".encode(), stderr=b"partial error")
+        with patch.object(examples.subprocess, "run", side_effect=failure):
+            captured = examples.invoke(["cli"], 1)
+        self.assertIsNone(captured["exit_code"])
+        self.assertEqual(captured["stdout"], "日")
+        self.assertEqual(captured["stderr"], "partial error")
+        self.assertIn("timed out", captured["error"])
+
+    def test_unexpected_failure_still_writes_both_artifacts(self):
+        with tempfile.TemporaryDirectory() as directory:
+            with patch("sys.argv", ["endor_examples.py", "--output-dir", directory]), \
+                 patch.object(examples, "collect", side_effect=ValueError("broken manifest")), \
+                 contextlib.redirect_stdout(io.StringIO()):
+                self.assertEqual(examples.main(), 1)
+            payload = json.loads(Path(directory, "endor-examples.json").read_text(encoding="utf-8"))
+            self.assertEqual(payload["status"], "FAIL")
+            self.assertEqual(payload["error"], "broken manifest")
+            self.assertIn("broken manifest", Path(directory, "endor-examples.md").read_text(encoding="utf-8"))
+        self.assertTrue(examples.fence("```\n日").startswith("````text\n"))
+
+
+if __name__ == "__main__":
+    unittest.main()