diff --git a/agent/legacy-rag/api/methodology_routes.py b/agent/legacy-rag/api/methodology_routes.py
old mode 100755
new mode 100644
index 07b43d4..06137a4
--- a/agent/legacy-rag/api/methodology_routes.py
+++ b/agent/legacy-rag/api/methodology_routes.py
@@ -29,7 +29,7 @@ PAPERS_CSV = PROJECT_ROOT / "data" / "evaluation" / "methodology_papers.csv"
 GOLD_CSV = PROJECT_ROOT / "data" / "evaluation" / "methodology_gold.csv"
 REVIEW_RESULTS_JSON = BATCH_DIR / "review_results.json"
 REVIEW_RESULTS_CSV = BATCH_DIR / "review_results.csv"
-# 2026-08-31 新工程适配:上传目录/产出根支持 env 覆盖
+# 2026-08-31 新工程适配:上传目录/产出根支持 env 覆盖(数据根迁移时无需改代码)
 UPLOAD_JOBS_DIR = Path(os.environ.get("METHODOLOGY_UPLOAD_JOBS_DIR", PROJECT_ROOT / "data" / "evaluation" / "methodology_upload_jobs"))
 UPLOAD_OUTPUT_ROOT = Path(os.environ.get("METHODOLOGY_UPLOAD_OUTPUT_ROOT", PROJECT_ROOT / "data" / "evaluation" / "0505"))
 DEFAULT_MINERU_CMD = str(conda_env_path("minerupy312") / "bin" / "mineru")
@@ -351,6 +351,22 @@ def mineru_gpu_preflight_error(
     )
 
 
+# 缺陷 #38（2026-09-11）：成功任务的 job.log 必然有 doc_analyze_streaming 行（176 实测）。
+# returncode≠0 且日志中无此行 = 启动期失败（pipeline 起来前就死了，真因被块缓冲吞掉）。
+# 判定条件（methodology 与 legacy-rag 两处 build_mineru_runtime/run_uploaded_methodology_job 同改）
+MINERU_PIPELINE_LOG_MARKER = "doc_analyze_streaming"
+MINERU_STARTUP_RETRIES = 1
+
+
+def job_log_has_pipeline_line(log_path: Path) -> bool:
+    """job.log 是否已出现 pipeline 日志行（def _job_log_has_pipeline_line）。"""
+    try:
+        text = log_path.read_text(encoding="utf-8", errors="replace")
+    except OSError:
+        return False
+    return MINERU_PIPELINE_LOG_MARKER in text
+
+
 def summarize_mineru_failure(log_path: Path, returncode: int) -> str:
     message = f"MinerU failed with exit code {returncode}; see {log_path}"
     if not log_path.exists():
@@ -368,6 +384,28 @@ def summarize_mineru_failure(log_path: Path, returncode: int) -> str:
     return message
 
 
+def summarize_mineru_startup_failure(log_path: Path, returncode: int, retried: bool) -> str:
+    """启动期失败（无 pipeline 行）的中文可读错误（缺陷 #38）。
+
+    附上 job.log 中可见的原始错误尾部，让用户能看到真因（修复前真因
+    被 stdout 块缓冲 + CLI 立即杀 API 进程永久丢失）。
+    """
+    tail = ""
+    try:
+        text = log_path.read_text(encoding="utf-8", errors="replace")
+        lines = [ln for ln in text.splitlines() if ln.strip()]
+        if lines:
+            tail = " | ".join(ln.strip() for ln in lines[-3:])[:500]
+    except OSError:
+        tail = ""
+    detail = f"原始日志尾部：{tail}" if tail else f"原始错误见日志：{log_path}"
+    action = "重试后" if retried else ""
+    return (
+        f"提取启动阶段失败{action}，可能受主机 GPU/显示重配置干扰，请重试"
+        f"（MinerU 退出码 {returncode}；{detail}）。"
+    )
+
+
 def build_mineru_runtime() -> tuple[list[str], dict[str, str], dict[str, str]]:
     env = os.environ.copy()
     env.setdefault("MINERU_MODEL_SOURCE", "modelscope")
@@ -384,6 +422,9 @@ def build_mineru_runtime() -> tuple[list[str], dict[str, str], dict[str, str]]:
     else:
         env.setdefault("CUDA_VISIBLE_DEVICES", "0")
     env.setdefault("TORCH_CUDNN_V8_API_DISABLED", "1")
+    # 缺陷 #38（2026-09-11）：子进程 API 的异常日志即写即落（不再被 stdout
+    # 块缓冲吞掉）——CLI 失败后立即杀 API 进程前真因已落到 job.log。
+    env["PYTHONUNBUFFERED"] = "1"
     # mineru 3.1.13 CLI 陷阱（部署 agent 实测）：带 MINERU_API_SHUTDOWN_ON_STDIN_EOF=1
     # 时，stdin EOF 会让 API 进程静默 shutdown——subprocess 调用必须 stdin=DEVNULL
     env.setdefault("MINERU_API_SHUTDOWN_ON_STDIN_EOF", "0")
@@ -566,7 +607,34 @@ def run_uploaded_methodology_job(
                     stdout=handle,
                     stderr=subprocess.STDOUT,
                 )
+            # 缺陷 #38：启动期失败自动重试 1 次——returncode≠0 且 job.log 无
+            # doc_analyze_streaming 行 = pipeline 起来前就失败（真因曾因块缓冲
+            # 丢失）。重试的 stdout 继续 append 到同一 job.log，日志/状态留
+            # retry 标记（G6-2：已有 pipeline 行的中途失败不得触发重试）。
+            startup_retried = False
+            if result.returncode != 0 and not job_log_has_pipeline_line(log_path):
+                startup_retried = True
+                with log_path.open("a", encoding="utf-8") as handle:
+                    handle.write(
+                        f"[retry] MinerU 启动期失败（exit {result.returncode}，"
+                        "无 pipeline 阶段日志），自动重试 1/1\n"
+                    )
+                    handle.write("$ " + " ".join(cmd) + "\n")
+                    handle.flush()
+                    result = subprocess.run(
+                        cmd,
+                        cwd=PROJECT_ROOT,
+                        env=mineru_env,
+                        text=True,
+                        stdin=subprocess.DEVNULL,
+                        stdout=handle,
+                        stderr=subprocess.STDOUT,
+                    )
             if result.returncode != 0:
+                if not job_log_has_pipeline_line(log_path):
+                    raise RuntimeError(
+                        summarize_mineru_startup_failure(log_path, result.returncode, retried=startup_retried)
+                    )
                 raise RuntimeError(summarize_mineru_failure(log_path, result.returncode))
             matches = sorted(mineru_out.rglob("*_content_list.json"))
             if not matches:
diff --git a/agent/legacy-rag/api/test_mineru_retry.py b/agent/legacy-rag/api/test_mineru_retry.py
new file mode 100644
index 0000000..cad974c
--- /dev/null
+++ b/agent/legacy-rag/api/test_mineru_retry.py
@@ -0,0 +1,328 @@
+"""test_mineru_retry.py —— MinerU 启动期失败重试与可读错误（缺陷 #38，验收 §4 + 回归 G6）。
+
+运行（worktree 根目录，无需部署、不触服务）：
+  /home/xzq/5090/teamWorkspace/zh/plantai-rag-new-venv/bin/python -m pytest agent/methodology/api/test_mineru_retry.py -v
+  /home/xzq/5090/teamWorkspace/zh/plantai-rag-new-venv/bin/python -m pytest agent/legacy-rag/api/test_mineru_retry.py -v
+说明：agent/methodology 与 agent/legacy-rag 共用同一份 routes（文件名同、内容同），
+同一测试文件复制两处、两处同跑（各用各树的 src/）。
+
+断言：
+1. build_mineru_runtime 返回 env 含 PYTHONUNBUFFERED=1；
+2. subprocess.run 首次失败 + job.log 无 pipeline 行 → 自动重试 1 次
+   （两次启动行、job.log 有 [retry] 标记，第二次 stdout 追加同一 job.log）；
+3. 重试仍失败 → status failed + 中文可读「提取启动阶段失败」+ 原始错误尾部；
+4. 已有 pipeline 行的中途失败 → 不重试，走原错误路径（含 CUDA OOM 中文错误）。
+"""
+from __future__ import annotations
+
+import json
+import subprocess
+from dataclasses import dataclass, field
+from pathlib import Path
+
+import pytest
+
+import api.methodology_routes as mr
+
+from api.methodology_routes import (
+    MINERU_PIPELINE_LOG_MARKER,
+    build_mineru_runtime,
+    job_log_has_pipeline_line,
+    run_uploaded_methodology_job,
+    summarize_mineru_startup_failure,
+)
+
+
+# ---------------------------------------------------------------------------
+# 固定夹具
+# ---------------------------------------------------------------------------
+
+@dataclass
+class FakeResult:
+    returncode: int
+
+
+@dataclass
+class RunnerRecorder:
+    """按调用顺序返回预设 returncode；每次调用把 stdout 写进所给 handle（模拟子进程日志）。"""
+
+    sequence: list
+    calls: list = field(default_factory=list)
+
+    def run(self, cmd, **kwargs):
+        n = len(self.calls)
+        self.calls.append({"cmd": cmd, "kwargs": kwargs})
+        rc = int(self.sequence[min(n, len(self.sequence) - 1)])
+        handle = kwargs.get("stdout")
+        if handle is not None and kwargs.get("text"):
+            handle.write(f"[fake stdout call#{n}] rc={rc}\n")
+            handle.flush()
+        return FakeResult(returncode=rc)
+
+
+@pytest.fixture
+def fake_success(monkeypatch):
+    """LLM 批次与 bbav5 配置阶段全部假成功（不触真实脚本/目录/真实函数）。
+
+    - gf_python_path / _preflight_executable 桩掉（不依赖真实 python 路径）
+    - build_bbav5_methodology_config 桩掉（不要求真实批次产物目录）
+    - figure_table 阶段走 routes 内部 try/except，不阻断
+    """
+    captured = {}
+
+    def fake_batch(cmd, **kwargs):
+        captured["batch_cmd"] = cmd
+        return FakeResult(returncode=0)
+
+    monkeypatch.setattr(mr, "gf_python_path", lambda: Path("/usr/bin/python3"))
+    monkeypatch.setattr(mr, "_preflight_executable", lambda cmd, hint, name: None)
+    monkeypatch.setattr(mr.subprocess, "run", fake_batch)
+    monkeypatch.setattr(
+        mr,
+        "build_bbav5_methodology_config",
+        lambda output_dir: captured.setdefault("config", {}),
+    )
+    return captured
+
+
+@pytest.fixture
+def job_root(monkeypatch, tmp_path):
+    """把 UPLOAD_JOBS_DIR 重定向到 tmp（绝不写真实数据目录）。"""
+    root = tmp_path / "upload_jobs"
+    monkeypatch.setattr(mr, "UPLOAD_JOBS_DIR", root)
+    return root
+
+
+def read_status(job_dir: Path) -> dict:
+    return json.loads((job_dir / "status.json").read_text(encoding="utf-8"))
+
+
+def run_job(monkeypatch, tmp_path, runner: RunnerRecorder, source_path: Path, fake_success, job_root):
+    """以 runner 重绑 subprocess.run 跑完整任务流（fake_success 保证后续阶段假成功）。
+
+    注意：fake_success 已把 mr.subprocess.run 桩为假批次；这里再次 setattr
+    为 runner（同一模块对象），此后所有子进程调用（mineru 两次 + LLM 批次）
+    都由 runner 按 sequence 依次应答——测试 sequence 需覆盖全部调用。
+    """
+    job_dir = job_root / "testjob"
+    # 模拟 mineru 成功产出：在 job/mineru 下伪造 *_content_list.json
+    # （测试不真正执行 mineru）
+    (job_dir / "mineru").mkdir(parents=True, exist_ok=True)
+    (job_dir / "mineru" / "fake_content_list.json").write_text("{}", encoding="utf-8")
+    # GPU 预检禁用（否则 nvidia-smi 会消耗 runner 的 sequence 槽位）
+    monkeypatch.setenv("METHODOLOGY_MINERU_GPU_PREFLIGHT", "0")
+    monkeypatch.setattr(subprocess, "run", runner.run)
+    run_uploaded_methodology_job(
+        job_id="testjob",
+        source_path=source_path,
+        paper_id="paper1",
+        family="",
+        output_dir=tmp_path / "out",
+        llm_model="deepseek-v4-flash",
+        llm_request_timeout=75,
+        sandbox=False,
+        sandbox_run_id="",
+    )
+    return job_dir
+
+
+@pytest.fixture
+def pdf_source(tmp_path):
+    src = tmp_path / "sample.pdf"
+    src.write_bytes(b"%PDF-1.4\nfake pdf bytes for startup-failure test\n")
+    return src
+
+
+@pytest.fixture
+def json_source(tmp_path):
+    src = tmp_path / "sample.json"
+    src.write_text("{}", encoding="utf-8")
+    return src
+
+
+# ---------------------------------------------------------------------------
+# 修复点 1：env 注入 PYTHONUNBUFFERED
+# ---------------------------------------------------------------------------
+
+def test_build_mineru_runtime_env_unbuffered(monkeypatch):
+    monkeypatch.delenv("MINERU_CMD", raising=False)
+    cmd, env, runtime = build_mineru_runtime()
+    assert env.get("PYTHONUNBUFFERED") == "1"
+    assert cmd  # 默认 MINERU_CMD 路径（不要求存在，build 阶段不校验）
+
+
+# ---------------------------------------------------------------------------
+# 修复点 2：启动期失败自动重试 1 次
+# ---------------------------------------------------------------------------
+
+def test_startup_failure_retries_once(monkeypatch, tmp_path, pdf_source, fake_success, job_root):
+    runner = RunnerRecorder(sequence=[1, 0, 0])  # mineru 首调失败 → 重试成功 → LLM 批次成功
+    job_dir = run_job(monkeypatch, tmp_path, runner, pdf_source, fake_success, job_root)
+
+    log_text = (job_dir / "job.log").read_text(encoding="utf-8")
+    startup_lines = [ln for ln in log_text.splitlines() if ln.startswith("$ ")]
+    assert len(startup_lines) >= 2, f"应出现两次 mineru 启动（1 次自动重试），实际：{startup_lines}"
+    assert "[retry]" in log_text, "job.log 应有可识别重试标记"
+    # 第二次 subprocess.run 的 stdout 追加到同一 job.log
+    assert "[fake stdout call#0]" in log_text
+    assert "[fake stdout call#1]" in log_text
+
+    status = read_status(job_dir)
+    assert status.get("status") == "completed"
+
+
+def test_startup_failure_retry_then_fail_readable_error(monkeypatch, tmp_path, pdf_source, fake_success, job_root):
+    runner = RunnerRecorder(sequence=[1, 1])  # 两次都失败，且无 pipeline 行
+    job_dir = run_job(monkeypatch, tmp_path, runner, pdf_source, fake_success, job_root)
+
+    log_text = (job_dir / "job.log").read_text(encoding="utf-8")
+    assert len([ln for ln in log_text.splitlines() if ln.startswith("$ /")]) == 2  # 两次 mineru 启动
+    assert "[retry]" in log_text
+    # pipeline 判定：仅检查子进程输出行（启动命令行本身含 marker 关键字）
+    stdout_lines = [ln for ln in log_text.splitlines() if "[fake stdout" in ln or ln.startswith("[retry]")]
+    assert MINERU_PIPELINE_LOG_MARKER not in "\n".join(stdout_lines)
+
+    status = read_status(job_dir)
+    assert status.get("status") == "failed"
+    message = str(status.get("message") or "")
+    assert "提取启动阶段失败" in message, f"中文可读错误缺失：{message}"
+    assert "退出码" in message and "1" in message
+    assert "原始日志尾部" in message or "原始错误见日志" in message
+    assert message  # 非空
+
+
+def test_startup_failure_with_marker_retries_once_then_no_second_retry(monkeypatch, tmp_path, pdf_source, job_root):
+    """G6-2 边界：第一次失败无 marker → 触发重试；重试后才有 marker 但 rc≠0 →
+    不二次重试，走 summarize_mineru_failure（有 marker 的中途失败路径）。"""
+    sequence = [1, 1]
+    calls = []
+
+    def run_with_marker(cmd, **kwargs):
+        n = len(calls)
+        rc = int(sequence[min(n, len(sequence) - 1)])
+        calls.append({"cmd": cmd, "kwargs": kwargs})
+        handle = kwargs.get("stdout")
+        if handle is not None and kwargs.get("text"):
+            if n == 1:  # 第二次调用（重试）时写入 pipeline 行
+                handle.write(f"[fake] {MINERU_PIPELINE_LOG_MARKER} started\n")
+            handle.write(f"[fake stdout call#{n}] rc={rc}\n")
+            handle.flush()
+        return FakeResult(returncode=rc)
+
+    monkeypatch.setenv("METHODOLOGY_MINERU_GPU_PREFLIGHT", "0")
+    monkeypatch.setattr(subprocess, "run", run_with_marker)
+    monkeypatch.setattr(mr, "_preflight_executable", lambda cmd, hint, name: None)
+    monkeypatch.setattr(
+        mr,
+        "summarize_mineru_failure",
+        lambda log_path, rc: f"CUDA out of memory fake path rc={rc}",
+    )
+    job_dir = job_root / "testjob"
+    run_uploaded_methodology_job(
+        job_id="testjob",
+        source_path=pdf_source,
+        paper_id="paper1",
+        family="",
+        output_dir=tmp_path / "out",
+        llm_model="deepseek-v4-flash",
+        llm_request_timeout=75,
+        sandbox=False,
+        sandbox_run_id="",
+    )
+    log_text = (job_dir / "job.log").read_text(encoding="utf-8")
+    assert MINERU_PIPELINE_LOG_MARKER in log_text
+    assert log_text.count("[retry]") == 1, "有 pipeline 行后不得再次重试"
+    # 中途失败（有 marker）→ 原错误路径（CUDA 中文错误）
+    status = read_status(job_dir)
+    assert status.get("status") == "failed"
+    assert "CUDA" in str(status.get("message") or "")
+
+
+# ---------------------------------------------------------------------------
+# 修复点 4：有 pipeline 行的中途失败不重试
+# ---------------------------------------------------------------------------
+
+def test_mid_failure_with_pipeline_line_no_retry(monkeypatch, tmp_path, pdf_source, job_root):
+    sequence = [1]
+    calls = []
+
+    def run_with_marker(cmd, **kwargs):
+        n = len(calls)
+        rc = int(sequence[min(n, len(sequence) - 1)])
+        calls.append({"cmd": cmd, "kwargs": kwargs})
+        handle = kwargs.get("stdout")
+        if handle is not None and kwargs.get("text"):
+            if n == 0:  # 第一次调用即写 pipeline 行
+                handle.write(f"[fake] {MINERU_PIPELINE_LOG_MARKER} started\n")
+            handle.write(f"[fake stdout call#{n}] rc={rc}\n")
+            handle.flush()
+        return FakeResult(returncode=rc)
+
+    monkeypatch.setenv("METHODOLOGY_MINERU_GPU_PREFLIGHT", "0")
+    monkeypatch.setattr(subprocess, "run", run_with_marker)
+    monkeypatch.setattr(mr, "_preflight_executable", lambda cmd, hint, name: None)
+    monkeypatch.setattr(
+        mr,
+        "summarize_mineru_failure",
+        lambda log_path, rc: f"CUDA out of memory fake path rc={rc}",
+    )
+    job_dir = job_root / "testjob"
+    run_uploaded_methodology_job(
+        job_id="testjob",
+        source_path=pdf_source,
+        paper_id="paper1",
+        family="",
+        output_dir=tmp_path / "out",
+        llm_model="deepseek-v4-flash",
+        llm_request_timeout=75,
+        sandbox=False,
+        sandbox_run_id="",
+    )
+    log_text = (job_dir / "job.log").read_text(encoding="utf-8")
+    assert MINERU_PIPELINE_LOG_MARKER in log_text
+    assert "[retry]" not in log_text, "已有 pipeline 行的中途失败不得触发重试"
+    # 原始启动块 = 1 条 `$ env ...` + 1 条 `$ <cmd>`；无重试追加
+    assert len([ln for ln in log_text.splitlines() if ln.startswith("$ /")]) == 1
+    status = read_status(job_dir)
+    assert status.get("status") == "failed"
+    assert "CUDA" in str(status.get("message") or "")
+
+
+# ---------------------------------------------------------------------------
+# JSON 直接投递（跳过 mineru 段）不重试、不破坏状态机
+# ---------------------------------------------------------------------------
+
+def test_json_source_skips_mineru_no_retry(monkeypatch, tmp_path, json_source, fake_success, job_root):
+    runner = RunnerRecorder(sequence=[0])  # 仅 LLM 批次一次
+    job_dir = run_job(monkeypatch, tmp_path, runner, json_source, fake_success, job_root)
+    log_text = (job_dir / "job.log").read_text(encoding="utf-8")
+    assert "[retry]" not in log_text
+    assert MINERU_PIPELINE_LOG_MARKER not in log_text
+    status = read_status(job_dir)
+    assert status.get("status") == "completed"
+
+
+# ---------------------------------------------------------------------------
+# 工具函数单元
+# ---------------------------------------------------------------------------
+
+def test_job_log_has_pipeline_line(tmp_path):
+    log_path = tmp_path / "job.log"
+    assert job_log_has_pipeline_line(log_path) is False
+    log_path.write_text("noise\n", encoding="utf-8")
+    assert job_log_has_pipeline_line(log_path) is False
+    log_path.write_text(f"2026-09-11 [INFO] {MINERU_PIPELINE_LOG_MARKER} begin\n", encoding="utf-8")
+    assert job_log_has_pipeline_line(log_path) is True
+
+
+def test_summarize_startup_failure_readable(tmp_path):
+    log_path = tmp_path / "job.log"
+    log_path.write_text("line one\nline two\nTraceback: boom\n", encoding="utf-8")
+    msg = summarize_mineru_startup_failure(log_path, 1, retried=True)
+    assert "提取启动阶段失败" in msg
+    assert "重试后" in msg
+    assert "退出码 1" in msg
+    assert "Traceback: boom" in msg
+    missing = summarize_mineru_startup_failure(tmp_path / "nope.log", 1, retried=False)
+    assert "提取启动阶段失败" in missing
+    assert "重试后" not in missing
diff --git a/agent/methodology/api/methodology_routes.py b/agent/methodology/api/methodology_routes.py
old mode 100755
new mode 100644
index e9f1f7a..06137a4
--- a/agent/methodology/api/methodology_routes.py
+++ b/agent/methodology/api/methodology_routes.py
@@ -351,6 +351,22 @@ def mineru_gpu_preflight_error(
     )
 
 
+# 缺陷 #38（2026-09-11）：成功任务的 job.log 必然有 doc_analyze_streaming 行（176 实测）。
+# returncode≠0 且日志中无此行 = 启动期失败（pipeline 起来前就死了，真因被块缓冲吞掉）。
+# 判定条件（methodology 与 legacy-rag 两处 build_mineru_runtime/run_uploaded_methodology_job 同改）
+MINERU_PIPELINE_LOG_MARKER = "doc_analyze_streaming"
+MINERU_STARTUP_RETRIES = 1
+
+
+def job_log_has_pipeline_line(log_path: Path) -> bool:
+    """job.log 是否已出现 pipeline 日志行（def _job_log_has_pipeline_line）。"""
+    try:
+        text = log_path.read_text(encoding="utf-8", errors="replace")
+    except OSError:
+        return False
+    return MINERU_PIPELINE_LOG_MARKER in text
+
+
 def summarize_mineru_failure(log_path: Path, returncode: int) -> str:
     message = f"MinerU failed with exit code {returncode}; see {log_path}"
     if not log_path.exists():
@@ -368,6 +384,28 @@ def summarize_mineru_failure(log_path: Path, returncode: int) -> str:
     return message
 
 
+def summarize_mineru_startup_failure(log_path: Path, returncode: int, retried: bool) -> str:
+    """启动期失败（无 pipeline 行）的中文可读错误（缺陷 #38）。
+
+    附上 job.log 中可见的原始错误尾部，让用户能看到真因（修复前真因
+    被 stdout 块缓冲 + CLI 立即杀 API 进程永久丢失）。
+    """
+    tail = ""
+    try:
+        text = log_path.read_text(encoding="utf-8", errors="replace")
+        lines = [ln for ln in text.splitlines() if ln.strip()]
+        if lines:
+            tail = " | ".join(ln.strip() for ln in lines[-3:])[:500]
+    except OSError:
+        tail = ""
+    detail = f"原始日志尾部：{tail}" if tail else f"原始错误见日志：{log_path}"
+    action = "重试后" if retried else ""
+    return (
+        f"提取启动阶段失败{action}，可能受主机 GPU/显示重配置干扰，请重试"
+        f"（MinerU 退出码 {returncode}；{detail}）。"
+    )
+
+
 def build_mineru_runtime() -> tuple[list[str], dict[str, str], dict[str, str]]:
     env = os.environ.copy()
     env.setdefault("MINERU_MODEL_SOURCE", "modelscope")
@@ -384,6 +422,9 @@ def build_mineru_runtime() -> tuple[list[str], dict[str, str], dict[str, str]]:
     else:
         env.setdefault("CUDA_VISIBLE_DEVICES", "0")
     env.setdefault("TORCH_CUDNN_V8_API_DISABLED", "1")
+    # 缺陷 #38（2026-09-11）：子进程 API 的异常日志即写即落（不再被 stdout
+    # 块缓冲吞掉）——CLI 失败后立即杀 API 进程前真因已落到 job.log。
+    env["PYTHONUNBUFFERED"] = "1"
     # mineru 3.1.13 CLI 陷阱（部署 agent 实测）：带 MINERU_API_SHUTDOWN_ON_STDIN_EOF=1
     # 时，stdin EOF 会让 API 进程静默 shutdown——subprocess 调用必须 stdin=DEVNULL
     env.setdefault("MINERU_API_SHUTDOWN_ON_STDIN_EOF", "0")
@@ -566,7 +607,34 @@ def run_uploaded_methodology_job(
                     stdout=handle,
                     stderr=subprocess.STDOUT,
                 )
+            # 缺陷 #38：启动期失败自动重试 1 次——returncode≠0 且 job.log 无
+            # doc_analyze_streaming 行 = pipeline 起来前就失败（真因曾因块缓冲
+            # 丢失）。重试的 stdout 继续 append 到同一 job.log，日志/状态留
+            # retry 标记（G6-2：已有 pipeline 行的中途失败不得触发重试）。
+            startup_retried = False
+            if result.returncode != 0 and not job_log_has_pipeline_line(log_path):
+                startup_retried = True
+                with log_path.open("a", encoding="utf-8") as handle:
+                    handle.write(
+                        f"[retry] MinerU 启动期失败（exit {result.returncode}，"
+                        "无 pipeline 阶段日志），自动重试 1/1\n"
+                    )
+                    handle.write("$ " + " ".join(cmd) + "\n")
+                    handle.flush()
+                    result = subprocess.run(
+                        cmd,
+                        cwd=PROJECT_ROOT,
+                        env=mineru_env,
+                        text=True,
+                        stdin=subprocess.DEVNULL,
+                        stdout=handle,
+                        stderr=subprocess.STDOUT,
+                    )
             if result.returncode != 0:
+                if not job_log_has_pipeline_line(log_path):
+                    raise RuntimeError(
+                        summarize_mineru_startup_failure(log_path, result.returncode, retried=startup_retried)
+                    )
                 raise RuntimeError(summarize_mineru_failure(log_path, result.returncode))
             matches = sorted(mineru_out.rglob("*_content_list.json"))
             if not matches:
diff --git a/agent/methodology/api/test_mineru_retry.py b/agent/methodology/api/test_mineru_retry.py
new file mode 100644
index 0000000..cad974c
--- /dev/null
+++ b/agent/methodology/api/test_mineru_retry.py
@@ -0,0 +1,328 @@
+"""test_mineru_retry.py —— MinerU 启动期失败重试与可读错误（缺陷 #38，验收 §4 + 回归 G6）。
+
+运行（worktree 根目录，无需部署、不触服务）：
+  /home/xzq/5090/teamWorkspace/zh/plantai-rag-new-venv/bin/python -m pytest agent/methodology/api/test_mineru_retry.py -v
+  /home/xzq/5090/teamWorkspace/zh/plantai-rag-new-venv/bin/python -m pytest agent/legacy-rag/api/test_mineru_retry.py -v
+说明：agent/methodology 与 agent/legacy-rag 共用同一份 routes（文件名同、内容同），
+同一测试文件复制两处、两处同跑（各用各树的 src/）。
+
+断言：
+1. build_mineru_runtime 返回 env 含 PYTHONUNBUFFERED=1；
+2. subprocess.run 首次失败 + job.log 无 pipeline 行 → 自动重试 1 次
+   （两次启动行、job.log 有 [retry] 标记，第二次 stdout 追加同一 job.log）；
+3. 重试仍失败 → status failed + 中文可读「提取启动阶段失败」+ 原始错误尾部；
+4. 已有 pipeline 行的中途失败 → 不重试，走原错误路径（含 CUDA OOM 中文错误）。
+"""
+from __future__ import annotations
+
+import json
+import subprocess
+from dataclasses import dataclass, field
+from pathlib import Path
+
+import pytest
+
+import api.methodology_routes as mr
+
+from api.methodology_routes import (
+    MINERU_PIPELINE_LOG_MARKER,
+    build_mineru_runtime,
+    job_log_has_pipeline_line,
+    run_uploaded_methodology_job,
+    summarize_mineru_startup_failure,
+)
+
+
+# ---------------------------------------------------------------------------
+# 固定夹具
+# ---------------------------------------------------------------------------
+
+@dataclass
+class FakeResult:
+    returncode: int
+
+
+@dataclass
+class RunnerRecorder:
+    """按调用顺序返回预设 returncode；每次调用把 stdout 写进所给 handle（模拟子进程日志）。"""
+
+    sequence: list
+    calls: list = field(default_factory=list)
+
+    def run(self, cmd, **kwargs):
+        n = len(self.calls)
+        self.calls.append({"cmd": cmd, "kwargs": kwargs})
+        rc = int(self.sequence[min(n, len(self.sequence) - 1)])
+        handle = kwargs.get("stdout")
+        if handle is not None and kwargs.get("text"):
+            handle.write(f"[fake stdout call#{n}] rc={rc}\n")
+            handle.flush()
+        return FakeResult(returncode=rc)
+
+
+@pytest.fixture
+def fake_success(monkeypatch):
+    """LLM 批次与 bbav5 配置阶段全部假成功（不触真实脚本/目录/真实函数）。
+
+    - gf_python_path / _preflight_executable 桩掉（不依赖真实 python 路径）
+    - build_bbav5_methodology_config 桩掉（不要求真实批次产物目录）
+    - figure_table 阶段走 routes 内部 try/except，不阻断
+    """
+    captured = {}
+
+    def fake_batch(cmd, **kwargs):
+        captured["batch_cmd"] = cmd
+        return FakeResult(returncode=0)
+
+    monkeypatch.setattr(mr, "gf_python_path", lambda: Path("/usr/bin/python3"))
+    monkeypatch.setattr(mr, "_preflight_executable", lambda cmd, hint, name: None)
+    monkeypatch.setattr(mr.subprocess, "run", fake_batch)
+    monkeypatch.setattr(
+        mr,
+        "build_bbav5_methodology_config",
+        lambda output_dir: captured.setdefault("config", {}),
+    )
+    return captured
+
+
+@pytest.fixture
+def job_root(monkeypatch, tmp_path):
+    """把 UPLOAD_JOBS_DIR 重定向到 tmp（绝不写真实数据目录）。"""
+    root = tmp_path / "upload_jobs"
+    monkeypatch.setattr(mr, "UPLOAD_JOBS_DIR", root)
+    return root
+
+
+def read_status(job_dir: Path) -> dict:
+    return json.loads((job_dir / "status.json").read_text(encoding="utf-8"))
+
+
+def run_job(monkeypatch, tmp_path, runner: RunnerRecorder, source_path: Path, fake_success, job_root):
+    """以 runner 重绑 subprocess.run 跑完整任务流（fake_success 保证后续阶段假成功）。
+
+    注意：fake_success 已把 mr.subprocess.run 桩为假批次；这里再次 setattr
+    为 runner（同一模块对象），此后所有子进程调用（mineru 两次 + LLM 批次）
+    都由 runner 按 sequence 依次应答——测试 sequence 需覆盖全部调用。
+    """
+    job_dir = job_root / "testjob"
+    # 模拟 mineru 成功产出：在 job/mineru 下伪造 *_content_list.json
+    # （测试不真正执行 mineru）
+    (job_dir / "mineru").mkdir(parents=True, exist_ok=True)
+    (job_dir / "mineru" / "fake_content_list.json").write_text("{}", encoding="utf-8")
+    # GPU 预检禁用（否则 nvidia-smi 会消耗 runner 的 sequence 槽位）
+    monkeypatch.setenv("METHODOLOGY_MINERU_GPU_PREFLIGHT", "0")
+    monkeypatch.setattr(subprocess, "run", runner.run)
+    run_uploaded_methodology_job(
+        job_id="testjob",
+        source_path=source_path,
+        paper_id="paper1",
+        family="",
+        output_dir=tmp_path / "out",
+        llm_model="deepseek-v4-flash",
+        llm_request_timeout=75,
+        sandbox=False,
+        sandbox_run_id="",
+    )
+    return job_dir
+
+
+@pytest.fixture
+def pdf_source(tmp_path):
+    src = tmp_path / "sample.pdf"
+    src.write_bytes(b"%PDF-1.4\nfake pdf bytes for startup-failure test\n")
+    return src
+
+
+@pytest.fixture
+def json_source(tmp_path):
+    src = tmp_path / "sample.json"
+    src.write_text("{}", encoding="utf-8")
+    return src
+
+
+# ---------------------------------------------------------------------------
+# 修复点 1：env 注入 PYTHONUNBUFFERED
+# ---------------------------------------------------------------------------
+
+def test_build_mineru_runtime_env_unbuffered(monkeypatch):
+    monkeypatch.delenv("MINERU_CMD", raising=False)
+    cmd, env, runtime = build_mineru_runtime()
+    assert env.get("PYTHONUNBUFFERED") == "1"
+    assert cmd  # 默认 MINERU_CMD 路径（不要求存在，build 阶段不校验）
+
+
+# ---------------------------------------------------------------------------
+# 修复点 2：启动期失败自动重试 1 次
+# ---------------------------------------------------------------------------
+
+def test_startup_failure_retries_once(monkeypatch, tmp_path, pdf_source, fake_success, job_root):
+    runner = RunnerRecorder(sequence=[1, 0, 0])  # mineru 首调失败 → 重试成功 → LLM 批次成功
+    job_dir = run_job(monkeypatch, tmp_path, runner, pdf_source, fake_success, job_root)
+
+    log_text = (job_dir / "job.log").read_text(encoding="utf-8")
+    startup_lines = [ln for ln in log_text.splitlines() if ln.startswith("$ ")]
+    assert len(startup_lines) >= 2, f"应出现两次 mineru 启动（1 次自动重试），实际：{startup_lines}"
+    assert "[retry]" in log_text, "job.log 应有可识别重试标记"
+    # 第二次 subprocess.run 的 stdout 追加到同一 job.log
+    assert "[fake stdout call#0]" in log_text
+    assert "[fake stdout call#1]" in log_text
+
+    status = read_status(job_dir)
+    assert status.get("status") == "completed"
+
+
+def test_startup_failure_retry_then_fail_readable_error(monkeypatch, tmp_path, pdf_source, fake_success, job_root):
+    runner = RunnerRecorder(sequence=[1, 1])  # 两次都失败，且无 pipeline 行
+    job_dir = run_job(monkeypatch, tmp_path, runner, pdf_source, fake_success, job_root)
+
+    log_text = (job_dir / "job.log").read_text(encoding="utf-8")
+    assert len([ln for ln in log_text.splitlines() if ln.startswith("$ /")]) == 2  # 两次 mineru 启动
+    assert "[retry]" in log_text
+    # pipeline 判定：仅检查子进程输出行（启动命令行本身含 marker 关键字）
+    stdout_lines = [ln for ln in log_text.splitlines() if "[fake stdout" in ln or ln.startswith("[retry]")]
+    assert MINERU_PIPELINE_LOG_MARKER not in "\n".join(stdout_lines)
+
+    status = read_status(job_dir)
+    assert status.get("status") == "failed"
+    message = str(status.get("message") or "")
+    assert "提取启动阶段失败" in message, f"中文可读错误缺失：{message}"
+    assert "退出码" in message and "1" in message
+    assert "原始日志尾部" in message or "原始错误见日志" in message
+    assert message  # 非空
+
+
+def test_startup_failure_with_marker_retries_once_then_no_second_retry(monkeypatch, tmp_path, pdf_source, job_root):
+    """G6-2 边界：第一次失败无 marker → 触发重试；重试后才有 marker 但 rc≠0 →
+    不二次重试，走 summarize_mineru_failure（有 marker 的中途失败路径）。"""
+    sequence = [1, 1]
+    calls = []
+
+    def run_with_marker(cmd, **kwargs):
+        n = len(calls)
+        rc = int(sequence[min(n, len(sequence) - 1)])
+        calls.append({"cmd": cmd, "kwargs": kwargs})
+        handle = kwargs.get("stdout")
+        if handle is not None and kwargs.get("text"):
+            if n == 1:  # 第二次调用（重试）时写入 pipeline 行
+                handle.write(f"[fake] {MINERU_PIPELINE_LOG_MARKER} started\n")
+            handle.write(f"[fake stdout call#{n}] rc={rc}\n")
+            handle.flush()
+        return FakeResult(returncode=rc)
+
+    monkeypatch.setenv("METHODOLOGY_MINERU_GPU_PREFLIGHT", "0")
+    monkeypatch.setattr(subprocess, "run", run_with_marker)
+    monkeypatch.setattr(mr, "_preflight_executable", lambda cmd, hint, name: None)
+    monkeypatch.setattr(
+        mr,
+        "summarize_mineru_failure",
+        lambda log_path, rc: f"CUDA out of memory fake path rc={rc}",
+    )
+    job_dir = job_root / "testjob"
+    run_uploaded_methodology_job(
+        job_id="testjob",
+        source_path=pdf_source,
+        paper_id="paper1",
+        family="",
+        output_dir=tmp_path / "out",
+        llm_model="deepseek-v4-flash",
+        llm_request_timeout=75,
+        sandbox=False,
+        sandbox_run_id="",
+    )
+    log_text = (job_dir / "job.log").read_text(encoding="utf-8")
+    assert MINERU_PIPELINE_LOG_MARKER in log_text
+    assert log_text.count("[retry]") == 1, "有 pipeline 行后不得再次重试"
+    # 中途失败（有 marker）→ 原错误路径（CUDA 中文错误）
+    status = read_status(job_dir)
+    assert status.get("status") == "failed"
+    assert "CUDA" in str(status.get("message") or "")
+
+
+# ---------------------------------------------------------------------------
+# 修复点 4：有 pipeline 行的中途失败不重试
+# ---------------------------------------------------------------------------
+
+def test_mid_failure_with_pipeline_line_no_retry(monkeypatch, tmp_path, pdf_source, job_root):
+    sequence = [1]
+    calls = []
+
+    def run_with_marker(cmd, **kwargs):
+        n = len(calls)
+        rc = int(sequence[min(n, len(sequence) - 1)])
+        calls.append({"cmd": cmd, "kwargs": kwargs})
+        handle = kwargs.get("stdout")
+        if handle is not None and kwargs.get("text"):
+            if n == 0:  # 第一次调用即写 pipeline 行
+                handle.write(f"[fake] {MINERU_PIPELINE_LOG_MARKER} started\n")
+            handle.write(f"[fake stdout call#{n}] rc={rc}\n")
+            handle.flush()
+        return FakeResult(returncode=rc)
+
+    monkeypatch.setenv("METHODOLOGY_MINERU_GPU_PREFLIGHT", "0")
+    monkeypatch.setattr(subprocess, "run", run_with_marker)
+    monkeypatch.setattr(mr, "_preflight_executable", lambda cmd, hint, name: None)
+    monkeypatch.setattr(
+        mr,
+        "summarize_mineru_failure",
+        lambda log_path, rc: f"CUDA out of memory fake path rc={rc}",
+    )
+    job_dir = job_root / "testjob"
+    run_uploaded_methodology_job(
+        job_id="testjob",
+        source_path=pdf_source,
+        paper_id="paper1",
+        family="",
+        output_dir=tmp_path / "out",
+        llm_model="deepseek-v4-flash",
+        llm_request_timeout=75,
+        sandbox=False,
+        sandbox_run_id="",
+    )
+    log_text = (job_dir / "job.log").read_text(encoding="utf-8")
+    assert MINERU_PIPELINE_LOG_MARKER in log_text
+    assert "[retry]" not in log_text, "已有 pipeline 行的中途失败不得触发重试"
+    # 原始启动块 = 1 条 `$ env ...` + 1 条 `$ <cmd>`；无重试追加
+    assert len([ln for ln in log_text.splitlines() if ln.startswith("$ /")]) == 1
+    status = read_status(job_dir)
+    assert status.get("status") == "failed"
+    assert "CUDA" in str(status.get("message") or "")
+
+
+# ---------------------------------------------------------------------------
+# JSON 直接投递（跳过 mineru 段）不重试、不破坏状态机
+# ---------------------------------------------------------------------------
+
+def test_json_source_skips_mineru_no_retry(monkeypatch, tmp_path, json_source, fake_success, job_root):
+    runner = RunnerRecorder(sequence=[0])  # 仅 LLM 批次一次
+    job_dir = run_job(monkeypatch, tmp_path, runner, json_source, fake_success, job_root)
+    log_text = (job_dir / "job.log").read_text(encoding="utf-8")
+    assert "[retry]" not in log_text
+    assert MINERU_PIPELINE_LOG_MARKER not in log_text
+    status = read_status(job_dir)
+    assert status.get("status") == "completed"
+
+
+# ---------------------------------------------------------------------------
+# 工具函数单元
+# ---------------------------------------------------------------------------
+
+def test_job_log_has_pipeline_line(tmp_path):
+    log_path = tmp_path / "job.log"
+    assert job_log_has_pipeline_line(log_path) is False
+    log_path.write_text("noise\n", encoding="utf-8")
+    assert job_log_has_pipeline_line(log_path) is False
+    log_path.write_text(f"2026-09-11 [INFO] {MINERU_PIPELINE_LOG_MARKER} begin\n", encoding="utf-8")
+    assert job_log_has_pipeline_line(log_path) is True
+
+
+def test_summarize_startup_failure_readable(tmp_path):
+    log_path = tmp_path / "job.log"
+    log_path.write_text("line one\nline two\nTraceback: boom\n", encoding="utf-8")
+    msg = summarize_mineru_startup_failure(log_path, 1, retried=True)
+    assert "提取启动阶段失败" in msg
+    assert "重试后" in msg
+    assert "退出码 1" in msg
+    assert "Traceback: boom" in msg
+    missing = summarize_mineru_startup_failure(tmp_path / "nope.log", 1, retried=False)
+    assert "提取启动阶段失败" in missing
+    assert "重试后" not in missing