首次提交

2026-09-28 16:09:31 +08:00
commit 4a6839319a
71 changed files with 12144 additions and 0 deletions
--- a/scripts/apply_update.sh
+++ b/scripts/apply_update.sh
@@ -0,0 +1,167 @@
+#!/usr/bin/env bash
+# 一键更新脚本(在线上【项目根目录】运行)。
+# 作用: 用 _new/ 里的新代码覆盖改过的文件, 并把新配置项【合并】进现有 config.json
+#       (只加缺失的键, 绝不动你已有的密钥/路径/已调过的值), 然后自检。
+# 用法: 解压更新包到项目根目录后, 执行  bash apply_update.sh
+set -u
+
+ROOT="$(pwd)"
+TS="$(date +%Y%m%d_%H%M%S)"
+BAK="_backup_${TS}"
+NEW="_new"
+
+echo "==> 项目目录: $ROOT"
+
+if [ ! -d "tool" ] || [ ! -f "config/config.json" ]; then
+  echo "✗ 当前目录不像项目根(缺 tool/ 或 config/config.json)。请 cd 到项目根再运行。"
+  exit 1
+fi
+if [ ! -d "$NEW" ]; then
+  echo "✗ 找不到 $NEW/ 目录。请先把更新包解压到项目根目录。"
+  exit 1
+fi
+
+# 0. 更新前先把本机 数据.jsonl + 超长待重跑列表 备份到 COS(不等每天10点的自动备份)。
+#    更新代码本身不会碰 数据.jsonl, 这一步纯粹是"顺手让全局数据保持最新"(比如全局去重要用),
+#    COS 访问失败不阻断更新, 只提示。
+echo "==> 更新前备份 数据.jsonl / 超长待重跑列表 到 COS"
+python3 tools/upload_daily.py || echo "   ⚠ 备份失败(不影响本次更新继续), 稍后可手动重跑: python3 tools/upload_daily.py"
+
+FILES="main.py run_worker.py tool/process.py tool/merge.py tool/getsrt.py tool/whisper_manager.py tool/VideoProcessing/multithread_algo.py tool/Ocr/tencentcloud_ocr.py tool/job_cos.py config/config_utils.py"
+
+# 1. 备份
+echo "==> 备份到 $BAK/"
+mkdir -p "$BAK/tool/VideoProcessing" "$BAK/tool/Ocr" "$BAK/config"
+for f in $FILES; do
+  [ -f "$f" ] && cp -p "$f" "$BAK/$f"
+done
+cp -p config/config.json "$BAK/config/config.json"
+
+# 2. 覆盖代码文件
+echo "==> 更新代码文件"
+for f in $FILES; do
+  if [ ! -f "$NEW/$f" ]; then
+    echo "✗ 更新包缺少 $NEW/$f, 中止(已备份, 未改动)"; exit 1
+  fi
+  cp -p "$NEW/$f" "$f"
+  echo "   ✓ $f"
+done
+
+# 2b. 附带的助手脚本(有就拷, 没有不报错)
+for f in tools/拆分超长.py tools/seed_jobs.py tools/upload_daily.py tools/release_machine.py tools/patch_config.py; do
+  if [ -f "$NEW/$f" ]; then
+    mkdir -p "$(dirname "$f")"
+    cp -p "$NEW/$f" "$f" && echo "   ✓ $f"
+  fi
+done
+
+# 3. 合并配置(只加缺失键, 不覆盖已有值; 报告与推荐值不同的键)
+echo "==> 合并 config.json 新配置项(保留你已有的密钥/路径/设置)"
+python3 - "$ROOT/config/config.json" <<'PYEOF'
+import json, sys
+p = sys.argv[1]
+d = json.load(open(p, encoding="utf-8"))
+au = d.setdefault("audio", {})
+sf = d.setdefault("splitframe", {})
+sub = d.setdefault("subtitle", {})
+ocr = d.setdefault("tencentcloud", {}).setdefault("ocr", {})
+dp = d.setdefault("dispatch", {})
+_bkt = (d.get("tencentcloud", {}).get("cos", {}) or {}).get("bucket", "asr-dig-data-1253181912")
+
+NEW_KEYS = [
+    (ocr, "max_attempts",           4),
+    (ocr, "timeout",                30),
+    (ocr, "dispatch_log_every",     1000),
+    (au,  "cpu_threads",            2),
+    (sf,  "crop_bottom_ratio",      None),
+    (sub, "min_duration_seconds",   300),
+    (sub, "no_composite",           False),
+    (sub, "merge_workers",          8),
+    (sub, "max_tiles_per_composite", 24),
+    (sub, "ocr_max_image_side",     4000),
+    (sub, "cv_detect",              True),
+    (sub, "cv_use_region",          False),
+    (sub, "filter_static",          True),
+    (sub, "static_frac",            0.25),
+    (sub, "filter_lowconf",         True),
+    (sub, "lowconf_thr",            80),
+    (sub, "lowconf_filter_frac",    0.6),
+    (sub, "lowconf_min_det",        200),
+    (sub, "strip_static_prefix",    False),
+    (sub, "filter_watermark_pos",   True),
+    (sub, "wm_dominant_frac",       0.25),
+    (sub, "wm_min_count",           20),
+    (sub, "split_far_boxes",       True),
+    (sub, "max_duration_seconds",   5400),
+    (sub, "long_video_list",        "./超长待重跑.txt"),
+    (sub, "ocr_no_drop",            True),
+    (sub, "ocr_max_passes",         5),
+    (sub, "ocr_retry_sleep",        2),
+    (sub, "ocr_fail_list",          "./OCR失败待重跑.txt"),
+    (sub, "no_subtitle_reasons",    ["字幕太少","冗余文字过多","冗余文字太多","无字幕区域","非字幕内容","置信度太低","未检测","特殊字体","视频过短","视频过长"]),
+    (sf,  "ffmpeg_timeout",          3600),
+    (sf,  "ffmpeg_threads",          "auto"),
+    (sf,  "gpu_skip_indices",        []),
+    (sf,  "gpu_auto_select",         True),
+    (sf,  "gpu_busy_mem_mb",         4096),
+    (dp,  "enabled",                 True),
+    (dp,  "source",                  "cos"),
+    (dp,  "run_long_video",          False),
+    (dp,  "auto_fallback_long_video", True),
+    (dp,  "global_dedup_on_startup",  True),
+    (dp,  "machine_id",              ""),
+    (dp,  "jobs_root",               "cos://%s/_jobs" % _bkt),
+    (dp,  "shard_size",              2000),
+    (dp,  "keep_alive",              True),
+    (dp,  "idle_poll_seconds",       60),
+    (dp,  "claim_shuffle",           True),
+    (dp,  "claim_match",             ""),
+    (dp,  "heartbeat_seconds",       300),
+    (dp,  "daily_backup_path",       "cos://%s/_jobs/daily" % _bkt),
+    (dp,  "long_list_backup_path",   "cos://%s/_jobs/long_video" % _bkt),
+    (dp,  "monitor_root",            "cos://%s/_jobs/monitor" % _bkt),
+    (dp,  "long_list_clear_after_upload", True),
+    (dp,  "long_list_min_batch",     1000),
+    (dp,  "daily_backup_hour",       10),
+    (dp,  "stop_flag_key",           "_jobs/STOP"),
+]
+added, differ = [], []
+for sec, k, v in NEW_KEYS:
+    if k not in sec:
+        sec[k] = v; added.append((k, v))
+    elif sec[k] != v:
+        differ.append((k, sec[k], v))
+
+json.dump(d, open(p, "w", encoding="utf-8"), ensure_ascii=False, indent=2)
+print("   新增键:", ", ".join(f"{k}={v}" for k, v in added) or "(无, 已都存在)")
+if differ:
+    print("   ⚠ 以下键你线上已有值, 未改动(如想用推荐值请手动改):")
+    for k, cur, rec in differ:
+        print(f"       {k}: 当前={cur}  推荐={rec}")
+PYEOF
+if [ $? -ne 0 ]; then
+  echo "✗ 配置合并失败, 正在回滚 config.json"
+  cp -p "$BAK/config/config.json" config/config.json
+  exit 1
+fi
+
+# 4. 自检
+echo "==> 自检 import"
+if python3 -c "import main, run_worker, tool.job_cos, tool.merge, tool.process, tool.getsrt, tool.whisper_manager, tool.VideoProcessing.multithread_algo; print('   ✓ import OK')"; then
+  :
+else
+  echo "✗ import 失败! 正在回滚所有改动"
+  for f in $FILES; do cp -p "$BAK/$f" "$f"; done
+  cp -p "$BAK/config/config.json" config/config.json
+  echo "已回滚到更新前。备份在 $BAK/"
+  exit 1
+fi
+
+echo ""
+echo "============================================="
+echo "✓ 更新完成。备份在: $BAK/"
+echo "  回退: 把 $BAK/ 里的文件拷回原位即可。"
+echo ""
+echo "下一步: 先小批量验证, 没问题再上量。"
+echo "  python main.py --source-list ./text2 --max-workers 32 --executor process"
+echo "============================================="
--- a/scripts/benchmark.sh
+++ b/scripts/benchmark.sh
@@ -0,0 +1,108 @@
+#!/usr/bin/env bash
+# 新机器上线前的算力/IO摸底测试(跟 scripts/diag.sh 不是一回事 —— diag.sh 是排查"已经在跑但
+# 出问题"的机器；这个是给"还没开始跑正式任务、想先估算这台机器大概能跑多快"的新机器用的)。
+# 测: CPU/内存/GPU(含真实硬解测试)/磁盘顺序读写速度/到OCR端点的网络延迟, 汇总成一份文件。
+#
+# 用法(项目根目录): bash scripts/benchmark.sh
+# 产物: benchmark_YYYYmmdd_HHMMSS.txt (把这个文件发回来)
+
+cd "$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)" || exit 1
+OUT="benchmark_$(date +%Y%m%d_%H%M%S).txt"
+PY="$(command -v python || command -v python3)"
+FFMPEG="$($PY -c "from config.config_utils import get_ffmpeg; print(get_ffmpeg())" 2>/dev/null || command -v ffmpeg)"
+
+exec > >(tee "$OUT") 2>&1
+
+echo "================ 算力/IO摸底 $(date) ================"
+echo "主机: $(hostname)"
+echo ""
+
+echo "========== 1. CPU =========="
+echo "[核数] $(nproc)"
+if command -v lscpu >/dev/null 2>&1; then
+  lscpu | grep -E "Model name|CPU\(s\)|MHz|Thread|Socket"
+fi
+echo ""
+
+echo "========== 2. 内存 =========="
+free -h
+echo ""
+
+echo "========== 3. GPU =========="
+if command -v nvidia-smi >/dev/null 2>&1; then
+  echo "[列表]"
+  nvidia-smi --query-gpu=index,name,memory.total,memory.used,utilization.gpu,driver_version --format=csv
+  echo ""
+  echo "[真实硬解测试: ffmpeg合成测试图案, 用CPU软编码生成素材(不碰GPU), 再用 -hwaccel cuda 解码,"
+  echo " 每张卡各测一遍 —— 生产流程只用GPU【解码】下载好的视频, 从不用GPU编码, 所以这里故意不测nvenc编码,"
+  echo " 只测跟生产完全一致的硬解路径]"
+  if [ -x "$FFMPEG" ] || command -v "$FFMPEG" >/dev/null 2>&1; then
+    n_gpu=$(nvidia-smi --query-gpu=index --format=csv,noheader | wc -l)
+    testfile="/tmp/_bench_src.mp4"
+    "$FFMPEG" -hide_banner -loglevel error -f lavfi -i "testsrc=size=1920x1080:rate=30:duration=10" \
+      -c:v libx264 -preset veryfast -f mp4 "$testfile" -y 2>&1 | tail -3
+    if [ -f "$testfile" ]; then
+      for dev in $(seq 0 $((n_gpu - 1))); do
+        echo "  -- GPU $dev --"
+        t0=$(date +%s.%N)
+        "$FFMPEG" -hide_banner -loglevel error -hwaccel cuda -hwaccel_device "$dev" \
+          -i "$testfile" -f null - 2>&1 | tail -5
+        t1=$(date +%s.%N)
+        echo "    硬解读回耗时: $(echo "$t1 - $t0" | bc 2>/dev/null || echo "?")s (10秒素材, 越接近0越好, 明显>10s说明硬解没吃上或GPU很忙; 有报错说明硬解不可用, 会自动回退CPU软解)"
+      done
+      rm -f "$testfile"
+    else
+      echo "  CPU软编码生成测试素材都失败了(跟GPU无关, 检查ffmpeg本身/libx264), 跳过硬解测试"
+    fi
+  else
+    echo "  找不到 ffmpeg, 跳过硬解测试"
+  fi
+else
+  echo "  没有 nvidia-smi(不是GPU机器, 或驱动没装)"
+fi
+echo ""
+
+echo "========== 4. 磁盘顺序读写速度(当前目录所在盘) =========="
+TESTFILE="./_bench_io.tmp"
+echo "[写] 顺序写1GB"
+sync
+dd if=/dev/zero of="$TESTFILE" bs=1M count=1024 conv=fdatasync 2>&1 | tail -3
+echo "[读] 顺序读(先清缓存再读更准, 没权限清缓存就直接读)"
+( sync && echo 3 > /proc/sys/vm/drop_caches ) 2>/dev/null
+dd if="$TESTFILE" of=/dev/null bs=1M 2>&1 | tail -3
+rm -f "$TESTFILE"
+echo ""
+echo "[磁盘剩余空间]"
+df -h .
+echo ""
+
+echo "========== 5. 网络: 到OCR端点的延迟/连通率 =========="
+if [ -f config/config.json ] && [ -n "$PY" ]; then
+  "$PY" tools/查端点存活.py --tcp --timeout 3 2>&1
+else
+  echo "  跳过(找不到 python 或 config/config.json)"
+fi
+echo ""
+
+echo "========== 6. 网络: 到COS的连通性 =========="
+if command -v coscli >/dev/null 2>&1 || [ -x ./coscli-linux-amd64 ]; then
+  COSCLI="$(command -v coscli || echo ./coscli-linux-amd64)"
+  COSCFG="$($PY -c "from config.config_utils import config;print((config.get_splitframe_config() or {}).get('cosconfig_path','.cos.yaml'))" 2>/dev/null || echo .cos.yaml)"
+  JOBS_ROOT="$($PY -c "from config.config_utils import config;print((config.get_dispatch_config() or {}).get('jobs_root',''))" 2>/dev/null)"
+  if [ -n "$JOBS_ROOT" ]; then
+    echo "[列 $JOBS_ROOT 耗时]"
+    t0=$(date +%s.%N)
+    "$COSCLI" -c "$COSCFG" ls "$JOBS_ROOT/" 2>&1 | tail -5
+    t1=$(date +%s.%N)
+    echo "  耗时: $(echo "$t1 - $t0" | bc 2>/dev/null || echo "?")s (纯网络往返延迟参考, 不是下载带宽)"
+  else
+    echo "  dispatch.jobs_root 未配置, 跳过"
+  fi
+else
+  echo "  找不到 coscli, 跳过"
+fi
+echo ""
+
+echo "================ 摸底结束 ================"
+echo ""
+echo ">>> 报告已保存到: $OUT  (把这个文件发回来)"
--- a/scripts/build_bundle.sh
+++ b/scripts/build_bundle.sh
@@ -0,0 +1,48 @@
+#!/usr/bin/env bash
+# =============================================================================
+# 一键构建离线部署包（在【能联网】的打包机上跑）
+#
+#   setup_server.sh(装环境) → make_portable.sh(自包含) → pack.sh(打包)  三步合一
+#
+# 用法:
+#   bash build_bundle.sh cpu            # CPU 版包：johnvansickle ffmpeg，老 glibc 也能跑
+#   bash build_bundle.sh gpu            # GPU 版包：BtbN cuda ffmpeg，目标机需 glibc>=2.28
+#   bash build_bundle.sh gpu --whisper-gpu   # GPU 版 + 把 CUDA 库打进包(目标机无系统 CUDA 也能 whisper GPU)
+#
+# 产物: /tmp/ocr_offline_bundle.tar.gz
+#       拷到目标机后:  bash deploy.sh /tmp/ocr_offline_bundle.tar.gz /opt
+# =============================================================================
+set -euo pipefail
+cd "$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+
+VARIANT="${1:-cpu}"               # cpu | gpu
+WHISPER_GPU_FLAG=0
+[ "${2:-}" = "--whisper-gpu" ] && WHISPER_GPU_FLAG=1
+
+log() { echo -e "\n\033[1;36m##### $* #####\033[0m"; }
+
+case "$VARIANT" in
+  cpu) FFV=cpu ;;
+  gpu) FFV=gpu ;;
+  *) echo "用法: bash build_bundle.sh [cpu|gpu] [--whisper-gpu]"; exit 1 ;;
+esac
+
+SD="$(cd "$(dirname "$0")" && pwd)"
+log "构建 $VARIANT 版离线包 (whisper_gpu=$WHISPER_GPU_FLAG)"
+
+# 1) 装基础环境(已装过会跳过大部分)
+log "1/3 setup_server.sh（装 gcc/uv/python/依赖）"
+bash "$SD/setup_server.sh"
+
+# 2) 转自包含（python+模型+ffmpeg 进项目）
+log "2/3 make_portable.sh（FFMPEG_VARIANT=$FFV）"
+FFMPEG_VARIANT="$FFV" WHISPER_GPU="$WHISPER_GPU_FLAG" bash "$SD/make_portable.sh"
+
+# 3) 打包（按变体命名，cpu/gpu 两个包不会互相覆盖）
+OUT="/tmp/ocr_bundle_${VARIANT}.tar.gz"
+log "3/3 pack.sh（打成 tar → $OUT）"
+bash "$SD/pack.sh" "$OUT"
+
+echo
+echo "✅ 完成。产物: $OUT"
+echo "   拷到目标机后:  bash $SD/deploy.sh $OUT /opt"
--- a/scripts/deploy.sh
+++ b/scripts/deploy.sh
@@ -0,0 +1,75 @@
+#!/usr/bin/env bash
+# =============================================================================
+# 离线机部署脚本（在【不能联网】的机器上跑）
+#
+# 自动识别打包模式：
+#  A) 自包含包（含项目内 python/）：解压到【任意目标路径】，默认当前目录。
+#       bash deploy.sh 包.tar.gz [目标父目录]      # 例: bash deploy.sh bundle.tar.gz /opt
+#  B) venv 包：解压到 / 还原到相同绝对路径（要求同路径/同用户/同架构）。
+#       bash deploy.sh 包.tar.gz
+#
+# 前提：与打包机【同 CPU 架构(x86_64)】且【glibc 不比打包机老】。
+# 离线机不需要 uv / python / gcc / 联网。
+# =============================================================================
+set -euo pipefail
+
+BUNDLE="${1:-/tmp/ocr_offline_bundle.tar.gz}"
+TARGET="${2:-$(pwd)}"      # 自包含模式的解压目标父目录
+log()  { echo -e "\n\033[1;32m==> $*\033[0m"; }
+warn() { echo -e "\033[1;33m[警告] $*\033[0m"; }
+
+[ -f "$BUNDLE" ] || { echo "找不到包: $BUNDLE"; exit 1; }
+
+log "[1/3] 兼容性 + 模式识别"
+echo "  架构: $(uname -m)（需与打包机一致，一般 x86_64）"
+echo "  glibc: $(ldd --version 2>/dev/null | head -1 | awk '{print $NF}')（不能比打包机老）"
+
+# 看包里第一个条目：以 / 或 data/ 之类绝对路径开头 = venv 模式；以纯目录名开头 = 自包含
+# 注意：head 关管道会让 tar 收到 SIGPIPE 返回非零，配合 set -e/pipefail 会误退出，故加 || true
+FIRST="$(tar tzf "$BUNDLE" 2>/dev/null | head -1 || true)"
+echo "  包内首条目: $FIRST"
+
+PROJDIR=""
+if printf '%s' "$FIRST" | grep -qE '^(/|root/|home/|data/|usr/|opt/)'; then
+  # ---------- venv 模式：解压到 / ----------
+  log "[2/3] venv 模式 → 解压到 / (还原到原绝对路径)"
+  tar xzf "$BUNDLE" -C /
+  # 项目目录 = 第一条目里到 ocr_data 那一段的绝对路径(取含项目名的目录)
+  PROJDIR="/$(printf '%s' "$FIRST" | sed -E 's#(.*/[^/]*ocr[_-]?data)/.*#\1#')"
+  warn "venv 模式：必须解压到相同路径/同用户，否则 venv 失效"
+else
+  # ---------- 自包含模式：解压到目标目录 ----------
+  log "[2/3] 自包含模式 → 解压到 $TARGET"
+  mkdir -p "$TARGET"
+  tar xzf "$BUNDLE" -C "$TARGET"
+  TOP="$(printf '%s' "$FIRST" | cut -d/ -f1)"
+  PROJDIR="$TARGET/$TOP"
+fi
+echo "  项目目录: $PROJDIR"
+
+log "[3/3] 离线自检"
+cd "$PROJDIR" || { echo "进项目目录失败: $PROJDIR"; exit 1; }
+# shellcheck disable=SC1091
+source activate.sh
+
+ok=1
+ck() { if eval "$2" >/dev/null 2>&1; then echo "  ✓ $1"; else echo "  ✗ $1"; ok=0; fi; }
+ck "python"                "python --version"
+ck "import cv2(headless)"  "python -c 'import cv2'"
+ck "import numpy"          "python -c 'import numpy'"
+ck "import faster_whisper" "python -c 'import faster_whisper'"
+ck "import qcloud_cos"     "python -c 'import qcloud_cos'"
+ck "ffmpeg(bin/)"          "ffmpeg -version"
+ck "ffmpeg 支持 AV1"        "ffmpeg -hide_banner -decoders | grep -iE 'dav1d|av1'"
+ck "coscli"                "./coscli-linux-amd64 --help"
+MODEL="$(python -c "import json;print(json.load(open('config/config.json'))['audio'].get('model_size','small'))" 2>/dev/null || echo small)"
+ck "whisper 模型离线加载"   "python -c \"from faster_whisper import WhisperModel; WhisperModel('$MODEL',device='cpu',compute_type='int8')\""
+
+echo
+if [ "$ok" -eq 1 ]; then
+  log "全部就绪 ✅ 可离线运行"
+  echo "  cd $PROJDIR && source activate.sh"
+  echo "  python main.py --source-list ./text2 --max-workers 12"
+else
+  warn "有项目未通过(见上面 ✗)。常见原因：架构/glibc 不匹配；venv 模式下路径/用户不一致"
+fi
--- a/scripts/diag.sh
+++ b/scripts/diag.sh
@@ -0,0 +1,152 @@
+#!/usr/bin/env bash
+# 一键诊断: 收集机器/进程/内存/OOM/日志/数据 全部信息到一个文件, 发回给我分析。
+# 用法(在项目根目录):  bash scripts/diag.sh
+# 产物:  diag_YYYYmmdd_HHMMSS.txt
+
+cd "$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)" || exit 1
+OUT="diag_$(date +%Y%m%d_%H%M%S).txt"
+PY="$(command -v python || command -v python3)"
+
+# 所有输出同时进文件
+exec > >(tee "$OUT") 2>&1
+
+echo "================ 诊断报告 $(date) ================"
+echo "主机: $(hostname)  目录: $(pwd)"
+echo ""
+
+echo "========== 1. 机器资源 =========="
+echo "[内存 free -g]"; free -g
+echo ""
+echo "[CPU核数] $(nproc)"
+echo "[负载 uptime] $(uptime)"
+echo ""
+
+echo "========== 2. OOM 检查(夜里有没有被系统杀) =========="
+echo "[dmesg killed process]"
+( dmesg -T 2>/dev/null || dmesg 2>/dev/null ) | grep -iE "killed process|out of memory|oom-kill" | tail -20 || echo "  (无 / dmesg无权限)"
+echo "[/var/log/messages oom]"
+grep -iE "killed process|out of memory|oom" /var/log/messages 2>/dev/null | tail -10 || echo "  (无 / 无权限)"
+echo ""
+
+echo "========== 3. 进程状态 =========="
+echo "[main.py 进程数] $(ps -eo args 2>/dev/null | grep -c '[m]ain.py')"
+echo "[ffmpeg 进程数]  $(ps -eo comm 2>/dev/null | grep -c '[f]fmpeg')"
+echo "[forkserver/worker] $(ps -eo args 2>/dev/null | grep -c '[m]ultiprocessing\|[f]orkserver')"
+echo ""
+echo "[吃内存最多的10个进程 RSS]"
+ps -eo pid,rss,etime,args --sort=-rss 2>/dev/null | head -11
+echo ""
+
+echo "========== 4. Python / 环境 =========="
+echo "[python] $($PY --version 2>&1)  路径: $PY"
+echo "[which python] $(command -v python 2>/dev/null)"
+echo "[自包含python在不在] $(ls python/bin/python3 2>/dev/null && echo yes || echo no)"
+echo ""
+
+echo "========== 5. config 关键值 =========="
+if [ -f config/config.json ]; then
+  $PY - <<'PYEOF'
+import json
+try:
+    d=json.load(open("config/config.json",encoding="utf-8"))
+    s=d.get("subtitle",{}); sf=d.get("splitframe",{}); au=d.get("audio",{})
+    print("  ocr_workers=%s no_composite=%s merge_workers=%s max_tiles=%s"%(
+        s.get("ocr_workers"),s.get("no_composite"),s.get("merge_workers"),s.get("max_tiles_per_composite")))
+    print("  cv_detect=%s filter_static=%s filter_watermark_pos=%s split_far_boxes=%s"%(
+        s.get("cv_detect"),s.get("filter_static"),s.get("filter_watermark_pos"),s.get("split_far_boxes")))
+    print("  min_duration=%s crop_bottom_ratio=%s ffmpeg_threads=%s use_gpu=%s"%(
+        s.get("min_duration_seconds"),sf.get("crop_bottom_ratio"),sf.get("ffmpeg_threads"),sf.get("use_gpu")))
+    print("  audio.model_size=%s cpu_threads=%s endpoints数=%s"%(
+        au.get("model_size"),au.get("cpu_threads"),len(d.get("tencentcloud",{}).get("ocr",{}).get("endpoints",[]))))
+except Exception as e:
+    print("  读config失败:",e)
+PYEOF
+else
+  echo "  config/config.json 不存在"
+fi
+echo "[启动命令] $(ps -eo args 2>/dev/null | grep '[m]ain.py' | head -1)"
+echo ""
+
+echo "========== 6. app.log 分析 =========="
+if [ -f app.log ]; then
+  echo "[总行数] $(wc -l < app.log)"
+  echo "[关键词计数]"
+  for kw in 成功 无字幕 失败 暂停启动 抽帧完成 拼接完成 字幕完成 内存不足 BrokenProcess 崩溃 worker崩 Traceback Error 超时 timeout 冷却 端点; do
+    c=$(grep -c "$kw" app.log 2>/dev/null)
+    [ "${c:-0}" -gt 0 ] && printf "   %-14s %s\n" "$kw" "$c"
+  done
+  echo ""
+  echo "[最近 40 行]"
+  tail -40 app.log
+else
+  echo "  app.log 不存在"
+fi
+echo ""
+
+echo "========== 7. 数据.jsonl 分析 =========="
+if [ -f 数据.jsonl ]; then
+  $PY - <<'PYEOF'
+import json,statistics
+from collections import Counter,defaultdict
+from datetime import datetime,timedelta
+latest={}
+for line in open("数据.jsonl",encoding="utf-8"):
+    line=line.strip()
+    if not line: continue
+    try: r=json.loads(line)
+    except: continue
+    v=r.get("video_cos_path")
+    if v: latest[v]=r
+n=len(latest)
+st=Counter(r.get("status") for r in latest.values())
+print("  去重视频 %d | 状态 %s"%(n,dict(st)))
+if n: print("  成功率 %.1f%%"%(st.get("success",0)/n*100))
+# 阶段耗时(成功)
+tot=defaultdict(float); te=0
+for r in latest.values():
+    if r.get("status")!="success": continue
+    sd=r.get("stage_durations") or {}; te+=r.get("elapsed_seconds") or 0
+    for k in ["split_frame","merge","getsrt"]:
+        if k in sd: tot[k]+=sd[k]
+if te:
+    print("  阶段占比: 抽帧%.0f%% 拼接%.0f%% OCR+字幕%.0f%%"%(tot["split_frame"]/te*100,tot["merge"]/te*100,tot["getsrt"]/te*100))
+# OCR/帧
+rs=[ic.get("total_ocr_count")/ic.get("frame_count") for r in latest.values() if (ic:=r.get("image_counts") or {}).get("frame_count") and ic.get("total_ocr_count")]
+if rs: print("  OCR/帧 中位 %.3f"%statistics.median(rs))
+# 时长
+durs=sorted((r.get("duration") or 0) for r in latest.values() if r.get("duration"))
+if durs: print("  视频时长 中位%.0fs max%.0fs | >20分钟 %d个"%(statistics.median(durs),max(durs),sum(1 for d in durs if d>1200)))
+# 单视频平均耗时
+els=[r.get("elapsed_seconds") or 0 for r in latest.values() if r.get("status")=="success"]
+if els: print("  成功视频平均耗时 %.0fs 中位%.0fs"%(sum(els)/len(els),statistics.median(els)))
+# 每小时吞吐 + 有效并发(重叠法)
+busy=defaultdict(float); cnt=defaultdict(int)
+for r in latest.values():
+    try: s=datetime.fromisoformat(r["process_start"]); e=datetime.fromisoformat(r["process_end"])
+    except: continue
+    if e<=s: continue
+    cnt[s.strftime("%m-%d %H")]+=1
+    cur=s
+    while cur<e:
+        hk=cur.strftime("%m-%d %H"); nx=cur.replace(minute=0,second=0,microsecond=0)+timedelta(hours=1)
+        seg=min(e,nx); busy[hk]+=(seg-cur).total_seconds(); cur=seg
+print("  每小时(完成数 / 有效并发):")
+for hk in sorted(busy)[-8:]:
+    print("    %s : %d条 / 并发%.1f"%(hk,cnt[hk],busy[hk]/3600))
+PYEOF
+else
+  echo "  数据.jsonl 不存在"
+fi
+echo ""
+
+echo "========== 8. 磁盘 / GPU =========="
+echo "[磁盘 df -h .]"; df -h . 2>/dev/null
+echo "[split/ 大小] $(du -sh split 2>/dev/null || echo '无')"
+echo "[正在处理的视频数 split/] $(ls split 2>/dev/null | wc -l)"
+if command -v nvidia-smi >/dev/null 2>&1; then
+  echo "[GPU]"; nvidia-smi --query-gpu=index,memory.used,memory.total,utilization.gpu --format=csv,noheader 2>/dev/null
+fi
+echo ""
+echo "================ 诊断结束 ================"
+echo ""
+echo ">>> 报告已保存到: $OUT  (把这个文件发回来)"
--- a/scripts/make_portable.sh
+++ b/scripts/make_portable.sh
@@ -0,0 +1,127 @@
+#!/usr/bin/env bash
+# =============================================================================
+# 把环境改造成"项目内自包含"——Python、第三方库、模型全放进项目目录，
+# 之后整个项目文件夹可以【随便放到任意路径、任意用户】，不再依赖 /root 或固定路径。
+#
+# 在【能联网】的机器上跑一次（已 setup_server.sh 装好 .venv 后）：
+#   bash make_portable.sh                    # 默认 ffmpeg 用 GPU 版(BtbN，含 cuda/nvdec，也兼容 CPU)
+#                                            #   需运行机 glibc>=2.28；GPU 机器自动硬解、CPU 机器软解
+#   FFMPEG_VARIANT=cpu bash make_portable.sh # ffmpeg 用 CPU 版(johnvansickle，全静态，老 glibc 也能跑)
+# 之后:
+#   bash pack.sh          # 打包(此时只需打项目目录，已自包含)
+#   # 离线机解压到【任意路径】即可:  source activate.sh && python main.py ...
+# =============================================================================
+set -euo pipefail
+
+cd "$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+PROJ="$(pwd)"
+PIP_MIRROR="${PIP_MIRROR:-https://repo.huaweicloud.com/repository/pypi/simple}"
+
+log()  { echo -e "\n\033[1;32m==> $*\033[0m"; }
+warn() { echo -e "\033[1;33m[警告] $*\033[0m"; }
+have() { command -v "$1" >/dev/null 2>&1; }
+
+[ -d .venv ] || { echo "没有 .venv，请先 bash setup_server.sh"; exit 1; }
+# shellcheck disable=SC1091
+source .venv/bin/activate
+have uv || { echo "需要 uv（本机改造用，离线机不需要）"; exit 1; }
+
+# 0) 确保 bin/ffmpeg 可用。两种变体（用环境变量 FFMPEG_VARIANT 选）：
+#    cpu (默认) = johnvansickle 静态版，含 dav1d(CPU 软解 AV1)，兼容老 glibc，哪都能跑。
+#    gpu        = BtbN 版，含 cuda/nvdec(GPU 硬解)，给 Ampere+ 卡的机器用；需较新 glibc(>=2.28)。
+#    用法:  FFMPEG_VARIANT=gpu bash make_portable.sh
+# 默认下 GPU 版(含 cuda/nvdec，也兼容 CPU 软解)。要纯 CPU 版用 FFMPEG_VARIANT=cpu。
+FFMPEG_VARIANT="${FFMPEG_VARIANT:-gpu}"
+FF_NEED=1
+if [ "$FFMPEG_VARIANT" = "gpu" ]; then
+  FF_URL="${FFMPEG_GPU_URL:-https://github.com/BtbN/FFmpeg-Builds/releases/download/latest/ffmpeg-master-latest-linux64-gpl.tar.xz}"
+  # GPU 版(BtbN)要 glibc>=2.28 才能运行；本机若太老仅警告(可能在给别的新机器打包)
+  # 注意：管道里 head 会让前面命令 SIGPIPE，配合 set -e/pipefail 会误退出，故加 || true
+  _glibc="$(ldd --version 2>/dev/null | grep -oE '[0-9]+\.[0-9]+' | head -1 || true)"
+  if [ -n "$_glibc" ] && [ "$(printf '%s\n2.28\n' "$_glibc" | sort -V | head -1)" != "2.28" ]; then
+    warn "本机 glibc=$_glibc < 2.28，GPU 版 ffmpeg 在本机跑不了(GLIBC_2.28 not found)。"
+    warn "仅当打包给 glibc>=2.28 的机器时才用 GPU 版；本机自用请: FFMPEG_VARIANT=cpu bash make_portable.sh"
+  fi
+  # 用 grep(不带 -q)读完整输出再判，避免 grep -q 提前关管道触发 SIGPIPE+pipefail 误判
+  if [ -x "$PROJ/bin/ffmpeg" ] && "$PROJ/bin/ffmpeg" -hide_banner -hwaccels 2>/dev/null | grep -i cuda >/dev/null; then
+    log "[0/4] bin/ffmpeg 已是 GPU 版(支持 cuda)，跳过"; FF_NEED=0
+  else
+    log "[0/4] 下载 GPU 版 ffmpeg (BtbN，含 cuda/nvdec) → ./bin"
+  fi
+else
+  FF_URL="${FFMPEG_STATIC_URL:-https://johnvansickle.com/ffmpeg/releases/ffmpeg-release-amd64-static.tar.xz}"
+  if [ -x "$PROJ/bin/ffmpeg" ] && "$PROJ/bin/ffmpeg" -hide_banner -decoders 2>/dev/null | grep -iE 'dav1d|av1' >/dev/null; then
+    log "[0/4] bin/ffmpeg 已支持 AV1(CPU 版)，跳过"; FF_NEED=0
+  else
+    log "[0/4] 下载 CPU 版 ffmpeg (johnvansickle，含 dav1d) → ./bin"
+  fi
+fi
+if [ "$FF_NEED" -eq 1 ]; then
+  mkdir -p "$PROJ/bin"; tmp="$(mktemp -d)"
+  if have curl; then curl -L "$FF_URL" -o "$tmp/ff.tar.xz"; else wget -O "$tmp/ff.tar.xz" "$FF_URL"; fi
+  tar xf "$tmp/ff.tar.xz" -C "$tmp"
+  # johnvansickle 在 <dir>/ffmpeg，BtbN 在 <dir>/bin/ffmpeg —— 直接全树找，两种都兼容
+  ffb="$(find "$tmp" -type f -name ffmpeg | head -1)"
+  ffp="$(find "$tmp" -type f -name ffprobe | head -1)"
+  if [ -n "$ffb" ] && [ -n "$ffp" ]; then
+    cp "$ffb" "$ffp" "$PROJ/bin/" && chmod +x "$PROJ/bin/ffmpeg" "$PROJ/bin/ffprobe"
+    log "ffmpeg/ffprobe 已装到 ./bin（$FFMPEG_VARIANT 版）"
+  else
+    echo "解压后没找到 ffmpeg/ffprobe，请手动放入 ./bin"; exit 1
+  fi
+  rm -rf "$tmp"
+fi
+
+# 1) 把 uv 的独立 Python 复制进项目 ./python（python-build-standalone 本身可重定位）
+log "[1/4] 复制独立 Python 到项目内 ./python"
+PYSRC="$(dirname "$(dirname "$(readlink -f .venv/bin/python)")")"
+echo "  源: $PYSRC"
+rm -rf "$PROJ/python"
+cp -a "$PYSRC" "$PROJ/python"
+PYBIN="$PROJ/python/bin/python3"
+[ -x "$PYBIN" ] || PYBIN="$(ls "$PROJ"/python/bin/python3.* | head -1)"
+echo "  项目内 Python: $PYBIN"
+# 去掉 uv 的"外部管理(EXTERNALLY-MANAGED)"标记，否则 uv 拒绝往这个 Python 里装包
+rm -f "$PROJ"/python/lib/python3.*/EXTERNALLY-MANAGED
+
+# 2) 把依赖装进这个项目内 Python（脱离 .venv，用 headless opencv）
+log "[2/4] 安装依赖到项目内 Python"
+uv pip uninstall --python "$PYBIN" --break-system-packages opencv-python 2>/dev/null || true
+uv pip install --python "$PYBIN" --break-system-packages -r requirements.txt --index-url "$PIP_MIRROR" \
+  --only-binary numpy,scipy,scikit-learn,opencv-python-headless,Pillow,av \
+  || uv pip install --python "$PYBIN" --break-system-packages -r requirements.txt --index-url "$PIP_MIRROR"
+
+# WHISPER_GPU=1：把 CUDA 库(cublas/cudnn)也装进项目 Python，这样目标机【无需系统 CUDA】
+# 也能让 whisper 跑 GPU(ctranslate2 从 venv 里找库；activate.sh 会把它加进 LD_LIBRARY_PATH)。
+if [ "${WHISPER_GPU:-0}" = "1" ]; then
+  log "  额外装 CUDA 库(nvidia-cublas/cudnn-cu12)给 whisper GPU 用"
+  uv pip install --python "$PYBIN" --break-system-packages \
+    nvidia-cublas-cu12 nvidia-cudnn-cu12 --index-url "$PIP_MIRROR" \
+    || warn "CUDA 库安装失败；目标机若无系统 CUDA，whisper 将回退 CPU"
+fi
+
+# 3) 把 whisper 模型缓存搬进项目 ./models（用 HF_HOME 指过去）
+log "[3/4] 预下载 whisper 模型到项目内 ./models"
+MODEL="$(python -c "import json;print(json.load(open('config/config.json'))['audio'].get('model_size','small'))")"
+mkdir -p "$PROJ/models"
+HF_HOME="$PROJ/models" HF_HUB_OFFLINE=0 TRANSFORMERS_OFFLINE=0 \
+  HF_ENDPOINT="${HF_ENDPOINT:-https://hf-mirror.com}" \
+  HF_HUB_DISABLE_XET=1 \
+  "$PYBIN" -c "from faster_whisper import WhisperModel; WhisperModel('$MODEL',device='cpu',compute_type='int8'); print('模型已放进 ./models')"
+
+# 4) 删掉旧 .venv（不再需要，已被项目内 Python 取代）
+log "[4/4] 清理旧 .venv"
+deactivate 2>/dev/null || true
+rm -rf "$PROJ/.venv"
+
+cat <<EOF
+
+完成 ✅  现在项目已自包含，可随便放任意路径/用户：
+  ./python/   独立 Python + 所有第三方库
+  ./bin/      ffmpeg / ffprobe
+  ./models/   whisper 模型
+  coscli-linux-amd64
+
+验证:  source activate.sh && python -c "import cv2,faster_whisper;print('OK')"
+打包:  bash pack.sh
+EOF
--- a/scripts/monitor.sh
+++ b/scripts/monitor.sh
@@ -0,0 +1,33 @@
+#!/usr/bin/env bash
+# 每隔 N 秒采集 CPU/IO/负载/网络,带时间戳写入一个文件。
+# 用法:  bash monitor.sh [间隔秒,默认5] [输出文件,默认 monitor.log]
+# 停止:  Ctrl-C  或  kill $(cat monitor.pid)
+INTERVAL="${1:-5}"
+OUT="${2:-monitor.log}"
+echo $$ > monitor.pid
+
+echo "===== 监控开始 $(date) 间隔${INTERVAL}s 输出 $OUT =====" | tee -a "$OUT"
+command -v iostat >/dev/null 2>&1 && HAS_IOSTAT=1 || HAS_IOSTAT=0
+command -v sar   >/dev/null 2>&1 && HAS_SAR=1   || HAS_SAR=0
+[ "$HAS_IOSTAT" = 0 ] && echo "(无 iostat, 只记 vmstat; 装: yum/apt install sysstat)" | tee -a "$OUT"
+
+trap 'echo "===== 监控停止 $(date) =====" >> "$OUT"; rm -f monitor.pid; exit 0' INT TERM
+
+while true; do
+  {
+    echo "----- $(date '+%Y-%m-%d %H:%M:%S') -----"
+    echo "[vmstat] r=运行队列 b=阻塞(等IO) wa=iowait% id=空闲%"
+    vmstat 1 2 | tail -1
+    if [ "$HAS_IOSTAT" = 1 ]; then
+      echo "[iostat] %util>90=磁盘吃满  w/s=写次数  w_await=写延迟ms"
+      # 取第二份报告(间隔值, 非开机累计), 只留设备行
+      iostat -dx 1 2 | awk '/Device/{c++} c==2'
+    fi
+    if [ "$HAS_SAR" = 1 ]; then
+      echo "[net] 网卡收发 kB/s:"
+      sar -n DEV 1 1 | awk '/Average:/ && $2!="lo" && $2!="IFACE" {printf "   %s rx=%s tx=%s kB/s\n",$2,$5,$6}'
+    fi
+    echo ""
+  } >> "$OUT" 2>&1
+  sleep "$INTERVAL"
+done
--- a/scripts/pack.sh
+++ b/scripts/pack.sh
@@ -0,0 +1,84 @@
+#!/usr/bin/env bash
+# =============================================================================
+# 离线部署打包脚本（在【能联网】的机器上跑）
+#
+# 两种模式自动识别：
+#  A) 自包含模式（已跑过 make_portable.sh，项目内有 ./python + ./models）
+#     → 打包里只含项目这一个文件夹，离线机解压到【任意路径/任意用户】即可。
+#  B) venv 模式（只有 .venv）
+#     → 打项目 + /root 下的 uv Python + HF 缓存，离线机要解压到相同路径。
+#
+# 用法:  bash pack.sh
+# 产物:  /tmp/ocr_offline_bundle.tar.gz
+# =============================================================================
+set -euo pipefail
+
+cd "$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+PROJECT_DIR="$(pwd)"
+OUT="${1:-/tmp/ocr_offline_bundle.tar.gz}"
+
+log()  { echo -e "\n\033[1;32m==> $*\033[0m"; }
+warn() { echo -e "\033[1;33m[警告] $*\033[0m"; }
+have() { command -v "$1" >/dev/null 2>&1; }
+
+PBASE="$(basename "$PROJECT_DIR")"
+PPARENT="$(dirname "$PROJECT_DIR")"
+
+# 运行时数据排除（部署不需要，打进去又慢又大）。用项目目录名作前缀，两种模式通用。
+RT_DIRS=(split video composite coscli_output srt info info_error log)
+EXCLUDES=(--exclude='*.log' --exclude='__pycache__' --exclude='*.pyc'
+          --exclude="$PBASE/数据.jsonl")
+for d in "${RT_DIRS[@]}"; do EXCLUDES+=(--exclude="$PBASE/$d"); done
+
+# 选最快的压缩器：pigz 多线程(输出仍是标准 gzip，离线机普通 tar xzf 就能解)
+if have pigz; then
+  COMPRESS="pigz"; log "用 pigz 多线程压缩(快)"
+else
+  COMPRESS="gzip -1"; warn "未装 pigz，用 gzip -1；想更快可先 yum install -y pigz"
+fi
+
+if [ -x "$PROJECT_DIR/python/bin/python3" ]; then
+  # ---------- A) 自包含：只打项目目录(用相对名)，可解压到任意路径 ----------
+  log "自包含模式（项目内含 ./python + ./models）"
+  [ -d "$PROJECT_DIR/models" ] || warn "没看到 ./models，确认 make_portable.sh 跑完整了"
+  tar "${EXCLUDES[@]}" --use-compress-program="$COMPRESS" -cf "$OUT" \
+    -C "$PPARENT" "$PBASE"
+  MODE="自包含（解压到任意路径/用户）"
+else
+  # ---------- B) venv：打项目(绝对路径) + 外部 Python + HF 缓存 ----------
+  [ -d .venv ] || { echo "既没有 ./python 也没有 .venv，请先 setup_server.sh 或 make_portable.sh"; exit 1; }
+  # shellcheck disable=SC1091
+  source .venv/bin/activate
+
+  log "venv 模式：确保 headless opencv + 预下载模型"
+  uv pip uninstall opencv-python 2>/dev/null || true
+  uv pip install opencv-python-headless \
+    --index-url "${PIP_MIRROR:-https://repo.huaweicloud.com/repository/pypi/simple}" 2>/dev/null \
+    || echo "（headless 可能已装，忽略）"
+  MODEL="$(python -c "import json;print(json.load(open('config/config.json'))['audio'].get('model_size','small'))")"
+  HF_HUB_OFFLINE=0 TRANSFORMERS_OFFLINE=0 HF_ENDPOINT="${HF_ENDPOINT:-https://hf-mirror.com}" \
+    python -c "from faster_whisper import WhisperModel; WhisperModel('$MODEL',device='cpu',compute_type='int8'); print('模型已缓存')"
+
+  PYDIR="$(dirname "$(dirname "$(readlink -f .venv/bin/python)")")"
+  HFCACHE="${HF_HOME:-$HOME/.cache/huggingface}"
+  echo "  uv Python: $PYDIR    模型缓存: $HFCACHE"
+
+  log "打包项目(绝对路径) + Python + 模型缓存"
+  # 项目用绝对路径(还原到原位)，外部两目录同理
+  EXABS=(--exclude='*.log' --exclude='__pycache__' --exclude='*.pyc'
+         --exclude="${PROJECT_DIR#/}/数据.jsonl")
+  for d in "${RT_DIRS[@]}"; do EXABS+=(--exclude="${PROJECT_DIR#/}/$d"); done
+  tar "${EXABS[@]}" --use-compress-program="$COMPRESS" -cf "$OUT" \
+    -C / "${PROJECT_DIR#/}" -C / "${PYDIR#/}" -C / "${HFCACHE#/}"
+  MODE="venv（离线机需解压到相同路径/用户）"
+fi
+
+echo
+log "完成 ✅  $OUT  ($(du -h "$OUT" | cut -f1))"
+echo "模式: $MODE"
+cat <<EOF
+
+下一步：
+  scp $OUT root@离线机:/tmp/
+  bash deploy.sh /tmp/$(basename "$OUT")          # 自包含模式可加目标路径: bash deploy.sh 包 /opt
+EOF
--- a/scripts/set_ffmpeg.sh
+++ b/scripts/set_ffmpeg.sh
@@ -0,0 +1,26 @@
+#!/usr/bin/env bash
+# 一键切换项目 bin/ffmpeg 的版本(本机自用 cpu / 打 GPU 包 gpu)
+#   bash temp/set_ffmpeg.sh cpu   # johnvansickle 全静态，任何 glibc 都能跑(本机自用)
+#   bash temp/set_ffmpeg.sh gpu   # BtbN 含 cuda/nvdec，需 glibc>=2.28(给 GPU 机器)
+set -euo pipefail
+cd "$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"   # 项目根
+V="${1:-cpu}"
+BIN="$(pwd)/bin"; mkdir -p "$BIN"
+have(){ command -v "$1" >/dev/null 2>&1; }
+
+case "$V" in
+  cpu) URL="https://johnvansickle.com/ffmpeg/releases/ffmpeg-release-amd64-static.tar.xz" ;;
+  gpu) URL="https://github.com/BtbN/FFmpeg-Builds/releases/download/latest/ffmpeg-master-latest-linux64-gpl.tar.xz" ;;
+  *) echo "用法: bash temp/set_ffmpeg.sh [cpu|gpu]"; exit 1 ;;
+esac
+
+echo "下载 $V 版 ffmpeg ..."
+tmp="$(mktemp -d)"
+if have curl; then curl -L "$URL" -o "$tmp/ff.tar.xz"; else wget -O "$tmp/ff.tar.xz" "$URL"; fi
+tar xf "$tmp/ff.tar.xz" -C "$tmp"
+ffb="$(find "$tmp" -type f -name ffmpeg | head -1)"
+ffp="$(find "$tmp" -type f -name ffprobe | head -1)"
+cp "$ffb" "$ffp" "$BIN/"
+chmod +x "$BIN/ffmpeg" "$BIN/ffprobe"
+rm -rf "$tmp"
+echo "已切到 $V 版: $("$BIN/ffmpeg" -version 2>&1 | head -1 || echo '本机跑不了(若是 gpu 版+老 glibc，正常)')"
--- a/scripts/setup_server.sh
+++ b/scripts/setup_server.sh
@@ -0,0 +1,244 @@
+#!/usr/bin/env bash
+# =============================================================================
+# 视频字幕提取系统 — 服务器一键环境配置脚本
+#
+# 安装/配置：gcc 编译工具链、ffmpeg(含 AV1/dav1d 解码)、uv、Python 虚拟环境、
+#            项目依赖，并处理常见环境坑(opencv libGL、UTF-8 locale、HF/pip 镜像、
+#            OMP 线程、文件句柄上限)。
+#
+# 用法(在项目根目录):
+#   bash setup_server.sh
+# 跑完后:
+#   source activate.sh      # 激活环境(设置 venv + 各种环境变量)
+#   python main.py --source-list ./text2 --max-workers 12
+#
+# 幂等：可重复运行，已装的会跳过。
+# =============================================================================
+set -euo pipefail
+
+# ----------------------------- 可调参数 --------------------------------------
+PYTHON_VERSION="${PYTHON_VERSION:-3.11}"
+PIP_MIRROR="${PIP_MIRROR:-https://repo.huaweicloud.com/repository/pypi/simple}"
+HF_ENDPOINT_URL="${HF_ENDPOINT_URL:-https://hf-mirror.com}"
+FFMPEG_STATIC_URL="${FFMPEG_STATIC_URL:-https://johnvansickle.com/ffmpeg/releases/ffmpeg-release-amd64-static.tar.xz}"
+PROJECT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+VENV_DIR="$PROJECT_DIR/.venv"
+
+cd "$PROJECT_DIR"
+
+# 让本脚本(及自检)优先用项目 bin/ 里的 ffmpeg(我们装的静态版)
+export PATH="$PROJECT_DIR/bin:$PATH"
+
+# ----------------------------- 工具函数 --------------------------------------
+log()  { echo -e "\n\033[1;32m==> $*\033[0m"; }
+warn() { echo -e "\033[1;33m[警告] $*\033[0m"; }
+err()  { echo -e "\033[1;31m[错误] $*\033[0m" >&2; }
+have() { command -v "$1" >/dev/null 2>&1; }
+
+# root 则不用 sudo，否则用 sudo
+if [ "$(id -u)" -eq 0 ]; then SUDO=""; else SUDO="sudo"; fi
+
+# 检测包管理器
+PKG=""
+if have dnf;     then PKG="dnf"
+elif have yum;   then PKG="yum"
+elif have apt-get; then PKG="apt"
+else warn "未识别的包管理器，跳过系统包安装(请确保已手动装好 gcc/ffmpeg 等)"; fi
+
+# 解压后 coscli/bin 里的二进制常丢失 +x，先无条件加执行权限（放最前，确保一定执行）
+for b in coscli-linux-amd64 coscli bin/ffmpeg bin/ffprobe; do
+  [ -f "$PROJECT_DIR/$b" ] && chmod +x "$PROJECT_DIR/$b" 2>/dev/null || true
+done
+log "已给 coscli / bin 二进制加执行权限"
+
+# =============================================================================
+log "[1/7] 安装系统依赖(gcc 工具链 + opencv/运行库)"
+# =============================================================================
+# 关键环境坑：
+#  - opencv-python 在无图形界面的服务器上 import 会报 libGL.so.1 缺失 → 装 mesa-libGL/libgl1
+#  - 还需要 glib2(libgthread-2.0) 否则 opencv 也可能报错
+#  - gcc/gcc-c++/make：少数包没有 wheel 时要本地编译
+case "$PKG" in
+  dnf|yum)
+    # openblas-devel：万一 scipy/numpy 走源码编译时需要(优先还是用 wheel)
+    $SUDO $PKG install -y gcc gcc-c++ gcc-gfortran make git tar xz which \
+                          mesa-libGL glib2 openblas-devel || warn "部分系统包安装失败，继续"
+    ;;
+  apt)
+    $SUDO apt-get update -y || true
+    $SUDO apt-get install -y build-essential gfortran git tar xz-utils curl \
+                             libgl1 libglib2.0-0 libopenblas-dev || warn "部分系统包安装失败，继续"
+    ;;
+esac
+
+# =============================================================================
+log "[2/7] 安装 ffmpeg(含 AV1/dav1d 解码) → 项目目录 bin/"
+# =============================================================================
+# 直接下载静态构建(johnvansickle，自带 dav1d)到项目的 bin/，代码会自动优先用它
+# (config_utils 启动时把 bin/ 提到 PATH 最前)。不依赖系统 ffmpeg，也不用 sudo。
+# 注意：CPU 版(无 NVDEC)；要 GPU 硬解需另装 cuda 版 ffmpeg。
+FFMPEG_DIR="$PROJECT_DIR/bin"
+LOCAL_FFMPEG="$FFMPEG_DIR/ffmpeg"
+need_ffmpeg=1
+if [ -x "$LOCAL_FFMPEG" ] && "$LOCAL_FFMPEG" -hide_banner -decoders 2>/dev/null | grep -qiE 'dav1d|av1'; then
+  log "项目自带 ffmpeg 已存在且支持 AV1: $LOCAL_FFMPEG，跳过下载"
+  need_ffmpeg=0
+fi
+if [ "$need_ffmpeg" -eq 1 ]; then
+  if have curl || have wget; then
+    mkdir -p "$FFMPEG_DIR"
+    tmp="$(mktemp -d)"
+    log "下载 ffmpeg 静态包到 $FFMPEG_DIR ..."
+    if have curl; then curl -L "$FFMPEG_STATIC_URL" -o "$tmp/ff.tar.xz"
+    else wget -O "$tmp/ff.tar.xz" "$FFMPEG_STATIC_URL"; fi
+    tar -xf "$tmp/ff.tar.xz" -C "$tmp"
+    ffdir="$(find "$tmp" -maxdepth 1 -type d -name 'ffmpeg-*' | head -1)"
+    if [ -n "$ffdir" ]; then
+      cp "$ffdir/ffmpeg" "$ffdir/ffprobe" "$FFMPEG_DIR/"
+      chmod +x "$FFMPEG_DIR/ffmpeg" "$FFMPEG_DIR/ffprobe"
+      log "ffmpeg/ffprobe 已装到 $FFMPEG_DIR"
+    else
+      err "解压后未找到 ffmpeg 目录，请手动安装 ffmpeg"
+    fi
+    rm -rf "$tmp"
+  else
+    err "没有 curl/wget，无法下载 ffmpeg，请手动安装"
+  fi
+fi
+
+# =============================================================================
+log "[3/7] 安装 uv"
+# =============================================================================
+if ! have uv; then
+  curl -LsSf https://astral.sh/uv/install.sh | sh
+  # uv 默认装到 ~/.local/bin 或 ~/.cargo/bin
+  export PATH="$HOME/.local/bin:$HOME/.cargo/bin:$PATH"
+fi
+have uv || { err "uv 安装失败，请检查网络"; exit 1; }
+log "uv 版本: $(uv --version)"
+
+# =============================================================================
+log "[4/7] 用 uv 安装 Python ${PYTHON_VERSION}"
+# =============================================================================
+uv python install "$PYTHON_VERSION" || warn "uv python install 失败(可能已装)"
+
+# =============================================================================
+log "[5/7] 创建虚拟环境并安装依赖"
+# =============================================================================
+# 优先用 pyproject.toml + uv.lock(规范、可复现)；没有则回退 requirements.txt。
+# pyproject 的 [tool.uv] 已配置：禁止 av/numpy/sklearn 本地编译 + 清华镜像。
+# 用 uv venv + uv pip install(现代解析器)。--only-binary 让它自动挑有 wheel 的版本，
+# 避免某些受限镜像把新版本只给源码包导致本地编译/报错(不像 uv sync 的 no-build 会硬失败)。
+[ -d "$VENV_DIR" ] || uv venv --python "$PYTHON_VERSION" "$VENV_DIR"
+# shellcheck disable=SC1091
+source "$VENV_DIR/bin/activate"
+if [ -f requirements.txt ]; then
+  uv pip install -r requirements.txt --index-url "$PIP_MIRROR" \
+    --only-binary numpy,scipy,scikit-learn,opencv-python-headless,Pillow,av \
+  || uv pip install -r requirements.txt --index-url "$PIP_MIRROR"
+else
+  warn "未找到 requirements.txt，跳过依赖安装"
+fi
+
+# 激活 venv 供后续自检使用
+# shellcheck disable=SC1091
+[ -f "$VENV_DIR/bin/activate" ] && source "$VENV_DIR/bin/activate"
+log "Python: $(python --version 2>&1)  位置: $(command -v python)"
+
+# =============================================================================
+log "[6/7] 写出环境变量脚本 activate.sh"
+# =============================================================================
+# 常见环境坑统一在这里处理，跑项目前 source 一下即可
+# 用单引号 heredoc 让脚本体保持字面(不在生成时展开)，实现"自动定位所在目录"，
+# 这样整个项目可以搬到任意路径，activate.sh 仍能找到自己的 .venv / bin。
+cat > "$PROJECT_DIR/activate.sh" <<'EOF'
+#!/usr/bin/env bash
+# 跑项目前先: source activate.sh —— 自动定位项目目录，项目可放任意路径
+HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+
+# Python：优先项目内自包含的 ./python(make_portable.sh 产物，可随便放)；否则用 .venv
+if [ -x "$HERE/python/bin/python3" ]; then
+  export PATH="$HERE/bin:$HERE/python/bin:$PATH"
+else
+  export VIRTUAL_ENV="$HERE/.venv"
+  export PATH="$HERE/bin:$VIRTUAL_ENV/bin:$HOME/.local/bin:$HOME/.cargo/bin:$PATH"
+fi
+
+# UTF-8：项目有中文文件名(数据.jsonl)。优先选系统存在的 UTF-8 locale，避免 setlocale 警告；
+# 并强制 Python 用 UTF-8(即使系统没有 UTF-8 locale 也能正确处理中文文件名)。
+if locale -a 2>/dev/null | grep -qiE '^C\.(utf-?8|UTF-?8)$'; then
+  export LANG=C.UTF-8 LC_ALL=C.UTF-8
+elif locale -a 2>/dev/null | grep -qiE 'en_US\.(utf-?8|UTF-?8)'; then
+  export LANG=en_US.UTF-8 LC_ALL=en_US.UTF-8
+fi
+export PYTHONUTF8=1
+export PYTHONIOENCODING=utf-8
+
+# HuggingFace 镜像（联网下模型时用）
+export HF_ENDPOINT=__HF_ENDPOINT__
+# 模型缓存：项目内有 ./models 就用它(自包含包)，否则用默认 ~/.cache/huggingface
+[ -d "$HERE/models" ] && export HF_HOME="$HERE/models"
+# 只在模型【确实已缓存】时才强制离线，否则允许联网下载(避免"没模型又禁网"的死锁)
+if ls "${HF_HOME:-$HOME/.cache/huggingface}"/hub/models--*faster-whisper* >/dev/null 2>&1; then
+  export HF_HUB_OFFLINE=1
+  export TRANSFORMERS_OFFLINE=1
+fi
+
+# 若项目自带 CUDA 库(WHISPER_GPU=1 打包)，加进 LD_LIBRARY_PATH，让 ctranslate2 找到(whisper GPU)
+_nvlibs="$(ls -d "$HERE"/python/lib/python*/site-packages/nvidia/*/lib 2>/dev/null | tr '\n' ':')"
+[ -n "$_nvlibs" ] && export LD_LIBRARY_PATH="${_nvlibs}${LD_LIBRARY_PATH:-}"
+
+# 限制数值库线程：防止 whisper/numpy/openblas 每个都抢满核(多视频并发时互相踩)
+export OMP_NUM_THREADS=2
+export OPENBLAS_NUM_THREADS=2
+export MKL_NUM_THREADS=2
+
+echo "环境已激活: $(python --version 2>&1) | $HERE"
+EOF
+sed -i "s|__HF_ENDPOINT__|$HF_ENDPOINT_URL|" "$PROJECT_DIR/activate.sh"
+chmod +x "$PROJECT_DIR/activate.sh"
+
+# 提高文件句柄上限(高并发 + 海量小帧文件)，当前会话生效；永久需改 /etc/security/limits.conf
+ulimit -n 65535 2>/dev/null || warn "ulimit 提升失败(非致命)，如需永久生效请改 /etc/security/limits.conf"
+
+
+# =============================================================================
+log "[7/7] 环境自检"
+# =============================================================================
+ok=1
+check() { if eval "$2" >/dev/null 2>&1; then echo "  ✓ $1"; else echo "  ✗ $1"; ok=0; fi; }
+
+check "gcc"            "gcc --version"
+check "ffmpeg"         "ffmpeg -version"
+check "ffprobe"        "ffprobe -version"
+check "ffmpeg 支持 AV1" "ffmpeg -hide_banner -decoders | grep -iE 'dav1d|av1'"
+check "uv"             "uv --version"
+check "python"         "python --version"
+check "import cv2"     "python -c 'import cv2'"
+check "import numpy"   "python -c 'import numpy'"
+check "import sklearn" "python -c 'import sklearn'"
+check "import av(PyAV)" "python -c 'import av'"
+check "import faster_whisper" "python -c 'import faster_whisper'"
+check "import qcloud_cos"     "python -c 'import qcloud_cos'"
+
+# GPU 检测(可选，有则抽帧能硬解加速)
+if have nvidia-smi && nvidia-smi >/dev/null 2>&1; then
+  echo "  ✓ 检测到 NVIDIA GPU(抽帧可硬解加速)"
+else
+  echo "  · 未检测到 GPU(抽帧用 CPU，正常)"
+fi
+
+echo
+if [ "$ok" -eq 1 ]; then
+  log "全部就绪 ✅"
+else
+  warn "有项目未通过(见上面 ✗)，请按提示排查"
+fi
+
+cat <<EOF
+
+下一步：
+  source activate.sh
+  python main.py --source-list ./text2 --max-workers 12
+
+EOF
--- a/scripts/stop.sh
+++ b/scripts/stop.sh
@@ -0,0 +1,52 @@
+#!/usr/bin/env bash
+# 停止本项目所有进程(主程序 + 进程池worker + ffmpeg子进程 + coscli)。
+# 用法:  bash stop.sh           普通停止
+#         bash stop.sh -y        不确认直接停
+set -u
+
+echo "==> 即将停止的进程:"
+ps -eo pid,etime,rss,args 2>/dev/null | grep -E "main\.py|run_worker\.py|multiprocessing|forkserver|[f]fmpeg|[c]oscli" | grep -v grep | grep -v "stop.sh" || echo "  (没找到运行中的进程)"
+echo ""
+
+if [ "${1:-}" != "-y" ]; then
+    read -r -p "确认全部杀掉? [y/N] " ans
+    case "$ans" in
+        y|Y) ;;
+        *) echo "已取消"; exit 0 ;;
+    esac
+fi
+
+echo "==> 1. 杀主程序 + 进程池worker"
+pkill -9 -f "main.py"          2>/dev/null && echo "   ✓ main.py"
+pkill -9 -f "run_worker.py"    2>/dev/null && echo "   ✓ run_worker.py"
+pkill -9 -f "multiprocessing"  2>/dev/null && echo "   ✓ multiprocessing worker"
+pkill -9 -f "forkserver"       2>/dev/null && echo "   ✓ forkserver"
+
+echo "==> 2. 杀 ffmpeg 子进程(不会自动死, 必须单独杀, 否则占GPU/CPU)"
+pkill -9 ffmpeg                2>/dev/null && echo "   ✓ ffmpeg"
+pkill -9 ffprobe               2>/dev/null && echo "   ✓ ffprobe"
+
+echo "==> 3. 杀残留 coscli(正在传输的)"
+pkill -9 -f "coscli"           2>/dev/null && echo "   ✓ coscli"
+
+sleep 1
+echo ""
+echo "==> 检查残留:"
+LEFT=$(ps -eo pid,args 2>/dev/null | grep -E "main\.py|run_worker\.py|multiprocessing|forkserver|[f]fmpeg|[c]oscli" | grep -v grep | grep -v "stop.sh")
+if [ -z "$LEFT" ]; then
+    echo "  ✓ 全部已停止"
+else
+    echo "  ⚠️ 仍有残留, 再杀一次:"
+    echo "$LEFT"
+    echo "$LEFT" | awk '{print $1}' | xargs -r kill -9 2>/dev/null
+    echo "  已强制杀掉"
+fi
+
+# 可选: 看GPU是否释放
+if command -v nvidia-smi >/dev/null 2>&1; then
+    echo ""
+    echo "==> GPU 占用(确认ffmpeg/whisper已释放):"
+    nvidia-smi --query-gpu=index,memory.used,utilization.gpu --format=csv,noheader 2>/dev/null
+fi
+echo ""
+echo "停止完成。重启: bash scripts/启动.sh 30"
--- a/scripts/启动.sh
+++ b/scripts/启动.sh
@@ -0,0 +1,164 @@
+#!/usr/bin/env bash
+# =============================================================================
+# 一键启动(防 SIGHUP/外部kill 脱离终端 + 调好参数 + 启动前自检 + 防重复启动)
+#
+# 两种模式(MODE 环境变量切换):
+#   MODE=worker (默认) —— 常驻 worker, 从 COS 自动认领任务(run_worker.py)。不需要本地 text2。
+#   MODE=local         —— 旧模式, 处理本地 ./text2 下的任务txt(main.py)。
+#
+# 用法:
+#   bash scripts/启动.sh                       # 默认 worker 模式启动
+#   bash scripts/启动.sh 32                     # 临时把 max-workers 改成 32
+#   MODE=local bash scripts/启动.sh             # 用旧的本地 text2 模式
+#   MAX_WORKERS=20 MEM_GUARD=40 bash scripts/启动.sh
+#
+# 停止:  bash scripts/stop.sh -y
+# =============================================================================
+set -u
+cd "$(cd "$(dirname "$0")/.." && pwd)"
+ROOT="$(pwd)"
+
+# ---- 可调参数(命令行第1个参数 / 环境变量 均可覆盖) ----
+# MODE 未显式指定时, 读 config.dispatch.source 决定: local -> 本地text2, 其余 -> COS常驻
+if [ -z "${MODE:-}" ]; then
+  _src="$(python -c "from config.config_utils import config; print((config.get_dispatch_config() or {}).get('source','cos'))" 2>/dev/null || echo cos)"
+  [ "$_src" = "local" ] && MODE="local" || MODE="worker"
+fi
+MODE="${MODE:-worker}"                      # worker=COS认领常驻 / local=本地text2
+MAX_WORKERS="${1:-${MAX_WORKERS:-25}}"      # 0=不传给python,由main.py/run_worker.py自动按本机CPU核数选
+                                            # 共享机/跟OCR服务同机部署的, 建议显式传小一点的值(如24), 别用自动
+MAX_TASKS="${MAX_TASKS:-0}"                # 0=不回收worker(避免3.11回收缩水把并发缩到0)
+MEM_GUARD="${MEM_GUARD:-30}"               # 可用内存低于此(G)暂停启动新视频
+SOURCE_LIST="${SOURCE_LIST:-./text2}"      # 仅 local 模式: 任务txt目录
+EXECUTOR="${EXECUTOR:-process}"            # process=进程池(绕GIL) / thread
+LOG="${LOG:-app.log}"
+NICE="${NICE:-15}"                         # 处理进程CPU优先级(0~19,越大越低)
+LOG_MAX_MB="${LOG_MAX_MB:-200}"            # $LOG 超过这个大小(MB)就在启动前轮转一次, 0=不轮转
+
+# ---- 启动前轮转一次日志(常驻进程一跑数天/数周, >>追加写不会自己变小) ----
+_rotate_log() {
+  local f="$1" maxmb="$2"
+  [ "$maxmb" -le 0 ] && return 0
+  [ -f "$f" ] || return 0
+  local sz_mb=$(( $(stat -c%s "$f" 2>/dev/null || echo 0) / 1024 / 1024 ))
+  if [ "$sz_mb" -ge "$maxmb" ]; then
+    local ts; ts="$(date +%Y%m%d_%H%M%S)"
+    mv "$f" "${f}.${ts}" 2>/dev/null && echo "[日志轮转] $f (${sz_mb}M) → ${f}.${ts}, 重新开始写"
+    ls -t "${f}".[0-9]* 2>/dev/null | tail -n +6 | xargs -r rm -f   # 只留最近5份轮转备份
+  fi
+}
+_rotate_log "$LOG" "$LOG_MAX_MB"
+
+# 按模式确定入口与进程匹配模式
+if [ "$MODE" = "worker" ]; then
+  ENTRY="run_worker.py"
+  PROC_PAT='python .*run_worker\.py'
+else
+  ENTRY="main.py"
+  PROC_PAT='python .*main\.py'
+fi
+
+MAX_WORKERS_DISP="$MAX_WORKERS"
+[ "$MAX_WORKERS" = "0" ] && MAX_WORKERS_DISP="0(自动按CPU核数选, 实际值见下面python启动日志)"
+
+echo "=================================================="
+echo " 启动 $ENTRY  (MODE=$MODE)"
+echo "   max-workers=$MAX_WORKERS_DISP  max-tasks-per-child=$MAX_TASKS  mem-guard=${MEM_GUARD}G"
+echo "   executor=$EXECUTOR  nice=$NICE(让OCR服务优先)  日志→$LOG"
+[ "$MODE" = "local" ] && echo "   source-list=$SOURCE_LIST"
+echo "=================================================="
+
+# ---- 1. 防重复启动 ----
+if pgrep -af "$PROC_PAT" 2>/dev/null | grep -v 'multiprocessing' | grep -v 'grep' >/dev/null; then
+  echo "✗ 已经有 $ENTRY 在运行了, 不重复启动。"
+  echo "  查看: pgrep -af '$PROC_PAT' | grep -v multiprocessing"
+  echo "  如要重启: 先 bash scripts/stop.sh -y"
+  exit 1
+fi
+
+# ---- 2. 启动前自检 ----
+fail=0
+need() { command -v "$1" >/dev/null 2>&1 && echo "  ✓ $1" || { echo "  ✗ 缺少 $1 (需在 PATH 或项目目录)"; fail=1; }; }
+echo "[依赖检查]"
+need python
+need ffmpeg
+need ffprobe
+if command -v coscli >/dev/null 2>&1; then echo "  ✓ coscli"
+elif [ -x "./coscli-linux-amd64" ]; then echo "  ✓ ./coscli-linux-amd64"
+else echo "  ✗ 找不到 coscli / ./coscli-linux-amd64"; fail=1; fi
+
+echo "[配置/任务检查]"
+[ -f config/config.json ] && echo "  ✓ config/config.json" || { echo "  ✗ 缺 config/config.json"; fail=1; }
+
+if [ "$MODE" = "worker" ]; then
+  # worker 模式: 任务来自 COS, 不需要本地 text2; 检查 dispatch 配置与机器ID + COS 队列可达
+  chk=$(python - <<'PY' 2>&1
+import sys
+try:
+    from config.config_utils import config
+    from tool import job_cos
+    dp = config.get_dispatch_config() or {}
+    if not dp.get("jobs_root"):
+        print("NO_JOBS_ROOT"); sys.exit(0)
+    mid = job_cos.get_machine_id()
+    todo = job_cos.list_todo()       # 顺带验证 COS 队列可达
+    print("OK\t%s\t%d" % (mid, len(todo)))
+except Exception as e:
+    print("ERR\t%s" % e)
+PY
+)
+  case "$chk" in
+    OK*)   mid=$(echo "$chk" | cut -f2); n=$(echo "$chk" | cut -f3)
+           echo "  ✓ dispatch 配置OK  机器ID=$mid  todo剩余分片=$n"
+           [ "$n" = "0" ] && echo "  ⚠ todo 队列为空(worker会空闲等待; 先 python tools/seed_jobs.py 灌任务)";;
+    NO_JOBS_ROOT) echo "  ✗ config.dispatch.jobs_root 未配置(取任务地址)"; fail=1;;
+    *)     echo "  ✗ 访问 COS 任务队列失败: ${chk#ERR	}"; fail=1;;
+  esac
+else
+  if ls "$SOURCE_LIST"/*.txt >/dev/null 2>&1; then
+    n=$(ls "$SOURCE_LIST"/*.txt 2>/dev/null | wc -l)
+    echo "  ✓ $SOURCE_LIST 有 $n 个任务txt"
+  else
+    echo "  ✗ $SOURCE_LIST 下没有 .txt 任务文件"; fail=1
+  fi
+fi
+
+if [ "$fail" -ne 0 ]; then
+  echo ""
+  echo "✗ 自检未通过, 已中止(没启动)。修好上面 ✗ 的项再跑。"
+  exit 1
+fi
+
+# ---- 3. 启动(setsid+nohup+disown 三保险, SSH断开也杀不到) ----
+echo ""
+echo "[启动中...]"
+if [ "$MODE" = "worker" ]; then
+  setsid nohup nice -n "$NICE" python run_worker.py \
+    --max-workers "$MAX_WORKERS" \
+    --max-tasks-per-child "$MAX_TASKS" \
+    --mem-guard-gb "$MEM_GUARD" \
+    --executor "$EXECUTOR" >> "$LOG" 2>&1 &
+else
+  setsid nohup nice -n "$NICE" python main.py \
+    --source-list "$SOURCE_LIST" \
+    --max-workers "$MAX_WORKERS" \
+    --max-tasks-per-child "$MAX_TASKS" \
+    --mem-guard-gb "$MEM_GUARD" \
+    --executor "$EXECUTOR" >> "$LOG" 2>&1 &
+fi
+disown 2>/dev/null || true
+
+# ---- 4. 等几秒确认起来了 ----
+sleep 6
+if pgrep -af "$PROC_PAT" 2>/dev/null | grep -v 'multiprocessing' | grep -v 'grep' >/dev/null; then
+  echo "✓ 已启动。日志末尾:"
+  tail -n 5 "$LOG" 2>/dev/null | sed 's/^/    /'
+  echo ""
+  echo "  实时看日志:  tail -f $LOG"
+  echo "  数worker:    pgrep -af multiprocessing | wc -l   (稍等会接近 $MAX_WORKERS_DISP)"
+  echo "  停止:        bash scripts/stop.sh -y"
+else
+  echo "✗ 启动后没检测到进程, 查日志看原因:"
+  tail -n 20 "$LOG" 2>/dev/null | sed 's/^/    /'
+  exit 1
+fi
--- a/scripts/守护重启.sh
+++ b/scripts/守护重启.sh
@@ -0,0 +1,121 @@
+#!/usr/bin/env bash
+# =============================================================================
+# 守护重启: main.py 被外部 kill / 异常退出后, 自动拉起(共享机上挡不住别人kill, 就让它自己回来)。
+# 【带明确开关, 想关随时关, 不会和你抢】:
+#   关闭原理 = 放一个"暂停标记文件" .guard_paused, 守护看到它就什么都不做。
+#
+# 用法:
+#   bash scripts/守护重启.sh install     # 安装(每分钟检测一次, 没在跑就启动)
+#   bash scripts/守护重启.sh off         # ★想停就用这个★ 暂停自动重启 + 停掉当前进程
+#   bash scripts/守护重启.sh on          # 重新开启自动重启(下一分钟内会被拉起)
+#   bash scripts/守护重启.sh status      # 看状态(装没装/开没开/在没在跑)
+#   bash scripts/守护重启.sh uninstall   # 彻底卸载(删掉定时任务, 进程不动)
+#   bash scripts/守护重启.sh tick        # (cron 内部调用, 别手动跑)
+#
+# 启动参数想改, 改下面 START= 这一行即可。
+# =============================================================================
+set -u
+ROOT="$(cd "$(dirname "$0")/.." && pwd)"
+cd "$ROOT"
+
+PAUSE="$ROOT/.guard_paused"          # 这个文件存在 = 暂停自动重启
+LOCK="$ROOT/.guard.lock"
+GLOG="$ROOT/guard.log"
+CRON_TAG="# ocr-guard"               # crontab 里的标识, 用于增删去重
+
+# ★ 启动命令(按之前调优建议; 想改并发/参数改这里) ★
+# MODE=worker(默认): 常驻 worker 从 COS 认领任务; MODE=local: 旧的本地 text2 模式。
+# nice -n 15: 让 co-located 的 OCR 服务优先抢 CPU; max-tasks-per-child 20: 本机内存偏紧
+# (44G可用/184G总量), 定期回收worker释放常驻内存; 3.11回收有缩水风险(见CLAUDE.md), 但20这个
+# 量级配合 --mem-guard-gb 观察下来风险可控, 内存不紧的机器可以改回0(不回收)。
+if [ -z "${MODE:-}" ]; then
+  _src="$(python -c "from config.config_utils import config; print((config.get_dispatch_config() or {}).get('source','cos'))" 2>/dev/null || echo cos)"
+  [ "$_src" = "local" ] && MODE="local" || MODE="worker"
+fi
+MODE="${MODE:-worker}"
+if [ "$MODE" = "worker" ]; then
+  START='nice -n 15 python run_worker.py --max-workers 30 --max-tasks-per-child 20 --mem-guard-gb 30 --executor process'
+  PROC_PAT='python .*run_worker\.py'
+else
+  START='nice -n 15 python main.py --source-list ./text2 --max-workers 30 --max-tasks-per-child 20 --mem-guard-gb 30 --executor process'
+  PROC_PAT='python .*main\.py'
+fi
+
+# 判断父进程是否在跑(forkserver worker 命令行不含入口文件名, 不会误判)
+running() {
+  pgrep -af "$PROC_PAT" 2>/dev/null | grep -v 'multiprocessing' | grep -v 'grep' >/dev/null 2>&1
+}
+
+# app.log 轮转: 进程可能正在往它里面 >> 追加写, 不能直接 mv(会导致进程还在写旧inode, 新文件
+# 看起来永远是空的) —— 用 copytruncate: 先 cp 一份出去, 再原地截断; O_APPEND 的fd是按"当前
+# 文件大小"找写入位置的, 原地截断后进程会自动从offset 0继续写这个(同一个)文件, 不受影响。
+_rotate_applog() {
+  local f="$ROOT/app.log" maxmb=200
+  [ -f "$f" ] || return 0
+  local sz_mb=$(( $(stat -c%s "$f" 2>/dev/null || echo 0) / 1024 / 1024 ))
+  if [ "$sz_mb" -ge "$maxmb" ]; then
+    local ts; ts="$(date +%Y%m%d_%H%M%S)"
+    cp "$f" "${f}.${ts}" 2>/dev/null && : > "$f" \
+      && echo "$(date '+%F %T') [日志轮转] app.log(${sz_mb}M) → app.log.${ts}, 原地清空续写" >> "$GLOG"
+    ls -t "${f}".[0-9]* 2>/dev/null | tail -n +6 | xargs -r rm -f   # 只留最近5份
+  fi
+}
+
+case "${1:-status}" in
+  install)
+    ( crontab -l 2>/dev/null | grep -vF "$CRON_TAG"
+      echo "* * * * * cd $ROOT && bash scripts/守护重启.sh tick $CRON_TAG" ) | crontab -
+    rm -f "$PAUSE"
+    echo "✓ 守护已安装(每分钟检测), 自动重启: 开启"
+    echo "  想停: bash scripts/守护重启.sh off"
+    ;;
+
+  uninstall)
+    crontab -l 2>/dev/null | grep -vF "$CRON_TAG" | crontab -
+    echo "✓ 已卸载守护定时任务(不再自动重启)。当前进程未动。"
+    echo "  如要同时停进程: bash scripts/stop.sh -y"
+    ;;
+
+  off)
+    touch "$PAUSE"
+    echo "✓ 已暂停自动重启(标记 $PAUSE)。正在停止当前进程..."
+    bash "$ROOT/scripts/stop.sh" -y
+    echo "✓ 已停止, 且守护不会再把它拉起。恢复运行: bash scripts/守护重启.sh on"
+    ;;
+
+  on)
+    rm -f "$PAUSE"
+    echo "✓ 自动重启已开启。若当前没在跑, 一分钟内会被拉起。"
+    ;;
+
+  tick)
+    # 暂停中 → 什么都不做(这就是 off 之所以能稳关掉的原因)
+    [ -f "$PAUSE" ] && exit 0
+    # 防并发: 同一时刻只允许一个 tick 启动
+    exec 9>"$LOCK"
+    flock -n 9 || exit 0
+    _rotate_applog   # 每分钟顺带检查一次, 跟要不要重启无关(常驻不重启时全靠这个兜底)
+    running && exit 0
+    echo "$(date '+%F %T') 检测到未运行 → 启动" >> "$GLOG"
+    setsid nohup $START >> "$ROOT/app.log" 2>&1 &
+    ;;
+
+  status)
+    echo "项目目录: $ROOT"
+    if crontab -l 2>/dev/null | grep -qF "$CRON_TAG"; then
+      echo "守护定时任务: ✓ 已安装(每分钟检测)"
+    else
+      echo "守护定时任务: ✗ 未安装"
+    fi
+    if [ -f "$PAUSE" ]; then
+      echo "自动重启:     ⏸ 已暂停 (.guard_paused 存在 → 不会拉起)"
+    else
+      echo "自动重启:     ▶ 开启"
+    fi
+    if running; then echo "main.py:      运行中"; else echo "main.py:      未运行"; fi
+    ;;
+
+  *)
+    echo "用法: install | on | off | uninstall | status | tick"
+    ;;
+esac