修复EasyOCR CUDA OOM冻结+时间戳day=0崩溃+OCR图片分目录img/raw+processed

- 冻结根因: CUDA OOM后worker不退出, 损坏context持续占显存, 父进程不杀子进程
  反复发图→线程堆积+显存耗尽→桌面冻结
- worker: readtext CUDA/kernel错误→立即退出释放显存(faulerhandler已有兜底)
- 父进程: 收到CUDA错误→立即杀子进程+30s冷却; 连续2次CUDA错误→EASYOCR_FORCE_CPU=1
  强制CPU模式启动(显存不足降级防冻结)
- 时间戳day=0崩溃: _is_timestamp_old跨午夜用replace(day-1), 每月1号day=1→day=0
  抛ValueError day is out of range for month→微信OCR整轮失败; 改用timedelta(days=±1)
- OCR调试图片回源码目录: img/raw(原图PNG) + img/processed(处理后JPG) 分目录存储
- 崩溃日志回源码根 easyocr_crash.log (不写%LOCALAPPDATA%)
- _clean_old_ocr_debug 默认 keep=20 (回退上一轮激进keep=6)
- .gitignore: 忽略 img/ 与 worker_stderr/stdout.txt

easyocr_worker.py

@@ -18,18 +18,12 @@ import os
 import traceback
 from datetime import datetime
 
-# ── 崩溃日志文件：写入 %LOCALAPPDATA%/windowmsgbot/，避开源码目录（防 360 误报 + 父进程方便取） ──
+# ── 崩溃日志文件：写入源码目录根（easyocr_crash.log），统一管理、便于父进程读取 ──
 def _get_crash_log_path() -> str:
     try:
-        local = os.environ.get("LOCALAPPDATA") or os.environ.get("APPDATA")
-        base = os.path.join(local, "windowmsgbot") if local \
-            else os.path.join(os.path.expanduser("~"), "windowmsgbot")
-    except Exception:
         base = os.path.dirname(os.path.abspath(__file__))
-    try:
-        os.makedirs(base, exist_ok=True)
     except Exception:
-        pass
+        base = "."
     return os.path.join(base, "easyocr_crash.log")
 
 CRASH_LOG = _get_crash_log_path()
@@ -81,10 +75,12 @@ except Exception as e:
 
 
 def _init_reader():
-    """初始化 EasyOCR reader（GPU 优先）。"""
+    """初始化 EasyOCR reader（GPU 优先，父进程可强制 CPU）。"""
     try:
         import torch
-        use_gpu = torch.cuda.is_available()
+        # 父进程 EASYOCR_FORCE_CPU=1 → 强制 CPU（显存反复 OOM 时降级，防冻结）
+        _force_cpu = os.environ.get("EASYOCR_FORCE_CPU") == "1"
+        use_gpu = (not _force_cpu) and torch.cuda.is_available()
     except Exception:
         use_gpu = False
 
@@ -152,6 +148,12 @@ def _main():
             sys.stdout.write(out + "\n")
             sys.stdout.flush()
         except Exception as e:
+            estr = str(e)
+            # ⭐ CUDA OOM / kernel error：GPU context 已损坏，继续跑只会反复卡死占显存
+            #   → 立即退出进程释放显存，父进程检测到 stdout 关闭后自动重启/降级 CPU
+            if "cuda" in estr.lower() or "out of memory" in estr.lower():
+                _write_crash_log(f"{datetime.now()} [easyocr_worker] CUDA 错误，主动退出释放显存: {estr[:200]}")
+                _flush_exit(f"CUDA 错误，主动退出释放显存: {estr[:200]}")
             out = json.dumps({"ok": False, "error": str(e)}, ensure_ascii=False)
             sys.stdout.write(out + "\n")
             sys.stdout.flush()