fix: 修复 MarkdownSplitter 死循环 + 添加 GPU 自动检测

- ingest.py: _split_single_paragraph 中分隔符距 start 小于 overlap 时
  start 会回退为负数，Python rfind 负索引导致无限循环
  修复: start = max(start + 1, next_start) 确保始终前进
- embedder.py: LocalEmbedder 自动检测 CUDA，优先使用 GPU
- ingest_obsidian.py: MAX_SIZE 调整为 200KB 适配 GPU 嵌入
- 清理临时日志文件

src/core/embedder.py

@@ -52,21 +52,33 @@ class Embedder(Protocol):
 
 # -- 本地模型 --
 class LocalEmbedder:
-    """sentence-transformers 本地模型嵌入器."""
+    """sentence-transformers 本地模型嵌入器. 自动检测 GPU."""
 
     def __init__(self, config: EmbedConfig):
         from sentence_transformers import SentenceTransformer
+
+        # 自动检测 GPU
+        try:
+            import torch
+            device = "cuda" if torch.cuda.is_available() else "cpu"
+            if device == "cuda":
+                logger.info("检测到 GPU: %s", torch.cuda.get_device_name(0))
+        except ImportError:
+            device = "cpu"
+
         self._config = config
         try:
             self._model = SentenceTransformer(
-                config.local_model, local_files_only=True
+                config.local_model, local_files_only=True, device=device
             )
         except Exception:
             logger.info("模型未缓存, 通过镜像下载 %s", config.local_model)
             old_endpoint = os.environ.get("HF_ENDPOINT")
             os.environ["HF_ENDPOINT"] = _HF_MIRROR
             try:
-                self._model = SentenceTransformer(config.local_model)
+                self._model = SentenceTransformer(
+                    config.local_model, device=device
+                )
             finally:
                 if old_endpoint is not None:
                     os.environ["HF_ENDPOINT"] = old_endpoint

src/core/ingest.py

@@ -155,7 +155,9 @@ class MarkdownSplitter:
             part = text[start:break_point].strip()
             if part:
                 parts.append(part)
-            start = break_point - self.overlap if self.overlap > 0 else break_point
+            # 确保 start 始终前进（避免分隔符距 start 小于 overlap 时 start 回退导致死循环）
+            next_start = break_point - self.overlap if self.overlap > 0 else break_point
+            start = max(start + 1, next_start)
         return parts