feat: enhance Milkdown editor and file system functionality

- Normalize line endings in Markdown export for DOCX files.
- Improve selection serialization to Markdown with better handling of empty documents.
- Add a new `updateFile` function to the file system for updating file properties.
- Introduce video transcoding capabilities using FFmpeg, supporting various video formats.
- Update AGENTS.md for clearer plugin structure and responsibilities.
- Add scoped styles for TreeNodeItem component to improve UI consistency.
- Implement cross-origin isolation headers in Vite configuration for enhanced security.
- Remove obsolete test_cross.py file.

backend/prompt.py

@@ -302,23 +302,24 @@ Requirements:
 - Non-empty and meaningful
 - Concise unless structure needs more
 - Follows markdown rules in system prompt
+- Use real line breaks instead of spelled-out escape sequences unless PREFIX or SUFFIX clearly requires that text
 
 === BOUNDARY DECISION GUIDE ===
 
 Step 1: Check PREFIX_ENDS_WITH_NEWLINE
 If false, ask: "Does output need to start on a new line?"
   - YES if PREFIX ends with: ":", "steps:", "items:", heading text, or complete sentence before heading
-  - If YES: start output with \\n
+    - If YES: make the first character of OUTPUT a real newline
 
 Step 2: Check SUFFIX_STARTS_WITH_NEWLINE  
 If false, ask: "Does output need to end with a newline?"
   - YES if SUFFIX starts with: heading (##), new paragraph, or list marker
-  - If YES: end output with \\n
+    - If YES: make the last character of OUTPUT a real newline
 
 Step 3: Choose newline type
-  - Use \\n\\n for: new paragraphs, before headings, starting lists
-  - Use \\n for: continuing within blocks, list items, table cells
-  - Exception: inside code fences, use \\n freely
+    - Use a blank line for: new paragraphs, before headings, starting lists
+    - Use a single line break for: continuing within blocks, list items, table cells
+    - Exception: inside code fences, use real newline characters freely
 
 === CONTEXT NOTES ===
 - OCR metadata (e.g., <OCR:description>) is hidden context, never copy to output