fix: normalize CRLF to LF in agent file parser

On Windows, .lines() strips \n but keeps \r, causing byte offset
miscalculation in frontmatter closing --- detection. The extracted YAML
content gets corrupted and serde_yaml parsing fails silently.

Normalize line endings before parsing to fix all 5 failing tests:
- test_all_built_in_agents_parseable
- test_explore_agent_disallows_write_tools
- test_general_purpose_has_all_tools
- test_scan_agents_no_dir
- test_scan_agents_with_extra_dirs_dedup

Author: KonghaYao

rust-agent-middlewares/src/claude_agent_parser.rs

@@ -162,6 +162,8 @@ pub fn parse_agent_file(content: &str) -> Option<ClaudeAgent> {
 
 /// 内部实现，返回具体错误信息
 fn parse_agent_file_inner(content: &str) -> Result<ClaudeAgent, String> {
+    // Normalize line endings to LF to avoid CRLF-related byte offset issues on Windows.
+    let content = content.replace("\r\n", "\n");
     let content = content.trim();
 
     if !content.starts_with("---") {
@@ -180,7 +182,7 @@ fn parse_agent_file_inner(content: &str) -> Result<ClaudeAgent, String> {
             after_open
                 .lines()
                 .take(i + 1)
-                .map(|l| l.len() + 1)
+                .map(|l| l.len() + 1) // +1 for the '\n' stripped by .lines()
                 .sum::<usize>()
         })
         .ok_or_else(|| "未找到闭合的 '---' 分隔符".to_string())?;