make the test suite and openapi generator platform-independent

Windows checkouts turned LF into CRLF (core.autocrlf=true), and several
tests plus the openapi generator hardcoded LF or POSIX path separators.
The result was a suite that was green on Linux CI but red on Windows, and
an openapi.json whose /api/hermes/* routes silently disappeared when
generated on Windows.

Pin the repository to LF via .gitattributes, strip shebangs before
Vite's SSR transform, and key route files by their posix path. Add a
manually triggered macOS verification workflow so the same checks run
on a self-hosted macOS runner.

tests/server/group-chat-history-window.test.ts

@@ -162,8 +162,12 @@ describe('group chat history windows', () => {
     expect(storage.getMessagesForContext('room-1')).toHaveLength(160)
     expect(latest?.totalTokens).toBe(expectedTotalTokens)
     expect(storage.getRoom('room-1')?.totalTokens).toBe(expectedTotalTokens)
-  })
+  }, 60_000)
 
+  // 这个用例故意写 501 条消息来触发 prune（保留 500 条）。而
+  // saveMessageAndRefreshRoom 每写一条都要重读整份转录并重新计一遍 token，
+  // 单测本身就是 O(n²)：单独跑约 1.3s，全量并行时 CPU 被抢，很容易撞上
+  // Vitest 默认的 5s 上限。这是重测试，不是卡死 —— 显式放宽超时。
   it('preserves snapshot summary tokens when the snapshot anchor was pruned from retained history', () => {
     const storage = groupServer.getStorage()
     storage.saveRoom('room-1', 'Room 1')
@@ -189,7 +193,7 @@ describe('group chat history windows', () => {
     expect(storage.getContextSnapshot('room-1')?.lastMessageId).toBe('msg-1')
     expect(latest?.totalTokens).toBe(expectedTotalTokens)
     expect(storage.getRoom('room-1')?.totalTokens).toBe(expectedTotalTokens)
-  })
+  }, 60_000)
 
   it('uses the full context transcript for the final AgentClient room estimate', async () => {
     const messages = Array.from({ length: 160 }, (_value, index) => ({