refactor(backend): 启动环境初始化与资源项目化（零跨项目依赖）

- main.py：uv run main.py 一键启动——读取 backend/.env、TORCH_HOME/NLTK_DATA 指向项目内缓存、certifi SSL + NO_PROXY 网络直连、默认开启 s2s，启动横幅
- vendor/：补丁后 s2s 云化栈移入项目（backend/vendor/s2s-cloud），pyproject sources 改相对路径
- .gitignore：忽略 .torch-cache / nltk_data 缓存
- AGENTS.md：进程管理铁律（AI 不启动/重启服务，由用户操作）

backend/vendor/s2s-cloud/tests/test_tts_input_coalescing.py

@@ -0,0 +1,42 @@
+from queue import Queue
+from threading import Event
+
+from speech_to_speech.pipeline.control import SESSION_END
+from speech_to_speech.pipeline.messages import EndOfResponse, TTSInput
+from speech_to_speech.TTS.qwen3_tts_handler import Qwen3TTSHandler
+
+
+def _make_handler():
+    handler = object.__new__(Qwen3TTSHandler)
+    handler.queue_in = Queue()
+    handler.queue_out = Queue()
+    handler.stop_event = Event()
+    return handler
+
+
+def test_coalesce_pending_tts_input_merges_ready_sentences_and_absorbs_response_end():
+    handler = _make_handler()
+
+    handler.queue_in.put(TTSInput(text="Second sentence.", language_code="en"))
+    handler.queue_in.put(TTSInput(text="Third sentence.", language_code="en"))
+    handler.queue_in.put(EndOfResponse())
+
+    text, lang, saw_end = handler._coalesce_pending_tts_input(TTSInput(text="First sentence.", language_code="en"))
+
+    assert text == "First sentence. Second sentence. Third sentence."
+    assert lang == "en"
+    assert saw_end is True
+    remaining = handler.queue_in.get_nowait()
+    assert isinstance(remaining, EndOfResponse)
+
+
+def test_coalesce_pending_tts_input_stops_before_control_messages():
+    handler = _make_handler()
+
+    handler.queue_in.put(SESSION_END)
+    text, lang, saw_end = handler._coalesce_pending_tts_input(TTSInput(text="Hello.", language_code="en"))
+
+    assert text == "Hello."
+    assert lang == "en"
+    assert saw_end is False
+    assert handler.queue_in.get_nowait() == SESSION_END