fix(runtime): recover awake ownership and batch model failures

tests/unit/test_llm_pause.py

@@ -6,6 +6,40 @@ import pytest
 from artemis.services import llm
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("replace_pause", [False, True])
+async def test_cancelled_wait_clears_only_its_own_pause(tmp_path, monkeypatch, replace_pause):
+    pause = tmp_path / "pause"
+    pause.write_text("old request")
+    entered = asyncio.Event()
+
+    async def wait(_seconds):
+        entered.set()
+        await asyncio.Event().wait()
+
+    monkeypatch.setattr(llm.asyncio, "sleep", wait)
+    task = asyncio.create_task(llm._wait_for_resume(pause))
+    await entered.wait()
+    if replace_pause:
+        pause.write_text("a newer request owns this pause")
+    task.cancel()
+    with pytest.raises(asyncio.CancelledError):
+        await task
+    assert pause.exists() is replace_pause
+
+
+@pytest.mark.asyncio
+async def test_pause_deadline_clears_file(tmp_path, monkeypatch):
+    from unittest.mock import AsyncMock
+
+    pause = tmp_path / "pause"
+    pause.write_text("failed request")
+    monkeypatch.setattr(llm.settings, "LLM_PAUSE_TIMEOUT_SECONDS", 1)
+    monkeypatch.setattr(llm.asyncio, "sleep", AsyncMock())
+    assert await llm._wait_for_resume(pause) is False
+    assert not pause.exists()
+
+
 @pytest.mark.asyncio
 async def test_timeout_wrapper_cancels_llm_call_when_caller_is_cancelled():
     started = asyncio.Event()