feat: publish GPU model lifecycle events

tests/test_gpu_lifecycle.py

@@ -0,0 +1,75 @@
+from __future__ import annotations
+
+from datetime import datetime, timezone
+import json
+import time
+
+import pytest
+
+from obliteratus.gpu_lifecycle import (
+    GpuLifecyclePublisher,
+    MemoryUsage,
+    measure_torch_memory,
+)
+
+
+def test_fake_supervisor_observes_order_identity_and_recoverable_state(tmp_path):
+    publisher = GpuLifecyclePublisher(
+        tmp_path,
+        heartbeat_seconds=0.01,
+        clock=lambda: datetime(2026, 1, 2, tzinfo=timezone.utc),
+        run_id="run-1",
+    )
+    memory = MemoryUsage(allocated_bytes=10, reserved_bytes=12, device_count=1)
+    publisher.loading("org/model")
+    publisher.resize(memory)
+    publisher.ready(memory)
+    time.sleep(1.05)
+    publisher.release(reason="test_complete")
+    assert publisher.release(reason="duplicate") is None
+
+    events = [json.loads(line) for line in (tmp_path / "events.jsonl").read_text().splitlines()]
+    assert [event["event"] for event in events] == [
+        "loading", "resize", "ready", "heartbeat", "release",
+    ]
+    assert [event["sequence"] for event in events] == list(range(1, 6))
+    assert len({event["event_id"] for event in events}) == 5
+    assert all(event["run_id"] == "run-1" for event in events)
+    assert events[1]["reserved_vram_bytes"] == 12
+    current = json.loads((tmp_path / "current.json").read_text())
+    assert current["event"] == "release"
+    assert current["reason"] == "test_complete"
+
+
+def test_disabled_publisher_is_noop():
+    publisher = GpuLifecyclePublisher(None)
+    assert publisher.loading("model") is None
+    assert publisher.resize(MemoryUsage()) is None
+    assert publisher.ready() is None
+    assert publisher.heartbeat() is None
+    assert publisher.release() is None
+
+
+def test_runtime_directory_must_exist(tmp_path):
+    with pytest.raises(ValueError, match="must already exist"):
+        GpuLifecyclePublisher(tmp_path / "missing")
+
+
+def test_runtime_directory_rejects_symlink(tmp_path):
+    real = tmp_path / "real"
+    real.mkdir()
+    link = tmp_path / "link"
+    link.symlink_to(real, target_is_directory=True)
+    with pytest.raises(ValueError, match="must not be a symlink"):
+        GpuLifecyclePublisher(link)
+
+
+def test_measure_torch_memory_aggregates_devices():
+    class FakeCuda:
+        is_available = staticmethod(lambda: True)
+        device_count = staticmethod(lambda: 2)
+        memory_allocated = staticmethod(lambda index: (index + 1) * 10)
+        memory_reserved = staticmethod(lambda index: (index + 1) * 20)
+
+    usage = measure_torch_memory(type("Torch", (), {"cuda": FakeCuda})())
+    assert usage == MemoryUsage(allocated_bytes=30, reserved_bytes=60, device_count=2)