feat: known-good CPU audio + 1080p VAAPI + KMS display wall

Working snapshot before the OpenVINO MetricGAN GPU spike:
thread-capped SpeechBrain DelaySum+MetricGAN on CPU, hop 200 ms,
video hold for A/V sync, Arc VAAPI H.264, headless kmssink wall.
32 tests passing. Do not treat this as a GPU audio path.

tests/test_av_sync.py

@@ -0,0 +1,23 @@
+"""A/V hold queue: video is delayed to match the audio hop."""
+from collections import deque
+
+from run_worker import drain_held_frames
+
+
+def test_drain_nothing_until_hold():
+    q = deque([(0.0, b"a"), (0.05, b"b")])
+    assert list(drain_held_frames(q, now=0.1, hold_s=0.2)) == []
+    assert len(q) == 2
+
+
+def test_drain_ready_frames():
+    q = deque([(0.0, b"a"), (0.05, b"b"), (0.25, b"c")])
+    got = list(drain_held_frames(q, now=0.26, hold_s=0.2))
+    assert got == [b"a", b"b"]
+    assert list(q) == [(0.25, b"c")]
+
+
+def test_no_hold_drains_all():
+    q = deque([(1.0, b"x")])
+    assert list(drain_held_frames(q, now=1.0, hold_s=0.0)) == [b"x"]
+    assert not q