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_display_grid.py

@@ -0,0 +1,81 @@
+"""5x4 camera mosaic helpers."""
+import numpy as np
+
+from display_grid import (
+    CameraGrid,
+    camera_ids,
+    display_name,
+    is_camera_identity,
+    letterbox_bgra,
+    placeholder,
+)
+
+
+def test_camera_ids():
+    ids = camera_ids(20)
+    assert ids[0] == "cam-01" and ids[-1] == "cam-20"
+    assert len(ids) == 20
+
+
+def test_is_camera_identity():
+    assert is_camera_identity("cam-07")
+    assert is_camera_identity("cam-07", "cam")
+    assert not is_camera_identity("speaker")
+    assert not is_camera_identity("display-wall")
+
+
+def test_display_name():
+    assert display_name("cam-07") == "CAM-07"
+    assert display_name("cam-01") == "CAM-01"
+
+
+def test_letterbox_preserves_canvas_size():
+    src = np.zeros((720, 1280, 4), dtype=np.uint8)
+    src[:] = (0, 0, 255, 255)
+    out = letterbox_bgra(src, 384, 270)
+    assert out.shape == (270, 384, 4)
+    assert out.dtype == np.uint8
+
+
+def test_grid_render_size_and_slots():
+    g = CameraGrid(width=1920, height=1080, cols=5, rows=4)
+    assert g.tile_w == (1920 - g.gap * 6) // 5
+    assert g.tile_h == (1080 - g.gap * 5) // 4
+    assert g.inner_w < g.tile_w
+    assert g.inner_h < g.tile_h
+    assert len(g.identities) == 20
+    buf = g.render()
+    assert len(buf) == 1920 * 1080 * 4
+    frame = np.zeros((720, 1280, 4), dtype=np.uint8)
+    frame[:] = (255, 0, 0, 255)
+    g.set_frame("cam-07", frame)
+    buf2 = g.render()
+    assert buf2 != buf
+    g.clear("cam-07")
+    assert len(g.render()) == len(buf)
+
+
+def test_empty_tiles_have_name_placeholder():
+    g = CameraGrid(width=1920, height=1080, cols=5, rows=4)
+    # Name is drawn in the inner region — not a flat fill.
+    tile = g._tiles["cam-01"]
+    assert tile.shape == (g.tile_h, g.tile_w, 4)
+    assert tile.sum() > 0
+    inner = tile[g.border : g.border + g.inner_h, g.border : g.border + g.inner_w]
+    # Center of an empty tile is not uniform (text glyphs).
+    assert inner.std() > 1.0
+
+
+def test_gap_separates_tiles():
+    g = CameraGrid(width=1920, height=1080, cols=5, rows=4, gap=8)
+    canvas = np.frombuffer(g.render(), dtype=np.uint8).reshape(1080, 1920, 4)
+    # Pixel in the gutter between col 0 and col 1 is canvas background, not a tile.
+    x = g.gap + g.tile_w + g.gap // 2
+    y = g.gap + 4
+    assert tuple(canvas[y, x]) == (10, 10, 10, 255)
+
+
+def test_placeholder_draws_text():
+    img = placeholder(640, 360, ["Screen share", "stub"])
+    assert img.shape == (360, 640, 4)
+    assert img.sum() > 0