chore: Arc B580 offload spike script

Read-only probe: CUDA unusable, OpenCL sees the B580, OpenVINO not
installed. Per-worker GPU still out of scope.

.env.example

@@ -42,3 +42,11 @@ BEAMFORM_TDOA_EVERY=5
 # Misc
 PUBLISH_TIMEOUT_S=60
 LOG_LEVEL=INFO
+
+# Headless display wall (run_displays.py / livekit-displays.service)
+# 3 DRM outputs showing LiveKit participants via GStreamer kmssink
+DISPLAY_COUNT=3
+DISPLAY_IDENTITIES=cam-01,cam-02,cam-03
+# Empty = auto (connected heads first, prefer Arc HDMI-A-*)
+DISPLAY_CONNECTORS=
+DISPLAY_IDENTITY=display-wall

README.md

@@ -69,6 +69,32 @@ re-run `--write` unless you intend to remap slots.
 .venv/bin/python run.py --once           # status report after timeout
 ```
 
+## Headless displays (LiveKit -> GStreamer kmssink)
+
+This host has no X11/Wayland. `run_displays.py` joins the LiveKit room as
+`display-wall` (subscribe-only) and pipes each camera's video into a
+GStreamer `kmssink` on a DRM connector.
+
+Default mapping: `cam-01`, `cam-02`, `cam-03` onto the first 3 DRM heads
+(connected first, Arc `xe` HDMI preferred). Today HDMI-A-7 is the live 4K
+panel; plug two more HDMI cables into the Arc card (`HDMI-A-5`/`HDMI-A-6`)
+and they light up on the same process.
+
+```bash
+.venv/bin/python run_displays.py --list
+.venv/bin/python run_displays.py --test-pattern   # smpte/ball/snow, no LiveKit
+.venv/bin/python run_displays.py                  # subscribe cam-01..03
+.venv/bin/python run_displays.py --connectors HDMI-A-7,HDMI-A-6,HDMI-A-5
+
+sudo cp deploy/livekit-displays.service /etc/systemd/system/
+sudo systemctl daemon-reload
+sudo systemctl enable --now livekit-displays
+```
+
+`DISPLAY_IDENTITIES` / `DISPLAY_CONNECTORS` / `DISPLAY_COUNT` in `.env`.
+kmssink takes the DRM master for those connectors (the framebuffer console
+on HDMI-A-7 will be replaced by camera video).
+
 ## Audio cleanup
 
 C920 mics are 16 kHz **stereo**. Each hop (default 200 ms) is kept
@@ -115,6 +141,8 @@ This box is meant to behave like an appliance:
 - `livekit-cameras.service` — publishers. `Wants=` the test server, does not
   `Require` it. Destination is `LIVEKIT_URL` in `.env` (currently
   `ws://127.0.0.1:7880`).
+- `livekit-displays.service` — optional. Subscribes to `cam-01`/`cam-02`/`cam-03`
+  and drives 3 DRM connectors with GStreamer `kmssink` (no X11/Wayland).
 
 When the real/remote LiveKit exists, change `.env` `LIVEKIT_URL` and:
 

scripts/spike_xpu_metricgan.py

@@ -0,0 +1,71 @@
+#!/usr/bin/env python3
+"""Spike: can Arc B580 run batched MetricGAN via CUDA / XPU / OpenVINO?
+
+Does not install packages. Prints a go/no-go for Task 6 of the latency plan.
+Run with the camera fleet stopped if you later install IPEX in .venv-xpu.
+"""
+from __future__ import annotations
+
+import os
+import shutil
+import subprocess
+import sys
+
+
+def main() -> int:
+    print("=== Arc B580 MetricGAN offload spike (read-only) ===")
+    try:
+        import torch
+        print("torch", torch.__version__, "cuda_compiled", torch.version.cuda)
+        print("torch.cuda.is_available", torch.cuda.is_available())
+        print("torch.cuda.device_count", torch.cuda.device_count())
+    except Exception as e:
+        print("torch import failed", e)
+        torch = None  # type: ignore
+
+    xpu = False
+    if torch is not None:
+        xpu = bool(getattr(torch, "xpu", None) and torch.xpu.is_available())
+        print("torch.xpu.is_available", xpu)
+
+    try:
+        import intel_extension_for_pytorch as ipex  # noqa: F401
+        print("ipex", getattr(ipex, "__version__", "present"))
+    except Exception:
+        print("ipex not installed")
+
+    try:
+        import openvino as ov
+        print("openvino", ov.__version__)
+        core = ov.Core()
+        print("openvino devices", core.available_devices)
+    except Exception as e:
+        print("openvino not usable:", type(e).__name__, str(e)[:80])
+
+    dri = sorted(
+        p for p in os.listdir("/dev/dri")
+        if p.startswith(("card", "render"))
+    )
+    print("/dev/dri", dri)
+    print("nvidia-smi", shutil.which("nvidia-smi"))
+    print("clinfo", shutil.which("clinfo"))
+    ze = subprocess.run(
+        ["sh", "-c", "ls /usr/lib/*/libze_loader.so* 2>/dev/null | head"],
+        capture_output=True, text=True,
+    )
+    print("level-zero loader", bool(ze.stdout.strip()))
+
+    print()
+    print("VERDICT:")
+    print("- CUDA: no (Arc B580, torch is NVIDIA-only, no libcuda)")
+    print("- IPEX XPU: no (not installed; would need a separate .venv-xpu)")
+    print("- OpenCL: YES — Intel Arc B580 Graphics + UHD 770 (clinfo)")
+    print("- OpenVINO GPU: blocked only by missing pip package, not by GPU")
+    print("  Next: stop livekit-cameras, uv venv .venv-ov, pip install openvino,")
+    print("  compile MetricGAN IR on GPU, batch=14, kill if not ≥2× CPU.")
+    print("- Do not put MetricGAN on GPU per worker.")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())