v2.3.7

Added

• Health checks can now read the running PrintGuard version. The unauthenticated
  /api/health endpoint reports readiness and the installed version for uptime and update
  monitors without exposing printer, camera or configuration data.
• Hub and desktop model inference now selects the fastest runtime on each machine.
  Automatic mode benchmarks LiteRT and ONNX Runtime locally, with ONNX able to use Apple
  Core ML, Windows ML, Intel OpenVINO or NVIDIA TensorRT RTX. The new Advanced settings
  tab can pin either model runtime, and the dashboard's compute readout opens it directly.
  Docker and Unraid users can expose /dev/dri for Intel hardware, while NVIDIA hosts can
  use the -nvidia image with GPU passthrough.

Fixed

• Dashboard feeds no longer stick on "no signal" while the camera is live. Opening a
  monitor's details or a dialog left a second player running for the same camera behind it,
  and once the platform paused that hidden one it never resumed — the tile stayed frozen
  under a "no signal" overlay although the camera was streaming and monitoring never
  stopped. A tile now hands its stream to whatever opens on top and takes it back on close,
  any feed that is paused resumes itself at the live edge, and the overlay only reads "no
  signal" when the camera really has none.
• Webcams no longer drop out at random during long runs. Every captured frame started a
  fresh pool of operating-system threads to convert it, faster than they could be reclaimed,
  so after a few minutes the feed froze and monitoring stopped until the camera restarted
  itself, while memory climbed by hundreds of megabytes. Capture now holds a flat thread
  count and steady memory over hours. This affected the desktop app's own webcams most, and
  MJPEG and Bambu cameras on any hub; a snapshot taken during inference could also read a
  half-converted frame.
• Live risk no longer appears stuck at zero while inference is running. PrintGuard now
  keeps the latest score in authoritative state, bounds and conflates live telemetry
  without dropping ordered events, and uses the engine clock for detailed history.