glanced 0.3.0

The release with attention.sock.

Added

• Attention mode. A third socket, attention.sock, streams head pose (present, yaw, pitch, conf — derived values, never frames or landmarks) to any connected client at --attention-fps (default 8). It uses the landmarker only — no ArcFace, no arming, no enrollment — so it works for anyone with a webcam. The camera is held only while someone is subscribed and is handed over to an unlock scan on request, so an auth request always has priority. glancectl attention prints the stream; glancectl daemon --no-attention turns the socket off. This is what omarchy-attention (a blur shield without screen capture) subscribes to.
• Camera.frames(min_interval) drops frames without decoding them, which is how attention runs the landmarker at 8 fps on a 30 fps camera.

Install

There is no AUR package. From a checkout of this tag:

python -m venv .venv && .venv/bin/pip install -e '.[runtime]'
.venv/bin/glancectl fetch-model
packaging/install.sh            # --no-plugin for the daemon only

126 tests pass on this tag.