Persistent secure daemon to reduce per-auth startup latency.

Every authentication attempt (sudo, login, etc.) spawns a completely fresh pam.py process that has to:

1. Start the Python interpreter
2. Import cv2, numpy, and dlib (all fairly heavy C++-backed libraries)
3. Load the face recognition models from disk
4. Only then open the camera and begin scanning

In practice this adds roughly 2+ second of pure setup overhead before the camera even activates, on top of the actual scan time. On my system (Zorin OS 18 / Ubuntu 24.04, tested on the 2.6.1 PPA build), I found myself finishing typing my password before the camera LED even turned on, simply because the cold-start overhead outlasted my typing speed.

Proposed solution

A small persistent, yet secure, background daemon (e.g. a systemd service) that:

• Starts at boot (or on first use) and imports cv2/numpy/dlib once
• Loads face models into memory once and keeps them active
• Keep pam.py bytecode in access instead of compiling it everytime.
• Listens on a local Unix socket for authentication requests

This should make the camera activate near-instantly on each sudo/login call, since the expensive one-time setup already happened at boot.

Security should be kept intact

• Security surface: a daemon that can be queried to test face matches needs care so it can't be abused outside of a legitimate PAM auth flow (e.g. socket permissions restricted to root, request validation, no way to just "ask" it to authenticate without a real PAM-triggered call)
• Camera lifecycle / privacy: should the daemon hold the camera open continuously, or open on-demand and release after each attempt/idle timeout? Continuous access has an obvious privacy downside.
• Has this architecture been discussed or attempted before? I didn't find an existing issue covering this specifically, apologies if I missed one.

Environment (for context)

• Zorin OS 18 (Ubuntu 24.04 base)
• Howdy 2.6.1 (installed via ppa:boltgolt/howdy)
• Standard RGB webcam (no IR), /dev/video0, HOG detection mode

Happy to help implement this if it's an actual issue. If this is already been addressed, I'd be glad to know. I really liked Howdy and want to do my part.