Sanad Package 3 — Facial Recognition + Places + Memories (SCAFFOLD)

Perception + memory only; no motion. Dashboard on :8013. License features: face_rec, places, memory, mask.

Planned (sanad-recognition container):

• Faces / VIP DB — vision/face_gallery.py (enroll/upload/capture/describe, VIP flag). Recognition is Gemini-side in-context (primer images) — no local ML.
• Places — vision/zone_gallery.py visual place recognition (the /go nav action stays disabled here; it belongs to P4).
• Memories — NEW persistent visitor-profile store (attributes, notes, last-seen, linked face_id); feeds personalized-greeting primers to comms.
• Face/Mask — drives the shared sanad-mask for expressions on recognition.
• Consumes the frames topic from Sanad_Core/camera; emits recognition.event / place.event.

Depends on Sanad_Core (camera, shared sanad-mask). See the plan for details.

Build model (when implemented): self-contained like P1 — a vendor/ copy of the Sanad engine + a standalone Dockerfile (FROM python:3.10-slim) + its own docker-compose.yml, refreshed by a sync_vendor.sh. No sanad-base. (Camera frames + shared mask are the cross-container dependencies.)