memory: face-rec service for Cicada pending Prime's four calls (spec + embedder addendum from brokkr)

persistent-memory.md

@@ -158,6 +158,26 @@ hermes-gateway, Booth, ttyd, zellij-web and dev-backup are running.
   checked; nothing has been rebooted.
 → `persistent-memory.d/2026-09-25-nh3-ml1-live.md`
 
+### ⏳ Face-recognition service for Cicada — WAITING ON PRIME (2026-09-26 0116)
+
+brokkr-smithy-dev relayed operator approval (althing threads `01M3ECEPDHQDS4QYFS2Z5P4W7M`
+spec, `01M3ECJ3EV7M4F793Y1B3JN5HF` addendum).
+- API: enroll, recognize, verify, add_samples, identities CRUD, health, metrics.
+- Store: SQLite gallery with in-memory cosine search. The gallery MUST be in the
+  private backup rotation (operator ruling).
+- Addendum: SCRFD detector + a **swappable embedder**. buffalo_l ArcFace ships
+  first; AdaFace ir101-webface12m and LVFace are the bench candidates, and the
+  threshold is calibrated per embedder on real frames.
+- No face-rec service exists on the fleet (checked).
+- Four calls surfaced to Prime:
+  - repo: its own under `pfi` (name his call) or eshpfi;
+  - placement: esh-ml1 recommended (same site as HA and the house);
+  - backup path: default is a nightly SQLite `.backup` to a restic-covered path,
+    since the GPU CTs are outside vzdump;
+  - process: full pipeline recommended, or reduced.
+- License flagged once: InsightFace packs are non-commercial research; YuNet+SFace
+  is the clean alternative.
+
 ### ⭐ esh-ml1 = the fleet's utility GPU host (2026-09-25): embed/rerank (TEI) + reward (vLLM)
 
 TEI is the fleet embed/rerank engine (Prime), LOAD-SHARED across esh-ml1 and