Client-side ML on mobile: on-device YOLO card-nano + TFLite voice (deliver the ~12MB nano model)

[ahmetabdullahgultekin]

From the (superseded) ROADMAP_CLIENT_APPS.md Phase-4 maintenance backlog — the one genuinely-open, still-relevant item. Move card detection + voice preprocessing on-device so raw media stays on the phone (privacy + CPU-only host).

• The web already runs card detection in-browser (web-app Implement iOS HMAC actuals (CommonCrypto) — iOS Phase 2 #106) but ships the 51 MB YOLOv8m (mislabeled "nano"); the true ~12 MB YOLOv8n must be delivered as a binary + SHA256 in the manifest (it's gitignored, not in any repo). Same model is needed for mobile on-device card detection.
• TFLite voice (Resemblyzer-equivalent) on-device for the VOICE factor, paired with the client-side-embedding flags already scaffolded on the backend (default OFF).

Operator-blocked on producing the retrained YOLOv8n weights (re-exporting best.pt can't shrink an m-architecture). Track as a gpuless feature; ship behind the existing client-side-embedding kill-switches.