Deploying face verification on edge surveillance platforms requires reducing repeated frame inference and precision cost without damaging biometric accuracy. This paper evaluates HAPT-EdgeFace++, a learned hardware-aware temporal routing framework built around frozen EdgeFace-XS. A 5,700-parameter PolicyNet receives 15 inexpensive quality, novelty, temporal-state and hardware-cost features and selects SKIP, INT8, FP16 or FP32. Oracle supervision is generated on CelebA from identity-fidelity and normalized hardware cost, while aligned LFW remains held out for final verification. PolicyNet reaches 94.32% oracle top-1 agreement and 94.27% balanced accuracy. FP16 preserves the 99.75% clean LFW baseline, whereas the exported INT8-QDQ backbone collapses to 57.57%, exposing a quantization failure that strongly shapes the optimal policy. On five-frame synthetic surveillance tracks, the learned policy without rescue retains 99.28% mixed-condition accuracy with 48.28% fewer backbone calls and 75.59% lower normalized precision cost; under severe degradation it retains 99.63% with 40.00% fewer calls and 66.61% lower cost. In contrast, the rescue-enabled final HAPT-EdgeFace++ configuration reaches 93.10% on mixed tracks and 99.27% on severe tracks, showing that heuristic rescue can over-route to a poor INT8 engine. PCA-256 INT8 templates reduce storage by 87.30% with only 0.067 percentage-point accuracy loss. These results support learned temporal routing and compact templates, but not the notebook's predefined overall deployment-superiority claim; INT8 backbone quantization and rescue-policy calibration remain the key limitations.
 

edge face recognition, adaptive inference, temporal frame skipping, mixed precision, quantization, PolicyNet, biometric template compression.