deploy: critical safety fixes (negation leakage, age-gate bypass, trust boundary)

Single-model ZeroGPU: local Qwen3-4B inference, named-patient demographics, Egyptian Arabic red-flag triggers

Guided history-taking interview: three Gemma 4 26B personas, one model, deterministic safety sentinel

Stop guard over-blocking benign narrative; add screenshots and judge CTA

Close safety-grill false negatives and harden model-claim guard

Deploy Dental SOAP hackathon app