feat: WITNESS_BACKEND env (providers|llama_cpp), default providers

feat(witness): add InferenceProvidersWitnessEngine backend

quality: explicit chatml chat format for Qwen2.5

quality: tighter sampling + longer replies on GPU

infra: CUDA build for T4 GPU + revert to base Qwen2.5-3B-Instruct

fix(ux): unbuffer SSE + thinking cursor + per-witness question chips

M6: swap to fine-tuned Qwen2.5-3B witness (LoRA SFT on Modal H100)

Days 4-7: cases 0/2/3 + gothic UI + verdict card + endless/daily

initial push: M1 Belief Engine + M2 Witness Engine + M3 vertical slice (Cases 1)