CascadeExit: Adaptive early-exit speculative decoding for LLM inference acceleration. 1.76x speedup on Llama-3.2-3B with 0.51% parameter overhead via confidence-calibrated cascade routing