Add SDP-pruned, WebGPU-compatible MeloTTS model (v2) — removes GatherND-based stochastic duration predictor; deterministic durations, bit-identical audio on CPU

docs(extension): rename Plus → GPU Lite tier (matches extension UI; mobile still Plus)

docs(extension): document Plus tier (neural-melo-en) + engine tier table

melo extension: upstream LICENSE (MIT)

melo extension: tokens (phoneme → id map)

melo extension: lexicon (enriched, incl. em-dash silence rule)

melo extension: upload fp32 model

Upload 27 files