Darveht/zenvion-voice-detector-v0.4

v0.4.1: fix config/labels/inference/training/datasets; add documented untrained-init checkpoint

Fix: logits concat in ZenvionOutput, id2label(109), noise(5)/intent(15)/accent(20), v0.3→v0.4, add weight-gen script

Remove test file

v0.4: Add missing config files, label mapping, vocab, CITATION + full README rewrite

test write access

Update to v0.4: fixed architecture, new training pipeline, inference, dataset loader & evaluation scripts