fix(pipeline): wire TextEncoder style_key into VectorEstimator.uncond_masker (mirror gitea 052b24d)

feat: streaming + custom voice mixing API

quality: VE style_key fix — closes 500-800 Hz centroid gap on dark voices

stability: bump default seed to 99 + re-gen samples

fix(critical): add missing residual in DurationPredictor.sentence_encoder

fix: DP residual + re-gen samples (Whisper validation 86% overlap)

security: strip absolute paths revealing dev machine + private monorepo

v0.1.0 — initial release (MLX-native Supertonic 3, ~x100 RTF on M4)