The CoreMLCustomSTFT tail omits the overlap-add/COLA normalization torch.istft applies, leaving raw output exactly 1.5x the PyTorch reference (FluidAudio#699, FluidAudio#852). KokoroTail_v2.mlmodelc folds the interior COLA envelope (1.5, periodic Hann, win 20 / hop 5) into the iSTFT deconv weights; graph unchanged. Originals are kept at KokoroTail.mlmodelc so existing consumers are unaffected until they opt in. A/B-verified: old/new output ratio 1.500000, corr 1.0.
fix(vocoder): rank-stable anchor for macOS 14 Espresso (FluidAudio#836)
Replace reduce_mean(keep_dims=false) -> rank-0 -> expand_dims with a single reduce_mean(axes=[1,2]) producing [1] directly. macOS 14's Espresso runtime mangles the rank-0 intermediate into rank 2, failing every prediction under all compute-unit presets. anchor is discarded by all hosts; x_pre verified bit-exact vs previous artifacts. Applies to ANE, ANE-zh, ANE-ja (ja vocoder is byte-identical to en).
