Style-conditioned duration branch

The PowerShell FP32 reference now composes Kokoro's three alternating bidirectional LSTM and AdaLayerNorm blocks in DurationEncoder, a separate predictor bidirectional LSTM, the 512-to-50 duration projection, and the frame-to-token gather. It accepts 512-channel token features plus a 128-value style vector and returns token durations and aligned 640-channel predictor features. The source contract is pinned kokoro/modules.py and kokoro/model.py at dfb907a02bba8152ca444717ca5d78747ccb4bec.

Test-KokoroDurationAdaLayerNorm.ps1 tests normalization and style re-concatenation. Test-KokoroDurationEncoder.ps1 exercises the three stock-weight layers. Test-KokoroDurationBranch.ps1 exercises an analytic zero-parameter composition and a two-token composition using the verified stock checkpoint's encoder, predictor LSTM, and duration-projection weights. The gate checks shapes, finite output, and bounded alignment; it does not compare the branch numerically with an independent stock oracle.

This branch starts from supplied token features, not admitted phonemes. A full ALBERT-to-duration gate, text encoder, F0/N, harmonic source, remaining decoder, and waveform are still required before PCM or speaker playback.