Text encoder FP32 reference

Invoke-KokoroTextEncoder.ps1 implements the stock batch-one TextEncoder from kokoro/modules.py at dfb907a02bba8152ca444717ca5d78747ccb4bec: token embedding, three channel-first weight-normalized Conv1D blocks with channel LayerNorm and LeakyReLU (slope 0.2), followed by a bidirectional LSTM. Dropout is inactive in eval mode. The model-neutral convolution uses PyTorch's pinned weight-normalization contract at 2b3ec34829036a65cd9d1398ea72a0167dc37470.

Test-KokoroWeightNormConv1d.ps1 passes an analytic center-tap test and matches the pre-existing AdaIN-specific convolution reference. Test-KokoroTextEncoder.ps1 passes a bounded analytic composition and a two-token stock-weight shape/finite gate after verifying the pinned checkpoint. The output layout is channel-first [512, tokens], ready for the already-gated frame-to-token gather when the full branches are joined.

This does not establish full-length numerical parity, text-to-phoneme admission, F0/N, decoder execution, PCM, or audible speech.