Learned generator primitives

The repository's existing bounded PowerShell FP32 AdaINResBlock1 reference (three AdaIN/Snake/dilated-convolution residual pairs) remains unchanged. This milestone adds weight-normalized transposed convolution for both generator upsamplers and ordinary strided convolution for the two source-spectrum injections. These are separate verified primitives; the complete learned generator is not yet composed or numerically gated.

Source identity: Kokoro dfb907a02bba8152ca444717ca5d78747ccb4bec, kokoro/istftnet.py (AdaINResBlock1, Generator), and pinned PyTorch 2b3ec34829036a65cd9d1398ea72a0167dc37470 for transposed-convolution weight layout and weight normalization.

Executable gates (the new upsampler and noise-convolution gates verify the stock checkpoint digest):

• pwsh -NoProfile -File tools/Test-KokoroAdaInResBlock1.ps1: existing independent analytic three-pass composition and missing-parameter rejection — pass. This is not a stock-weight numerical gate.
• pwsh -NoProfile -File tools/Test-KokoroWeightNormTransposeConv1d.ps1 -CheckpointPath <pinned checkpoint>: analytic scatter/stride and both stock upsamplers, producing 20 and 120 frames — pass.
• pwsh -NoProfile -File tools/Test-KokoroNoiseConv1d.ps1 -CheckpointPath <pinned checkpoint>: analytic stride/padding and both stock spectrum convolutions, producing 20 and 121 frames — pass.

The second spectrum path has 121 frames because centered STFT includes its edge frame; the learned feature path has 120 frames and the stock generator reflection-pads one frame before adding the source path. The frame count is verified, but the addition and subsequent residual stack are not yet gated. No independent PyTorch numerical differential, full-generator run, PCM, or speaker output is claimed.