Decoder prelude and core reference

The bounded PowerShell FP32 decoder reference now carries aligned text features and double-rate F0/N curves through the stock input preparation, encode block, three same-rate AdaIN decode blocks, and final AdaIN upsample block. It returns 512-channel generator features at twice the duration frame rate. It does not yet run Kokoro's harmonic source, generator, iSTFT, or PCM.

Source identity: Kokoro dfb907a02bba8152ca444717ca5d78747ccb4bec, kokoro/istftnet.py (Decoder.forward, AdainResBlk1d, UpSample1d). The checkpoint identity and digest are enforced by the gates.

Executable gates:

• pwsh -NoProfile -File tools/Test-KokoroDecoderPrelude.ps1 -CheckpointPath <pinned checkpoint>: analytic stride-two and channel layout checks plus stock-weight shape/finite check — pass.
• pwsh -NoProfile -File tools/Test-KokoroAdaInResBlock1d.ps1 -CheckpointPath <pinned checkpoint>: analytic learned shortcut without upsample plus stock decoder encode shape/finite check — pass.
• pwsh -NoProfile -File tools/Test-KokoroF0NAdaInResBlock.ps1 -CheckpointPath <pinned checkpoint>: existing F0/N regression after extracting the shared AdaIN operation — pass.
• pwsh -NoProfile -File tools/Test-KokoroDecoderCore.ps1 -CheckpointPath <pinned checkpoint>: stock-weight execution across all decoder core blocks, output shape and finite values — pass.

The first full-core run exposed a reference-only channel bound of 1,024 in the depthwise transposed convolution. The stock final decode block requires 1,090 input channels; the bound was raised to 2,048, retaining the existing operation-count bound, and the full-core gate then passed. This was a bound correction, not a change to model math.

These checks establish topology, tensor compatibility, basic analytic semantics, and finite short-input execution. They do not establish an independent decoder numerical differential, normal utterance throughput, waveform synthesis, audible speech, or device performance. No external model runtime was used in the gate.