Connected acoustic branches

Invoke-KokoroAcousticBranches.ps1 now carries one admitted token sequence through the bounded PowerShell ALBERT reference, the stock 768-to-512 projection, the style-conditioned duration branch, and a single alignment map. The same token IDs enter the stock text encoder; its channels are gathered by that map. The aligned duration features enter the F0/N branch. The 256-element voice row is split exactly as in Kokoro's forward_with_tokens: first 128 elements for decoder style, last 128 for duration and F0/N style.

Source identity: Kokoro dfb907a02bba8152ca444717ca5d78747ccb4bec, kokoro/model.py (KModel.forward_with_tokens) and kokoro/modules.py (ProsodyPredictor, TextEncoder, DurationEncoder). The ALBERT reference also traces Transformers 8ac2b916b042b1f78b75c9eb941c0f5d2cdd8e10.

Gate: pwsh -NoProfile -File tools/Test-KokoroAcousticBranches.ps1 -CheckpointPath <pinned checkpoint> passed using the digest-verified stock checkpoint. The gate used three boundary-delimited tokens, speed 100 to keep the scalar duration within its bound, one ALBERT repeat, and one layer in each configurable encoder. It checked a one-frame-per-token map, style-half selection, aligned text dimensions, doubled F0/N time dimension, and finite values. The complete component gates were also executed in the test.

This is a structural and finite-output gate, not an independent numerical differential of the connected model. It does not establish normal-speed or full-utterance execution, decoder synthesis, PCM, audible speech, or device performance. No Python, QNN, ONNX, or external model runtime ran in the gate.