Bounded ALBERT FP32 reference

The repository now has PowerShell correctness references for the ALBERT input embedding sum and layer normalization, embedding-to-hidden projection, multi-head self-attention with key mask and residual normalization, and the gelu_new feed-forward and residual normalization. A composition repeats the shared ALBERT layer as specified by Kokoro's pinned configuration. A bounded linear primitive covers the following bert_encoder projection. These are model computations, not Hexagon lowering or live speech.

Behavioral sources: kokoro/model.py and kokoro/modules.py at dfb907a02bba8152ca444717ca5d78747ccb4bec; Transformers models/albert/modeling_albert.py, configuration_albert.py, and activations.py at 8ac2b916b042b1f78b75c9eb941c0f5d2cdd8e10. The stock checkpoint was verified against lib/manifest.json before reading tensors. No Python, QNN, ONNX, or C# implementation is involved.

Executable gates:

• Test-KokoroAlbertEmbeddings.ps1: analytic embedding and projection arithmetic, input bounds, stock tensor shapes, and finite stock output.
• Test-KokoroAlbertAttention.ps1: analytic attention, residual norm, key mask, stock tensor shapes, and finite stock-weight output.
• Test-KokoroAlbertFeedForward.ps1: analytic residual and gelu_new branch, stock tensor shapes, and finite stock-weight output.
• Test-KokoroAlbertEncoder.ps1: repeated-layer composition and one connected stock-weight ALBERT layer from token IDs to hidden output.
• Test-KokoroBertEncoderProjection.ps1: analytic and stock-weight linear projection shape and finite output.

These gates do not yet compare full 12-layer ALBERT output numerically against an independent oracle, nor do they establish practical whole-model runtime. Duration, text-encoder, F0/N, source, decoder, and final waveform remain absent from the owned phoneme-to-PCM path. No audible speech claim follows from this result.