Stock AdaIN residual block differential

Status: 64-frame differential passed; short-shape discrepancy remains.

tools/New-KokoroAdaInR0Fixture.ps1 verified the checkpoint against lib/manifest.json, loaded generator resblock 3 parameters with the repository's PowerShell checkpoint reader, and computed the full three-pass PowerShell FP32 reference. The input is deterministic, 128-channel, and all frames are valid. The style vector is zero. This is a small numerical fixture, not speech or a claim about normal voice conditioning. tools/Prepare-Resblock.ps1 prepared QNN static tensors from the same pinned checkpoint and style file. The inputs, reference outputs, and static tensors remain only in ignored build/ directories.

The isolated src/runspace/R0Emit.ps1 graph finalized and executed on the physical S23 diagnostic app. Its all-one mask removes the valid-frame mask difference; its QNN graph still uses its historical FP16-headroom AdaIN formulation. These measurements do not establish which implementation is correct:

The 64-frame SNR exceeds the harness's 20 dB threshold and passes the same-weight, same-input differential before block lowering. It does not prove all lengths, all styles, or full speech. The historical 62.10 dB result used a different full-length reference and was not used to pass this gate.

Stagewise QNN outputs at eight frames agreed with PowerShell through the first complete residual pass (57.57 dB SNR). The next AdaIN and Snake outputs agreed at 53.60 and 52.61 dB, then the first dilation-3 Conv1D fell to 2.40 dB. At 16 frames, that same convolution scored 4.05 dB. At 64 frames, it scored 53.86 dB; the full block scored 52.99 dB. A dilation-1 alternative for the short fixture also failed, and sampled fused static weights applied on the host matched the PowerShell Conv1D output within 1.81e-8. This localizes the observed short-shape discrepancy to the QNN graph/backend execution of that dilated convolution or an as-yet-unidentified interface detail. It does not establish the backend's internal cause. Keep 8 and 16 frames as regression fixtures and separately validate any product short- utterance behavior against stock semantics.

The device runner restored the diagnostic app's pre-test startup scripts, result, and four r0 files; a post-test byte comparison matched all seven backups. The backups remain in private app storage. The additional reference stage files remain in the diagnostic app's r0 directory but have a different length from the restored original input and therefore do not activate the stagewise trace on that original fixture. A copy of the repository's generic native-binding module was also staged because the installed diagnostic app did not have it; that added file remains there. No product APK, model assembly, or speech path was changed by this test.

Gates run: Test-KokoroAdaIn.ps1, Test-KokoroAdaInStyle.ps1, Test-KokoroAdaInConv1d.ps1, Test-KokoroAdaInResBlock1.ps1, Test-KokoroAdaInCheckpoint.ps1, script parsing, fixture shape and finite output, checkpoint digest, and post-test restoration comparison.