Squash Function

VOICE_V1 uses a tanh z-score squash for projected PCA values.

For each retained PCA axis, xtask projects every token embedding, then records the population mean and population standard deviation of those projected values. Runtime code will convert a projected value to a bounded knob coordinate by subtracting the frozen mean, dividing by the frozen standard deviation, and applying owned mathx::tanh.

This choice keeps the asset payload small: two f64 statistics per axis plus the squash function identifier. It also preserves smooth ordering near the semantic center while bounding outliers without a hard percentile cliff.

T-52 validation

Finalized for VOICE_V1. After the integrated voice tuning pass, the tanh z-score squash still lands the semantic axes inside the intended pitch/vowel/contour/warble ranges without making common tokens feel pinned to the extremes. No asset regeneration was needed: the squash function and frozen statistics now carried by assets/dootdoot_asset_v1.doot remain the VOICE_V1 contract.