[ewerton336]

MaxAbsoluteNormalization, TaxicabNormalization, EuclideanNormalization and PNormalization divide by the norm of the vector. For an all zero vector the norm is 0, so every element becomes NaN.

llama.cpp handles this case in common_embd_normalize (common/common.cpp):

const float norm = sum > 0.0 ? 1.0 / sum : 0.0f;

so a zero vector stays all zeros. This change does the same: each method returns the vector unchanged when the norm is 0 (and the ReadOnlySpan overload returns a zero filled array).

Added tests in SpanNormalizationExtensionsTests that run every mode on new float[3] through the array and span overloads and expect [0, 0, 0]. They failed before the fix ([NaN, NaN, NaN]) and pass now. No model needed.

[martindevans]

LGTM. Thanks for fixing this