RandomRotation

Random rotation for TurboQuant encoding.

TurboQuant uses random orthogonal rotations to spread quantization error uniformly across dimensions before scalar quantization. This is the first stage of the TurboQuant pipeline.

The rotation is deterministic given a seed, so the same rotation can be reproduced during decoding without storing the full rotation matrix.

Implementation uses fast random Hadamard-like rotations (random sign flips

• structured permutation) rather than full O(d^2) matrix multiplication. This gives O(d log d) rotation cost.