A CANONICAL ANSWER · FROM THE VINDEX KNOWLEDGE GRAPH
A scale prices the codes
A scale converts small codes back into real magnitudes: a group of weights shares one, and each weight's code multiplies it. A zero point shifts the grid so it need not be centred on zero. They are the metadata a scheme spends its bytes on — and they are never free: NVFP4's one eight-bit scale per sixteen weights is precisely the half bit that turns 'four-bit' into 4.5 bits per weight.
SUPPORTED · derived — 72 bits per 16 weights: 64 code bits + 8 scale bits