Clarification of Oct-Encoded Per-Vertex Normals

The paper that the spec refers to that defines oct encoding of normals talks about the snorm8 data type, which they define on page 6 as having an integer representation of -127 to +127. This spec uses unsigned char for the values, which is 0 to 255. Do we subtract 128, or do we cast to signed char? And either way we can end up with -128, so how is that handled?

Start by comparing the referenced paper's snorm8 definition with the specification's unsigned-char representation. Resolve whether decoding uses an offset or a signed cast, including handling of -128. Done means the specification clearly documents the representation and edge-case behavior.

Written by the indexing model from the issue text.