This has come up in https://github.com/apache/arrow/pull/673 and in prior discussions.
The basic summary is that we should not write non-zero padding bytes in IPC messages. However: one cannot in general rely on the padding being non-zero when the data is in memory (for example: zero-copy slices of Arrow arrays/vectors).
I think it would be good to clarify this point in Layout.md – namely that what gets written to the wire should be deterministic. However, in-memory algorithms should not in general expect the padding region to have a particular value. As an example, a popcount on a validity bitmap would want to exclude padding bytes from the computation. Other elementwise SIMD operations are free to use the padding bytes as they wish, with a known caveat.