i'm a big fan of compact semi-human-readable sentinels in binary data formats where structure is flexible. i'm using 3 byte prefixes for a key-value store to indicate which table a record belongs to.

same reason i hate RFC numbers BIP numbes and kind numbers and nip numbers. they mean nothing without a decoder ring. they should be descriptive, or gtfo

in any case though, if you are going to need lexicographically sortable numbers, you want big-endian.

it's just nutty to use a number when the thing could be a mnemonic and is not going to be so numerous that you can't fit it into 4 letters.

btw, this is a long established convention, the Macintosh and Amiga system libraries both use this convention of 32 bit mnemonic keys, to indicate filetypes, as key/value structure keys, and so on.

humans are not fucking rolodexes. language is words, not numbers. it is irrelevant that words are numbers: numbers are not words.