quarantine 'em

(People are less likely to hand-edit them when they move steps around, which is what we want for stable IDs.)

And I guess you're using UTF-16 internally so that makes sense

Nah, utf8, but visual weight in the spec source is roughly as important to minimize as transferred page weight.

(Given that the spec is huge and nearly all ASCII, using utf16 would be kinda dumb.)

Then is there any particular reason for not using Base65536? It would yield shorter markers.

Only in the limit. We want more than a 6x margin, just in case, so we'd need 2 chars anyway. Also: much less font support.

Two other comments: (1) Base32768 uses CJK characters which may *not* look like gibberish to CJK readers, particularly in short strings

(2) It may not be easy to visually distinguish one Base32768 string from another if you are using them as a "visual checksum"