we could do with a grammar for tags

[richvdh]

The spec says 'The name of a tag MUST not exceed 255 bytes', but doesn't specify an encoding, so that doesn't make much sense as it stands.

Should we allow non-ascii? (they are used as display names, so potentially) Spaces? Are they case-sensitive?

They are specific to users so homograph attacks aren't too much of a worry.

[uhoreg]

Since it's created by users for their own use, it seems like just saying that it's a UTF-8 string should be fine. I would hope that clients would present users with a list of tags to select, rather than requiring them to type it in every time, so I don't know if it matters much if we say that they're case sensitive or not. Maybe we can say that clients MAY fold case, and consider different character sequences that render identically (e.g. combining characters vs. single codepoints) to be the same, so that clients that care can implement that, but we don't force clients to have to deal with it.

[KitsuneRal]

For the record, here's how Quaternion works. You're allowed to choose any tag that Quaternion knows from other rooms, or to enter new tags. Quaternion doesn't fold case; it only converts between m.favourite/m.lowpriority and human-friendly captions, and strips/adds u. prefix with user-defined tags.
Attempts to make any kind of unification lead to undesired results: for example, shortly after adding support of u. prefixes, and trying to unify legacy non-prefixed tags with prefixed ones in UI, I've had a report that Riot shows duplicate tags (both prefixed and non-prefixed) because Quaternion was only adding prefixed tags while non-prefixed ones were still used on other rooms. I believe it's better to leave users to their own devices (pun intended) and let them clean up their own mess they may produce with homographs etc.