You're completely right about the &, such an easy thing to overlook, and I definitely should have added tests for the attributes as well.

I would also add a test for the attribute keys, since people generally care about both the key and the value. There doesn't seem to be any unescape methods for keys but having escaped things in the key might just not be allowed for xml, I'm not an expert in that regard. Regardless having an assert_eq! with some key value would be nice.

Finally, I think the proper way to deal with coercing the Cow/BytesStart/whatever to a &[u8] (or any other deref impl) is to use .as_ref(), instead of &*var. Obviously both work, and I feel this is mostly just a style note than anything else, though at the same time .as_ref() is a lot easier to try than desperately adding &* to make coercions work (especially for people who don't know what's going on).

Tangent: Actually the whole & thing reminds me of a bit of how there are a bunch of invariants that the user needs to enforce. Like when creating the BytesStart, the bytestring must be escaped, and the same goes for attributes unless you use a &str instead of a bytestring, in which case it will escape for you. In theory we could do checks everytime and make unchecked bytestrings be some unsafe function, though that feels like a bad ergonomic and performance hit for what probably doesn't matter most of the time. Probably. This deserves its own separate issue though