The only current documented guidance on this is the following:

Discord may strip certain characters from message content, like invalid unicode characters or characters which cause unexpected message formatting. If you are passing user-generated strings into message content, consider sanitizing the data to prevent unexpected behavior. For example, to prevent unexpected mentions you could consider stripping @ characters or placing a zero-width space (U-200B) after them.

However, the behavior of collapsing characters has been observed with NFC-Normalized Unicode with all control characters pre-processed out. With this being the case, the above definition about it happening only with invalid Unicode characters, or with characters with unexpected formatting is clearly not correct.

This leaves no ability for a bot to take in web content, such as from RSS feeds without removing all @ characters, which changes the intent of a message. Changing this to insert zero-width spaces is also not universally acceptible, as there may be things in web content which users should be able to copy paste without unintended effects.

If you don't want to provide alternative ways such as message flags or just rejecting "bad" messages (see #1276), then please at least document the behavior in such a way that developers can properly account for it without breaking normal formatting.