(Edit: I'm talking about the proposal from previous threads of encrypting with a random key, not encrypting with the hash.)
1) Alice takes file P and hashes it to produce key K
2) Alice encrypts file P with key K to produce file C
3) Alice encrypts key K with key U (her user key) to produce key X
4) Alice sends key X and file C to Mega for storage
When Bob does the same steps on a matching file, the only things that differ are key U and key X -- these are his user key, and his user key for the data. That means that Mega can deduplicate file C freely, because they're identical.
Note that Mega never knows the hash of the original file -- key K -- or it'd be able to decrypt the files.
This implies that it is not convergent, unless dedup is somehow done with the "Meta MAC".
The trick lies in ensuring the key used for encryption is linked to the file, not the user. You then have to store a separate key for each encrypted file (and that encryption key is probably going to need to be wrapped by your actual user key for storage).
It seems like either the server must know the file's encryption key, or the key must be derivable from the contents... which is no good.
How could they de-dupe stuff without knowing what it is in the first place?