Actually, there is per-chunk HMAC support in the protocol and present revisions add HMAC for each chunk into Meta: https://github.com/Librevault/librevault-common/blob/master/...

This HMAC is computed over plaintext chunks and used to determine, if we already have this chunk in database (because of random IV we get different hash of ciphertext every time), not for verification.

And what is wrong about verifying integrity and authenticity by hash-then-sign?

Maybe nothing, but it's just kind of a weird, nonstandard way of doing things. Also it's not immediately obvious to me that your method meets modern security definitions like IND-CCA2.