Signatures are dependent on the data they are signing and my public key, my public key is fixed but the data I am signing is not. I can add and remove transactions to the block to change the output of the signature. (Technically with ECDSA signatures you don't even need to do that, you just change the nonce used in signing to get a different signature)
The generation signature only depends on the
previous block generation signature and your public account number. Nothing more.
In a double spend attack (including a "51% attack) the attacker would be the one generating the sequence of blocks. That means each block relies on the prior block also made by the attacker. The attacker signs a block and if it doesn't allow him to forge the next block, just keeps resigning it until it does (as pointed out a single digest can have an infinite number of unique signatures by changing the k value). The attacker attempts signatures until he produces a one which allows him to sign the next block as well. The attacker then moves on to the next block. If this seems kind of like a PoW it is.
Part of the generation signature (called hit) is used to determine a queue of forgers. First in this queue is allowed to forge. If he did not, it is the turn of the second in the queue and so on.
The attacker won't be publishing his chain until it is longer. As long as one of his accounts is valid for signing the next block (and thus somewhere in the queue even last) there will be nobody ahead of him in the queue that knows about the block. The network doesn't require a specific signer from the queue be used, it just favors a higher signer over a lower one but all signers are equally valid. If the attacker had* >51% of the network stake e will produce the longest/best chain. Note: it isn't actually a "queue" but this doesn't materially change the scenario.
As a side note: deterministic (or quasi deterministic) signing/minting/forging is an interesting idea. It has some advantages but it isn't some magical 51% proof shield and the "nothing at risk" issue around PoW remains unchanged when compared to other PoS systems.
* It is "had" not "has" because in PoS the critical resource is not a physical item, it is a record in the blockchain. A miner who no longer has any hashpower can no longer mine but a forger who had but no longer has a stake can forge a parallel chain starting from where he had the stake and double spending the tx resulting in him losing the stake. An attacker with 51% of the stake as of block X can sell that stake and still perform a 51% attack starting from block X using the stake he had but no longer has on the main chain.