When Worse is Better: Navigating the compression-generation tradeoff in visual tokenization