Well, here's the thing. I don't think it's for the exact input, it's for a type of inputs. Custom hash functions for specific types of data are a basic optimization technique and I find it odd you'd even suggest every hash function should be 'DoS-immune'. There's absolutely nothing 'real world' about this unless you think the world consists entirely of hostile inputs. In the real world, people absolutely optimize.

Your argument seems to be that that's not the intent of the benchmark which may be true but it's not clear from the rules provided at all. To me, it looks like the opposite is true - they talk about using a standard hash table and most of those allow user-specified hash functions.