1) The miner's main routine sends to the hashing routine 2 (SSE) (or 4 for AVX) inputs for checking, not one.
2) A hashing routine with 2 (or 4) inputs and 2 (or 4) outputs.
That is normally how SIMD code is written yes.
But as you may have seen in the asm file above, there's hardly any SIMD in there. It's mostly serial operations. Miner program sending one input, expects one output from the hashing subroutine, thus nothing to do in batches of "same".
If you think CPU miners can be optimized the do more SIMD processing, and you are probably right in at least some cases, then go optimize and make some money.
I believe it's not only CPU miners that are at stake here. Even signatures could be validated 2-4 at a time, on the same subroutine, to get SIMD benefits. So scaling is affected as well. It would need the main routine to be adjusted accordingly instead of sending 1 for processing (per thread) to 2-4 (per thread) and expect back 2-4. Heck, even cracking speed could be improved.