How does Bitcoin keep inputs unique at the consensus level

When you look at how transactions are set up and what Bitcoin's protocol actually does, it's kinda interesting to think about how it makes sure that duplicate inputs don’t mess things up at the consensus level. And what if there was a soft fork that could change this to let us replace inputs in the same block, kinda like some proposals for covenants? I’ve also been pondering this from a design standpoint. Why did they go with the witness structure (segwit) as a soft fork instead of pushing for a hard fork? Plus, what kind of issues might pop up if they swapped the witness Merkel root in the coinbase transaction rather than just extending the block header?