A "fork" of the blockchain is devastating. If you are unaware you have forked from the main chain you risk double spend attacks on all of your transactions (and mined block rewards become worthless). Thus any implementation needs to follow the exact same rules (including "bugs") for determining which transactions are valid.

That doesn't mean every bug needs to be implemented. It's possible there are DOS or remote code execution bugs in one implementation but not another. Diversity between clients is a good thing. If such a bug is discovered in one client it won't take down the entire network (except currently the vast majority of full nodes use the original bitcoind implementation)

Various different architectures may have benefits too (e.x. I like how btcd is splitting the blockchain and wallet pieces into separate processes)