Say 50% is CryptoNight and 50% is an ASIC friendly PoW, would this create more security? Maybe current miners would leave, creating less security? This might be an idea that would be better for the far future once Monero starts tail emission maybe?
To evaluate whether one alternative is more secure than the other we have to answer the question: Which alternative makes an attack more costly?
In the Cryptonight-only case the cost of an attack is that of gaining a 51% advantage over all mining nodes on the network.
In the Cryptonight+ASIC case you'd need to get some combination of regular miners and ASIC miners.
It gets a little "handwavy" at this point, because a true estimate would be impossible. But by definition an ASIC can only be used for mining, whereas a general purpose computer can be used for mining or other things. Therefore, it stands to reason that the capital investment in the general computer category will be much larger than the capital investment into ASICs.
Another way to think about this is that the general purpose computer has mining value + entertainment value + productivity value + ..., whereas the ASIC only has the mining value. Because of this regular users are willing to spend more to pay for a 'miner', whereas an attacker isn't, as he's only motivated by profit.
Furthermore, enthusiasts might mine at a loss (they want to secure the network and the computer is otherwise idle) with their general computer, but nobody is going to buy an ASIC (and pay for power) to run it at a loss.
For these reasons I think the capital investment required to attack a mining algorithm targeting a general purpose compute node will always be greater than a mining algorithm targeting ASICs only, or some combination of ASICs and general purpose compute nodes.
Say 50% is cryptonight and 50% is an asic friendly pow
That scenario would bring mining centralization very quickly because once an ASIC friendly PoW is introduced only miners using ASICs can effectively compete. This is true whether you are only using 1 PoW algorithm or several.
A better solution is to stick with a tested and working algorithms (CryptoNight for now) and only replace it once a better option is found.