It's a proxy for one particular style of coding questions: leetcode. It's nothing more, nothing less.
Calling it a proxy for general intelligence or coding in general is a joke. That's propaganda passed around the company to convince people that the practice should be continued. It strokes the employees own egos to think they passed some magic intelligence test so it motivates them to keep repeating that it measures intelligence/skill.
My assumption is that there are a lot more people who are studying for these interviews than you think, which is why the proxy is failing (or producing homogeneous results)
On the other hand, it's okay to target sufficient but not necessary. False negatives can be fine, so long as you are finding enough true positives.
Even if you are systemically missing out on a sub-pool who are good, it's not necessarily a bad thing - particularly if you are realistic about its limitations. To my mind the real problem isn't that you might miss good people (you will) but that if your filter is too constrained the good people you find might be too homogeneous.
IMO the biggest problem is that small startups copy the Google's approach while their hiring environment is completely different.
I work at Google, opinions are my own.
I think your assumption that Google interviews based on what you will do on the job is wrong. The reality is that for most SWE positions, we can't assume you will know the things you'd be using for the job because of all the internal tools. Thus, algorithm interviews are more or less a proxy for some combination of general intelligence and coding questions.
I'm not saying it's perfect and there are many things I don't like about it, but I think it's important to know what the goals are when making a critique of whether the process meets those goals.