As another extreme, Turbo Pascal's earlier implementations had only one pass. Just straight from Pascal to binary code with some optimizations too. I wonder what the minimum number of passes could be by today's optimization standards. Two? Three?
We often forget this alternative way (with its pros and cons).
I feel like it may have been possible to know the full spec of everything you were working on.
It's still perfectly possible to design languages [0] for efficient single pass compilation, but they won't look like C#, Swift or Rust.
But anyway, what you describe is maybe not entirely unlike equality saturation: https://www.cs.cornell.edu/~ross/publications/eqsat/
The idea is to apply a whole bunch of optimizations in parallel, but in a non-destructive fashion. So instead of changing "x * 5" to "x << 2 + x" or whatever, you just note that both are equivalent ways of expressing the same. Then you go on applying optimizations to both variants. At some point you stop and end up with a soup that contains many many different computations that all do the same thing. Then you apply a solver once to pick out an optimal variant.
The trick is to make this scale.
I see the last stage of compiling as a special skills which requires a lot of time, especially if you want to support multiple platforms. If you're thing is to create a great programming language, then your time is better spent on that rather than create a bad or ok backend supporting very few platforms.
Let's replace "laziness" with "carefully considered trade-offs" in this context.
i.e. "The fact that a lot of modern day compilers don't have built in assemblers etc. are probably more due to carefully considered trade-offs".