> In languages like C++ and Rust the majority of objects have trivial lifetimes and don't need reference counting at all.
Sure, that's why they can get away with an inefficient GC.
> Hence, even if per object the cost of RC may be higher than tracing, the fact that there are 1000x fewer objects to manage makes RC a huge win.
The "win" here is not RC at all but the fact that there's little reliance on GC and, of course, it's not really a "win" but something you buy by giving up on something else (i.e. a tradeoff), in this case abstraction ability (memory management details cannot be hidden as implementation details), and giving that up increases maintenance costs. So to get the footprint savings you have to pay -- either with performance, if you rely heavily on a refcounting GC, or with maintenance cost, if you don't. To get the performance you can choose to pay for RAM or for more development hours, but either way -- you have to pay.
Tracing GCs manage not to do any work for the vast majority of objects, those with trivial lifetimes -- they, too, don't manage the vast majority of objects (the allocator just bumps a pointer and the GC is never even aware of the existence of most of objects) -- by not freeing dead objects promptly which comes at the cost of spending more RAM.
In other words: good footprint, good performance, good abstraction (i.e. lower maintenance costs) -- pick two and pay by giving up on the third. C++/Zig/Rust pick the first two, Java/C#/Go pick the last two, and I guess Python picks the first and last.
There are, of course, other nuances; e.g. most refcounting GCs in common use leak memory as they are too crude to handle cycles, and not all tracing GCs are state-of-the-art.