The point is that it's very hard to predict how much of an impact the GC will have a priori. You can of course measure after the fact, and try to improve, but it's hard to architect your system in a way that you can more or less guarantee will get good GC performance (other than just not allocating any memory, of course). Malloc actually suffers from a similar problem: it's just hard to know a priori if your access patterns will work well with the internals of your allocator/GC.

IBM Metronome and similar approaches provide very strong guarantees (if you keep a certain bound on garbage creation ofc) at the expense of throughput.

What they lose is that they introduce also a certain guaranteed loss of available cpu time, due to optimizing for latency and real-time predictability.

100us counts as excellent latency? x) that's half a million CPU cycles!

That's fast enough for real-time audio, so yes, that's excellent. And that includes compaction and only a small hit to throughput. You can get sub-2us latency at a serious hit to throughput.