A memory leak is pretty terrible, and having to either a) disallow cycles or b) specially handle them in user code is also pretty terrible IMO. Loads of handy graph data-structures, all sorts of fun caching techniques just got either impossible or much more difficult. You definitely can work with those systems but it's like stepping into a time machine (one spitting you out around '94 at an MS COM presentation), why make do unless you have to?
I never claimed GC's solve all leaks either (they do a pretty good job with finalizer semantics honestly), but they're awfully awfully handy for solving very common (and sometimes tricky) ones.
I also did not mean to imply that our 40ms cut is a hard limit, it's a target average with a desire to minimize variability (Marc's work on which lead to this post); one that we're hitting pretty well. If you're working on a real-time system, you're right that you shouldn't be using these non-deterministic GCs.
They're not mainstream (that I'm aware of), but there are realtime GCs that can guarantee a pause of less that 1ms. Metronome is the one I'm familiar with.
Your last point seems flat wrong, btw. If you're in a GC environment with worst-case latency requirements (real requirements, not just nice-to-haves or 99.9th percentiles, or whatnot) of 40ms, then you're in a whole world of hurt. GC won't do that -- brush up on your C.