The slow pace of recentlog conversations is kind of refreshing. You actually have a chance to think about what you're going to say...

raph points me to the GC FAQ, which taught me a few things, but still left me with the impression that for all it's flaws, the "refcounts + smarts" GC Python currently uses is still about the best choice given the constraints.

You also say:

I haven't done benchmarking, but my intuition tells me that Python's approach is one of the poorest performing.

Uhh, then sorry, you lose. I've been surprised by flukes of compiler optimisation, cache behaviour, etc too many times now to believe anything about performance -- sometimes even with benchmarks to back them up.

One point to mention:

For four, I'm sure that Python's cycle detector is slower than most mark-sweep collectors, because of all the extra accounting it has to do.

Yes, but it doesn't have to run at all often. If mark & sweep is all you've got, then it matters more. For allocating and deallocating in a tight loop rc is actually quite nice; you get to reuse the same memory over and over.