Figured out why there was a regression: seems like the long-living objects were messing with GC object density a bit.

The net result of the last 2 changes is that we're faster just about everywhere, 2.7% overall, with ~10-17% on call-heavy benchmarks.

Over the weekend thus far, we're ~9% faster, with 35-60% speedups on call-heavy benchmarks... And I have one more idea to try.