RTK reports token savings, but our cost benchmarks disagree

The `rtk find` loop was fixed in 0.46.0. That's the pattern: it gets reported, we fix it.

We published our own reproduction after the JetBrains run: https://www.rtk-ai.app/blog/rtk-on-skillsbench/

Bash output is a small share of the bill, so that's the ceiling, and turn-count variance is bigger than the ceiling. Your 7% terminal share on Fable is the same finding from the other direction.

We're finalizing an integrated proxy inside rtk itself.