impact-and-audio: correct the frame-cost claim to a measured range
zero profiler scripts prints last and avg as the same number and --reset does
not reset it, so every reading from that tool is one frame. The first write-up
quoted two single samples as a 3x improvement. Replaced with ten samples over a
live bout: median 2.47 ms, 0.62-0.70 ms in frames where nothing resolves,
2.5-4.4 ms in the frame an exchange lands, on an arena that has grown from 173
to 244 entities since the baseline. The whole-frame number is the one that did
not regress: p50 100.97 ms against 103.02 ms.