zlrs found that its radio could reuse the same 128-attempt quality measurement more than once, repeatedly raising transmit power and restarting a long power-saving interval. The code looked responsive, but it was reacting to old evidence. The fix gives each measurement window a generation marker, so power control can act on it once and must wait for fresh traffic before changing course again.
The repository now carries a two-board runner that flashes both radios, controls faults through test-only firmware hooks, and checks recovery. In the reported run, 152 host checks passed, the boards exchanged 2 mebibytes in both directions without changing a byte in 82.7 seconds, and both finished connected at 2.6 megabits per second. They also recovered after feedback was deliberately removed, ending at the same power setting without test hooks enabled.
Radio policy is difficult to trust from a simulator because timing, interference, and device state all matter. Keeping this fixture runnable from the repository means the next change to smoothing or power thresholds can be judged by the same transfer and recovery evidence, instead of a bench note that nobody can reproduce six months later.