Monday's soak block asked the radio a temptation question. Priya stood up a second access point in the room - same network name, deliberately stronger where Alpha drives - and ran a fourteen-pass block to see whether the radio would leave the access point it's paired to for the better one. It never did. The log has a column for which physical access point the radio is attached to, and that column read the same address for every pass until we started forcing it off the air ourselves.
A setting we've written about before did that. Roaming - a radio hopping to the strongest access point with the right name - is a feature on a phone. On a body that streams its driving commands over the link, a roam means a scan, and a scan means leaving the channel mid-drive. June's ghost was exactly this: the factory default hunting a "better" network every sixty-three seconds. So the shipping config pins the radio to the access point it's paired to, and Monday was the first time that pin faced a genuinely stronger rival rather than an empty room. The robot refused the better offer, on purpose, every pass the rival was live - six of them, four with the rival both stronger and carrying traffic.
The second half of the block was rougher by design: four times, Priya used the network's own machinery to throw Alpha off the air mid-pass - the kind of disconnection a router reboot or a mesh system reshuffling itself produces in a real home. All four ended the same way. No fresh command inside the 70 ms window, so the dead-man cut the drive and the brakes brought it to a stop - braking, not coasting; we retired coasting in July. Four for four, and a new cause class opened in the run registry for link drops, so these halts get their own paperwork from now on.
The rulebook had its say too. The first drop got the full forty-minute diagnosis; the next two took the short path a written cause allows; and at the fourth, the recurrence cap fired and forced a full check anyway - even though the cause on record was Priya herself, doing it on purpose. The cap doesn't know who the attacker is. That's what it's for.
Recovery is the part with a finding in it. Three of the four reconnects took 0.8, 0.9 and 1.1 seconds - call it about a second. The fourth took six point six: the radio flapped through three attach attempts before it stuck. Priya's read on that number, verbatim: "That 6.6 seconds cost us waiting, not safety - Alpha was already halted before the flap started and stayed halted through all of it. And it is one event out of four forced drops, so treat it as the worst we have measured so far, not a ceiling we have proven; we want a lot more drops on the books before that number gets a rule." A robot that waits out a bad reconnect standing still is doing the right thing; how long a person should ever watch it wait is now a written question.
That block closes the list the range story opened in July. Same room, through a wall, a busy neighbour, a saturated channel, a microwave, and now a network trying to lure the radio away or throwing it off - each one measured, each one with its worst number written down. Six blocks, about twenty-eight and a half hours, a hundred and six passes, and the envelope Priya stated then still holds: same room and one wall is the located range.