the fleet math contradicts itself, 75/25 provisioning, an around 20 percent cost model and an android row that just multiplies 200 nodes by 12 dollars, so that 0.047 headline only works at 100 percent utilization. the ios bit mixes up appium going through port 4723 vs WDA on 8100 directly, and noreset just keeps app data, it doesn't keep a session warm, plus that window.chrome runtime line is dead 2019 fingerprinting that does nothing on mobile. redis multiplexes one connection per client so there's no pool to exhaust, the real issue is commands queuing during an outage, and maxRetriesPerRequest null makes that queue grow forever without a commandtimeout backstop. the concurrency code has actual bugs too, a check then act race on the account to device binding, no lease ttl so dead jobs wedge a node, and the in process busy set stops working once you run multiple coordinators, that needs SET NX. calling proxy rotation and account binding a way to stop bans is still evasion, not architecture