I like cloud inference when the Wi-Fi is real and I want to ship a weight update without touching fifty phones. I like on-device when aisle 7 eats packets and people still need an answer.
The hybrid that has worked: a small on-device embedding for offline top-N against a cached library, and a cloud pass when the network is there and the on-device confidence is mush.
Measure the aisle, not the office
Walk the building with the phone. The conference room is not the RF environment. I log network type, round-trip, and whether we fell back to edge.
If cloud p50 is fine and p95 is 4 seconds, operators will feel the p95. They do not live in your average.
Updates are an operations problem
Edge models need a download story that does not happen during a wave. Cloud models need a rollback. Both fail if you cannot ship a fix the same day a vendor changes packaging.
