At this point in testing, I've got a slot of data flowing into transmission queues (and then being thrown away :-) This means I'm taking every image, pointcloud, and pose with C API callbacks. I don't use the java tango api, and the java app is holding about 48MB of data on average and is configured for a large heap.
For some reason, I keep getting into a situation where I get a log message from the tango_client_api -
virtual void TangoMortician::binderDied(const android::wp&): Internal Error: TangoService died.
Following this, oddly enough I get a successful connect to the tango service, but everything's pretty much borked.
The only cure I have found is to reboot the tablet, but whatever my app is doing to annoy it causes the issue to return after a handful of debug cycles.