Segregation of GPU and CPU task allocation would be the ideal solution. The debt problem would be resolved, and you could set a low GPU cache and a higher CPU cache. It’s a requirement to have low cache for some GPU projects. This obviously exposes the clients to outages (local and server side), so such crunchers tend to hook up to several projects... A bigger problem than outages might be that Boinc thinks the host should be running GPU tasks for projects that don’t even support GPU’s, and vice versa. Boinc wide this results in many systems failing to get either CPU or GPU tasks.

Got repeated "World Community Grid Requesting new tasks for GPU" messages - for about an hour. Was using <zero_debts>n</zero_debts> and had completed a couple of restarts since using it.
Boinc would not ask for any CPU tasks from WCG, crunching a mix of WCG tasks. I put the cache up very high, and only had a few tasks in the queue, but 2 were for the Lattice Project!
It was these that threw the update system out completely; it messes with the update process in some way.
Suspended the Lattice Project and got WCG tasks straight away.
Most Boinc projects work fine together, but this one is troublesome.

[OT], but perhaps worth noting, the present MW tasks are a bit of a resource hog, resulting in sluggish performance and possibly increasing the elapsed time/CPU time delta; tasks run at 99% GPU utilization and then drops to zero, while it grabs CPU time with higher priority.