Unlike labels, which require fixed rules (or manual action) to place emails into them, the tabs sort emails algorithmically.
Doesn't seem right for me, but I guess I can see the draw for folks who don't want to deal with setting this stuff up themselves. (Which is probably most folks.)
We are working in user-defined contexts from day one (not all users fit into the 5-context model), but I think incorrect classification is the most important issue to address to minimize user frustration. We are approaching it by treating topic and context membership fuzzily and using graph visualization to aid in traversal of inexact queries.
I couldn't agree with you more: there are serious problems with binary membership and "automatic" filtering.