By using a cache with an eviction policy, you sort of already accept that the data you write at "t" might not exist at "t+1". So that solves the semantics and expectations part.

For the part where an overloaded cache drops Set requests, the most frequent Set requests are the one most likely to be admitted in between the dropped ones, so I guess this is also sort of taken care of.

I have no affiliation with the project, but reading and asking myself the same question as you, I reasoned what I wrote above.