Queued writes carry their pixels (a layer buffer shared with the RAM cache, a frame moved in)
and the writer thread converts and compresses them; rendering threads go on at once. An entry
larger uncompressed than the cache may hold is refused before it is queued.