It will help to assume that there is at least one group of evil people who are investing in Bitcon's demise. Not because there are, but because there might be. So let's assume they are making a set of a billion transactions, or a trillion, and maintaining currently-being-legitimately-used hashing power. When block size is large enough to frustrate other miners, this hash power (or some piece of it) will be experimentally shifted to solving a block containing an internally consistent subset of the prepared trasnsactions to fill it - experimentally at first, but on the active Bitcoin network. One seemingly random, bloated, useless (except for the universal timestamp) block will be created and the evil group will measure the effect on the mining community - client takedowns, market exits, and whatever else interests them. Then they lie in wait, perhaps let out one more to do another experiment, but with the goal of eventually catching us unawares and doing as much damage to morale as possible.
Good concrete descriptions of the threats against which we want to guard will be very helpful. Maybe there are already unit tests for such things or requests for miners' reactions to them (as opposed to just the software's behavior). My description might be a bit too long and perhaps not a very good example, but do we have a place where such examples can be constructed?