We should add some direct tests of LSM logic: e.g., that we don't throttle under "normal" conditions, and that compact reduces simple trees down to one level. For example, if a change means that LSM isn't able to flush chunks under load, that should make a test fail rather than just changing the performance of wtperf.

Either have Python code look at stats, or walk the metadata and/or the filesystem to check that the final LSM state matches what is expected. Verbose messages could be intercepted to indicate if there is any throttling?