Hey --
This API is great. It's essentially what we hacked up on top of today's input events.
A few high level thoughts (in addition to more specific comments on the doc)
(1) "is this input event important" (and if so, what are the expectations) is a key question. As an example, let's say a mousemove event had a input latency of 200 ms. If that event did not modify the DOM you don't actually care about it because there's no way for the user to perceive it.
(2) It may be helpful for developers to get a baseline rate. We do this by logging 1/N events even if they aren't over the logging threshold (we set N differently for mouse moves vs clicks). A setting for doing this could be helpful. You could also do something like "events that are > 16 ms are sampled 1/100" so that developers can get a picture of events that long without crazy overheads. Perhaps the API should tell the user what the sampling rate of any individual entry was. This way the user can do things like create a weighted average or histogram.
(3) help people debug the events that are slow -- eg a reference to the DOM node that had the event, the ability to have a performance.mark that is only saved if the event gets sampled, etc.