PerformanceMetricsStates doesn't appear to support arbitrarily long delays between writing and reading.

While it holds time ranges for for each state, that's isn't sufficient if, for example, there's a 3 frame delay for (not impossible for a frame that gets blocked on GPU composition, for example). You could have frames 1,2,3 where 1 & 3 have a single frame state key/value, but not 2.

Additionally, a couple other problems:

PerformanceMetricsStates#getIntervalStates uses nanoTime (via markSingleFrameStateForRemoval) to affect single frame states, which seems incorrect. It should be only using data from the times passed in.

It should instead be driven by the values passed in getIntervalStates(), since the time getIntervalStates is called may be some arbitrary point in the future. The behavior for single frame states should work even if they're not

getIntervalStates is destructive, so multiple JankStats instances will interfere (only the first will get single frame states). Instead, consider separating into two apis:

getIntervalStates() (as is, but just a getter, non destructive)
discardDataBeforeTimestamp() (discards data that happened before the end UI thread timestamp of last frame)