I've no idea what exactly happens when that +200 jump occurs, maybe there is some kind of load in that moment. Maybe some calculations are still being done when the miner is closing (if there was no load at that moment, why would it go to p0 after all and not to some low-power mode?). I've had very stable rigs hashing for 24+ hours but crashing when the miner is closed (and then, after forcing cuda apps to run in p0 on those systems, they're stable at those clocks for days/weeks). I've seen quite a few similar reports from other users, afaik it's a relatively common issue.