Thanks for the post. You rightly lay out the value of paying for longer term outcome and impacts that give the implementer autonomy to innovation. Unfortunately as you note much of pay for performance is still too focused on outputs. In areas where there is no known technology to achieve impacts, or that technology is highly context dependent (low external validity), paying for impacts can be a great discovery mechanism to move governments and donors from planners to searchers.
The impact evaluation problem you note is important and a valuable lesson, however it doesn't seem like it necessarily is an intrinsic problem to pay for performance. When done well and with enough forethought we can combine the best of pay for performance with impact evaluation.
Finally it's worth comparing the alternative to measuring results. Governments spend significant amounts of money doing costly and time consuming audits which really say very little about the actual results achieved. Measuring results rather than receipts doesn't have to be as costly as many believe as long as we can move governments and donors out of the realm of doing costly audits.