I accept your instinct on this as sufficiently authoritative to take at face value. Only you know what is really going on during the first phase scan.

What you've revealed in this thread strongly suggests it's reordering data, not merely reading it. My adaptation of your bitmap idea for use as a run-length indicator admitted popped into my head whilst I was presuming Phase I was simply a linear read operation.

It's still a fun problem, but I'm hard pressed to find anything else that might be optimized in Phase II based on metadata collected during Phase I.

I have no doubts you are taking into consideration all the conventional optimizations -- not that peer review hurts, even for the experienced, but that isn't the question you asked here. :-)