Generally if both head and tail are volatile, the head must be read prior to the tail in order to compute size. Otherwise size() can be negative which probably contradicts the general contract of the interface.
Reading head before tail ensures under no circumstances you can get ahead of the tail.

The issue is in CoalescingPowerOfTwoBuffer, size() should be {return (int) (-lastRead+nextWrite -1);}
The negative value can be observable by external threads outside SP/SC scenario.