[atlas] Core SPSA 3827 applied, tuned and untuned both on the zero-blend S4000 net, testing whether prod-tuned margins were mis-sized for it.

[sn1] full SPSA on 3638FF8F (FT 1024) after #3801 values: 40k iters x 2 pairs

[sn1] DFRC datagen d2 from BE9A5723 (kb4): 300k games N=15000 (genfens dfrc)

[sn1] gen c12 from BE9A5723 (kb4): 1.5M games N=15000

[sn1] gen e12 from BE9A5723 (kb4): 4.5M games N=5000 (volume-vs-depth experiment vs c12 at N=15000)

[sn1] DFRC datagen d3 from 85B5CDD6 (kb4, gen 11): 1M games N=15000 (genfens dfrc)

[atlas] The actual deployment question: zero-blend S4000 with its own tuned margins against production exactly as shipped.

[atlas] Do the zero-blend tune deltas transfer? Same parameters applied to current production, both sides on the production net.

[sn1] DFRC value: gens 6-11 + d1 (D9F68A8D) vs same without d1 (8535053F)

[sn1] gen 11: kb4 FT 1024 on c6a..c11 + DFRC d1 (~1.1B), 200 SB vs main (kb4, gens 5-10)

[sn1] gen 11: kb4 FT 1024 on c5a..c11 + DFRC d1 (~1.25B), 220 SB vs main (kb4, gens 5-10)

[sn1] gen c11 from DF8B55A5: 1.5M games N=15000

[sn1] DFRC datagen d1 from DF8B55A5: 300k games N=15000 (genfens dfrc)

[atlas] Core SPSA on the zero-blend S4000 net at its native scale, testing whether prod-tuned margins are mis-sized for it.

[sn1] capacity probe: 8 king buckets vs main's 4 (FT 1024, gens 5-10, 200 SB), FIXED NODES

[sn1] FT 1536 (gens 5-10) at 10+0.1; fixed nodes +31.6 #3837

[sn1] 4 king buckets (FT 1024, gens 5-10) WITH refresh cache at 10+0.1; fixed nodes +15.3 #3836

[sn1] 4 king buckets (FT 1024, gens 5-10) at 10+0.1; fixed nodes +15.3 #3836; no refresh cache

[sn1] capacity probe: FT 1024 + 4 king buckets (ranks 1, 2, 3-4, 5-8) on gens 5-10, 200 SB, FIXED NODES vs main (FT 1024, same data)

[sn1] capacity probe: FT 1536 on gens 5-10, 200 SB, FIXED NODES vs main (FT 1024, same data)

[sn1] gens 4-10: c4..c10 (~1.27B), 240 SB vs main (gens 5-9, 160 SB)

[sn1] gens 5-10: c5a..c10 (~1.1B), 200 SB vs main (gens 5-9, 160 SB)

[sn1] LR warm-up 10% vs control 3638FF8F (same data/recipe)

[sn1] Chess960/DFRC support (pure addition, bench-identical)

[sn1] gen c10 from 3638FF8F (FT 1024): 1.5M games N=15000

[sn1] gens 6-9: c6a..c9 (~820M), 160 SB vs main (c7a..c8b, 80 SB)

[sn1] gens 5-9: c5a..c9 (~980M), 160 SB vs main (c7a..c8b, 80 SB)

[sn1] gen 9: FT 1024 on c8, c8b, c9a, c9 (~450M)

[sn1] SPSA #3801 values (tuned on 5448A4FD) vs main (FT 1024)

[sn1] full SPSA on 5448A4FD: 40k iters x 2 pairs

[atlas] Zero-blend lineage with both tails against current production, the full candidate.

[sn1] gen c9 from 39D2FF42: 1.5M games N=15000