[sn1] TT: 4-entry buckets, age + depth replacement (was 1 entry, always replace)

[sn1] gen 15 net 4A59F07A (c6a..c16 + c16t + DFRC d1-d4, ~2.13B, 380 SB) vs main gen 14 AE339E48

[sn1] DFRC datagen d5 from A617C0B4 (kb4, gen 12): 600k games N=15000 (genfens dfrc)

[sn1] gen c17t from AE339E48 (gen 14; main 541f5d8): 1.5M games N=15000, played out with 5-man TB adjudication (new default after A/B #3902)

[sn1] at most 4 double extensions per line (was unlimited below 2 x root depth)

[sn1] gen 15 net 0979BF47 (c6a..c16 + DFRC d1-d4, 360 SB; c16 = control adjudication) vs main gen 14 AE339E48

[sn1] fifty-move TT guard only for |score| > 50 at halfmove >= 90 (was: all cutoffs refused)

[sn1] razoring on corrected static eval (was TT-adjusted eval)

[sn1] STC no-worse: hidden 16 -> 16 on gen 15 data (4F0AF6CD; c6a..c16 + d1-d4) vs main gen 14 AE339E48

[sn1] gen c17 from AE339E48 (gen 14; main 26c3286: Syzygy, draw fixes, TT damping fix): 1.5M games N=15000

[sn1] adjudication A/B: gen 15 net trained with c16t (played out, TB adjudication; 7981241C) vs with c16 (current adjudication; 0979BF47), same history and SB

[sn1] RFP, razoring, NMP gate on corrected static eval (was TT-adjusted eval)

[sn1] FIXED NODES: 16 -> 16 with sparsity 1e-6 switched on at SB 31 (B9FB6D46) vs without (2F99B9D8)

[atlas] Per-bucket hidden layers at prod scale S4000, four times the data per stack, after the S1000 version missed by 9.5.

[sn1] gen c16 from 43F3DD60 (gen 13): 1.5M games N=15000; adjudication A/B control (current settings)

[sn1] gen c16t from 43F3DD60 (gen 13): N=15000; adjudication A/B: TB adjudication 5-man required, no win adjudication

[sn1] STC no-worse: hidden layers 16 -> 16 (2F99B9D8; gen 14 data, warm-up, reordered; vectorised float layers) vs main gen 14 AE339E48

[sn1] fifty-move damping applied after the TT (raw evals in TT), qsearch TT cutoffs off at halfmove >= 90; same curve (non-regression)

[sn1] FIXED NODES: 16 -> 32 with sparsity penalty 1e-6 (A602863B; 16% fewer non-zero groups) vs without (36CECC57), gen 14 data; hidden1 both sides

[sn1] FIXED NODES: hidden layers 16 -> 16 (2F99B9D8) vs 16 -> 32 (36CECC57), gen 14 data, warm-up; hidden1 both sides

[sn1] STC no-worse: FT 768 -> 16 -> 32 (1C57005B; gen 14 data, warm-up, reordered; local nps 0.86 of main) vs main gen 14 AE339E48

[sn1] FIXED NODES: FT 768 -> 16 -> 32 (1C57005B) vs main gen 14 AE339E48 (FT 1024, no hidden layer), same data

[sn1] STC no-worse (retry): two hidden layers 16 -> 32 (36CECC57) with the float layers vectorised (AVX2 intrinsics; was scalar, nps 0.73 -> 0.78 of main) vs main gen 14

[sn1] SPEED ONLY (identical search): AVX2 16-bit pair kernel vs current AVX2 kernel, same net 36CECC57; read per-host nps (Milan), stop after a few thousand games

[sn1] STC no-worse: two hidden layers 16 -> 32 (36CECC57; gen 14 data, warm-up, reordered) vs main gen 14 AE339E48

[sn1] FIXED NODES: two hidden layers 16 -> 32 (36CECC57) vs one 16 (B8809FFD), gen 14 data, both warm-up; hidden1 both sides

[sn1] FIXED NODES: training length: hidden layer + warm-up 400 SB (43C07679) vs 260 SB (8A964D65), same data; hidden1 both sides

[sn1] STC no-worse: 8-neuron hidden layer + warm-up on gen-14 data (91057715; reordered) vs main gen 14 AE339E48

[sn1] draw fixes: insufficient material, mate at halfmove 100 (non-regression)