Fail-soft qsearch stand pat + RFP returns (eval+beta)/2

TT replacement rebased on SE master (non-regression; #29 was +5.5 +- 4.7 at the 8k cap on the old master)

History bonus/malus x8 (gravity now saturates), LMR hist divisor x8, no 4999 clamp in the quiet stage

LMR +1 for quiets when the TT move is a capture

Capture history [piece][to][victim]: MVV-LVA + hist/8 ordering, bonus/malus on cutoffs

Cut-node tracking (SF-style flags), LMR +1 at expected cut nodes

Continuation correction history [stm][last move piece][to], on top of pawn + non-pawn corrhist

TM: soft *= (1.25-0.05*stability) * (1.5-bestmove node share)*1.35

Razoring: non-PV, depth<=3, eval + 250*d < alpha -> qsearch, fail low if q <= alpha

ProbCut: non-PV, depth>=5, beta+200, SEE-filtered captures, qsearch then depth-4 verification

LMR: -1 at PV nodes, -1 for checking quiets

Non-pawn correction history (per colour, half weight each) on top of pawn corrhist

TT replacement: same key / older generation / depth+3 >= stored depth (was always replace)

Improving (static eval > 2 plies ago): RFP margin*(depth-improving), LMR +1 when not improving

qsearch TT: probe/cutoff, TT move first, store at depth 0 (never evicting main-search entries)

History pruning: non-PV, depth<=4, quiet non-killer with hist < -2048*depth

Singular extensions (depth>=8, sb = tt - 2d, (d-1)/2 search) + double ext (non-PV, 20 margin, max 6) + multicut + negative ext

NMP dynamic R = 4 + depth/4 + min((eval-beta)/200, 3)

net gen34_h512_cos40 vs gen4_h256_cos60: gen3+gen4 data (193.9M), hidden 512, 40-epoch cosine

net gen4_h256_cos60 vs gen3_h256_cos60: gen4 data (104.5M, cos60 + current search), same arch/recipe

net: cos60 + 10 epochs quantisation-aware fine-tuning (rounded-net val loss 0.46947 -> 0.46937)

audit 2: no repetitions across null moves (pliesFromNull), 50-move rule vs checkmate, mate stop only when depth >= mate length

net gen3_h256_cos60: same data/arch, 60 epochs warmup+cosine vs 15 epochs step decay (val 0.4703 -> 0.4692)

speed: SIMD SCReLU output + SIMD accumulator updates, aligned accumulators, single-SEE qsearch, plain-net make, TT prefetch, in-check reuse, LMR log table, -DNDEBUG. Same search (bench identical); +49% nps Xeon, +31% laptop (non-regression)

staged move picker (lazy TT/captures/killers/quiets/bad captures); ~10% nps at equal ordering (non-regression)