Search: add quiet History Pruning

[sebhofmann]

Summary

• prune extremely low-scoring quiet moves using the exact combined butterfly + continuation-history score from the staged move picker
• restrict pruning to non-root, non-check nodes after at least one legal move, at depth <= HistoryPruning_MaxDepth
• add UCI tuning controls (Use_HistoryPruning, HistoryPruning_MaxDepth, HistoryPruning_Margin) and a dedicated search-stat counter

Closes #6.

Correctness and instrumentation

• make clean && make both: pass
• ./perft_test.sh --fast: 29 passed, 0 failed, 4 slow cases skipped
• ./see_test.sh: 92 passed, 0 failed
• make test-nnue: pass
• ./symmetry_test.sh: existing failure (new: 10 pass / 4 warn / 6 fail; unchanged baseline: 11 pass / 2 warn / 7 fail)
• stats build, startpos depth 12: History 1,574; LMP 112,513; LMR 12,623; Futility 7,833 — triggers without dominating quiet pruning

LTC SPRT (manually aborted; no accepted Elo gain)

• TC: 60+0.6
• openings: /home/paschty/Downloads/2moves_v2.pgn
• paired colors: yes (-games 2 -repeat)
• H0/H1: 0 / +5 Elo
• alpha/beta: 0.05 / 0.05
• base commit: 056488eef9b51b70cb4b826ed028b280afd870fc
• test commit: e0846042675e40cd88bc0e83bf52a84fafdd343f

Final retained run (2026-08-12)

• last complete checkpoint (1,170 games): 101 / 133 / 936
• Elo estimate/error: -9.5 +/- 8.9 Elo
• LLR: -2.92, bounds [-2.94, +2.94]
• actual stop (1,172 games): 101 / 133 / 938
• decision: manually aborted / inconclusive; the lower SPRT boundary was not crossed
• interpretation: strong negative signal and no evidence of an Elo gain; this must not be presented as an accepted gain

Earlier aborted run (2026-08-11)

• last complete checkpoint (500 games): 55 / 48 / 397
• Elo estimate/error: +4.9 +/- 13.8 Elo
• LLR: 0.236, bounds [-2.94, +2.94]
• actual stop (505 games): 56 / 48 / 401
• decision: manually aborted / inconclusive

The PR remains open to preserve the implementation and test evidence. It should not be merged as an accepted Elo improvement based on these results.