Ember V1.4.0

What's Changed

• Removed game-specific root ordering, depth extensions and the Lazy SMP tactical verification helper by @ExxDreamerCode in c465244 and c34cadf
• Native ARM64 kernels for Ember V2 dense inference, SIMD256 selected by default on ARM64, verified generic release paths, scalar parity coverage and workload-scoped ARM profiling by @starius in #53
• Clock safety: reusable hard-deadline watchdog, search deadlines driven through a shared stop token, scheduler-stall clock reserve in bullet and node-counter-gated per-node wall-clock polls by @starius and @ExxDreamerCode in fed59aa, bc94802, 9abc886, c5620b1
• Added the OwnBook option gating the embedded book, with tooling and docs routed through it by @ExxDreamerCode in e61e2f6 and c1a08ca
• NNUE and eval speedups: buffered accumulator rows applied in one blocked pass and the discarded evaluation skipped at depth-zero nodes by @ExxDreamerCode in a07e152 and c615613
• Separated the tablebase band from the mate band and fuzzed the legal and pseudo move generators against the search validator by @ExxDreamerCode in 41b913b and a6b3647
• Upgraded the pinned Rust nightly to 2026-10-01, resolved the macOS PGO profiler via the Rust host triple and addressed the new Clippy lints by @starius in 6afc537, 225d34e and 12c2ccf
• Replaced the tablebase backend with maintained Pyrrhic: search-scoped Syzygy probing, a pinned compact table set for checks and PGO training, and complete Syzygy notices shipped in archives by @starius in #56
• Transposition table and UCI protocol: Release-ordered TT store/probe, race-safe Hash resize, fail-closed go parsing with a true go infinite, and child-line TT prefetch before recursing by @ExxDreamerCode in 8728c1d, 9d648a7, 724ebbb and 60f2d75
• Reworked clock budgeting into an independent model and added MoveOverhead as a UCI option alias by @ExxDreamerCode and @starius in 93b4e12 and c29e06a
• Fixed mate detection under the quiescence check cap by @starius in c3abbc7
• Correction history connected to every search evaluation site, trained from completed nodes with depth-gravity weighting and staged at search boundaries; unproven policies stay opt-in by @ExxDreamerCode and @starius in #57
• Capture and evasion moves now take LMR reductions with a separate history-scaled divisor by @ExxDreamerCode in 64ec2d5
• Quiet-history bonuses now scale linearly with depth instead of quadratically, so standard search depths actually learn history by @ExxDreamerCode in a7a20ea
• Developer tooling: retrying atomic state writes under transient locks, mandatory tuner-config and docs updates for new parameters, and the converter regression moved to Python tests by @ExxDreamerCode in e32d278, 4d7ac86 and 6a53d84
• Release 1.4.0 by @ExxDreamerCode in 7506152

Full Changelog: V1.3.1...V1.4.0

Rating:
V1.3.1 + 95 elo