feat(search): shared transposition table + Lazy SMP

[cjunius]

What

Two search features from the roadmap, both leaning on the Zobrist hash dragontoothmg already exposes (Board.Hash()):

Transposition table (internal/engine/transposition.go)

• Fixed-size, power-of-two table; size set by the new UCI Hash option (default 64 MiB).
• 16-byte slots, each a pair of atomic.Uint64 words accessed with Hyatt's lockless XOR trick (word0 = key ^ data) — safe for concurrent Lazy-SMP access with no mutex; a write torn across goroutines simply reads as a miss.
• data packs best move, an int32 score, depth, and an EXACT / LOWER / UPPER bound flag.
• Mate scores are rebased by ply on store and probe so "mate in N from the root" survives a table round-trip.
• negamax probes for a cut-off and stores its result with the correct bound; the stored move is ordered first in orderMoves even when the entry is too shallow to cut.

Lazy SMP (internal/engine/search.go)

• SearchParams.Threads workers run iterative deepening in parallel, each on its own copy of the board (dragontoothmg.Board is all value types), over the one shared table.
• Workers start at staggered depths so TT races push them into different subtrees; the deepest completed result wins.
• The UCI layer owns a persistent engine.TT (cleared on ucinewgame) and honours a new Threads option, capped at runtime.NumCPU() at search time.

Modelled on the same design in pyChess (lazy_smp.py / shared_tt.py), adapted to Go's shared-memory goroutines instead of multiprocessing.

Impact (Apple M4, depth from the opening)

depth	before	with TT (1 thread)
7	3.4s / 23.6M nodes	0.5s / ~3.1M nodes
8	30s / 161M nodes	1.3s / ~5.9M nodes

Basic Lazy SMP (no root-move splitting yet) adds parallel breadth and tactical robustness; fixed-depth wall-clock is roughly flat pending root splitting, which is noted in the roadmap and docs.

Tests

• transposition_test.go — bound/window semantics, depth-preferring replacement, ply rebasing round-trip, and a -race concurrent-access test.
• search_test.go — TT does not change the best move, warm TT cuts node count, and the Lazy-SMP path finds mate / wins material / respects movetime.
• go test -race ./..., golangci-lint (v2.13.2), govulncheck, gofumpt all clean.

🤖 Generated with Claude Code

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

❌ Patch coverage is 79.06977% with 36 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
internal/uci/uci.go	35.89%	25 Missing ⚠️
internal/engine/transposition.go	88.67%	6 Missing ⚠️
internal/engine/search.go	93.75%	5 Missing ⚠️

📢 Thoughts on this report? Let us know!