Skip to content

MOVEGEN-HANDOFF.md

Latest commit

 

History

History
69 lines (46 loc) · 1.98 KB

MOVEGEN-HANDOFF.md

File metadata and controls

69 lines (46 loc) · 1.98 KB
Raw

Movegen + core handoff

main @ 9302db0 — movegen closed, §A (Zobrist/Undo) merged.
Branch movgen-improvements — same as main after fast-forward.

Done

Item Status
Shift L2 / TOPO wall masks ✓ production
Lazy L3, wall_masks, split loops ✓
Perft bulk d1, gates exact ✓
§A const Zobrist, fused deltas, slim Undo ✓ merged main
§B pawn default ✓ O1Lookup — perft-proven fastest (see MOVEGEN.md)
Movegen multithread / GPU ✗ policy: never

Gates

perft 3 = 2_062_264
perft 4 = 247_569_030
perft 5 = 28_837_934_502   (sub-12s, public record)
perft 6 = 3_257_436_276_501
cargo test --release → all pass
titanium bench 3 20 → ~210–240M nps (honest)

O1 pawn — PRODUCTION DEFAULT (was research-only; promoted)

Superseded note: O1 was previously kept research-only. It is now the default because it is decisively faster and verified correct — do not be misled by any lingering "research only" text elsewhere.

  • generate_legal_moves_slice uses PawnGenMode::default() → O1Lookup.
  • Fastest at perft(4) in both plain and BMI2/PEXT builds; correct vs the oracle at d3/d4 and cross-verified at d5/d6.
  • Tables (~2MB) are a fixed offline artifact; regenerate / cold-start-generate with cargo run --bin movegen-o1-gen.
  • ShiftCanStep / Scalar remain only as portable bench/test baselines.

Next work (Fable or Cursor)

1. L3 flood fraction in search (not perft)

Profile wall-heavy replay positions — is §C incremental L3 worth a proof harness?

2. §C incremental L3

Blocked until harness spec + tests vs scalar flood (BUG-DIARY a1h/a5h).

3. §D completeness oracle

Batch exact solve + invariant hash collisions — research track.

4. Eval / search (STATE.md)

Distance cache, opening depth, Game A/B replays.

Do not redo

Movegen tables, topo tables, movegen threads, O1 as default without in-search proof.