CPU-only sparse traversal / terminal gate diagnostics. GPU backends and backends that do not expose these choices return an unavailable block.

Effective worker count for diagnostics. CPU backends should resolve `config.cpu_threads` (0=auto, 1=serial, N=clamped) inside prepare() and return tha

Built-in defaults. Convenience constructor.

Minimal URL component encoder. Avoids adding `urlencoding` just for one parameter. Matches `encodeURIComponent` (unreserved set: A-Z a-z 0-9 - _ . ~).

Built-in presets. Match the GUIDE.md table exactly. GPU=0 means "let the backend probe at runtime"; we don't gate on a hard VRAM number for arbitrary

Get the flush table for GPU upload (read-only pointer)

Get the parallel key table (prime products) for GPU upload. Required so the device-side probe can verify it landed on the right slot instead of return

Get the hash-based table for GPU upload

Get the unique5 table for GPU upload

Get all valid combos given dead cards (board + opponent cards)

Health check endpoint

Get the total node count (valid after build)

Tiny helper used only inside the walk lambda. children_offset(node) is what we want — wrapped here because the lambda accesses tree_ a lot and it's cl

v1.8.1+ phase-timing accessors. Each call returns the cumulative wall-clock time (in ms) spent in that iteration phase across the whole solve. Solver

============================================================================ v1.8.2 Phase 2: fused-kernel showdown for an OOP-traverser group. Bypasse

Helper: generate the rank bitmask for a card

Helper: get prime for a card's rank