[bal7hazar]

Lot P1 of origami_hexmap: the helpers each lot re-implemented locally move to the shared files, every module uses them, the remaining gas is taken where it was measurable, and the hexmap CI job drops from 8m23s to under 4m20s (3m43s to 4m16s on the last runs, the L8 facade tests included).

Shared primitives

• helpers/bits.cairo: Bits::bitwise (AND, XOR and OR from one application of the builtin, wrapping the private corelib libfunc), Bits::and / or / xor per limb, Bits::popcount now the L7 two-limb byte-count form, Bits::popcount_small, Bits::byte_counts, Bits::top_byte / low_byte (bounded_int divisions), and the generic Set<T> of L3 (WideSet, SmallSet). The local extern fn bitwise of caver, bfs, dial and their benches are gone.
• helpers/layout.cairo: Layout::expand / expand_small run the L1 dilation (limbs, bitwise triple, up/down from X = 2P - Pe with one product each) through the new 4-value Dilation that the layer loops carry; Layout::with_interior (3 shared lookups); Layout::edge_neighbours and Layout::neighbour_in, the edge-endpoint helpers of Bfs and Dial.
• Caver::keep_component floods the frontier through BfsInternal::component (the flood of Bfs::reachable without the checks): -5.6 % cave, -10.6 % maze, -11.5 % serpentine.
• helpers/rng.cairo: the pool refills below 2^64 (bias 2^-56 per pool instead of 2^-24, cost +0.04 % per carved tile); Rng::draw_byte, draw6, shuffle6 and split216 divide with core::internal::bounded_int::div_rem (1.10k vs 1.56k). The walker draws 6 triples from each limb of every permutation. Keeping a spare pool or a refill counter in Rng was measured: +2 % per tile in the recursive generators (one more field), not kept.

Gas (before = main at 7adb9c4, 256 seeds for the spreader)

Area	Change
Spreader	-2.7 % to -9.2 % (max over the seeds), -2.9 % to -9.8 % (mean) on every library benchmark; count 1 on dense boards about -9 %
Walker	-2.3 % to -4.3 %
Mazer / Digger (the maps change with the stream)	per carved tile over 8 seeds -1.5 % (order 0), -2.3 % (order 1); digger maze -2.0 % per tile
Caver	generate 0 % to -5.6 %, keep_component -5.6 % to -11.5 %
Dial	0 % to -7.6 %
Bfs	reachable -3.4 % to -4.8 %; search, distance, range unchanged
Primitives	popcount -27 %, next_below -14 %, Layout::expand -3.6 % (17x14), BFS layer step -6 %

One library benchmark costs more: bench_mazer_17x14_order_0 +0.8 %, because the new stream carves 90 tiles instead of 89 (-0.3 % per tile). bench_rng_draw (generic u128 bound) +0.4 %: one more refill per 100 draws with the 2^64 threshold. GAS.md (new section P1) has the before/after of every library benchmark, the measured losers and the bounds; 77 budgets are lowered.

L8 facade (merged while this PR was open)

The branch is rebased on eb0f4b1. map.cairo calls BfsInternal::and, expand and expand_small: they remain as one-line delegations to Bits::and and Dilation, so the facade is untouched. Its benchmarks move with the libraries: compute_distribution -6.6 %, scenarios -2 % to -5 %, new_maze +0.8 % (the 90-tile maze); budgets in tests/bench_map.cairo are left to L8.

Visibility

The public traits keep their documented functions: Mazer::generate, Digger::maze / corridor, Spreader::generate, Bfs::*, Dial::*, Caver::*, Walker::generate. The recursion, dig and selection helpers move to pub(crate) internal impls; BfsInternal, Back, Endpoint, the stores and goals, Carver, Heading, TurnTrait, BitSetTrait and Counts become pub(crate). The BfsInternal functions that misbehave with height < 3 or width >= 128 can no longer be reached from outside the crate.

CI

The pull-request job runs snforge test --package origami_hexmap --fuzzer-runs 64: the library fuzz benchmarks keep their budgets (max over 256 seeds) and check them on 64 seeds; the 100 spreader losers are #[ignore]d. A new non-blocking job, hexmap-full, runs everything with 256 seeds on pushes to main (snforge test --package origami_hexmap --include-ignored). Locally, with a fresh build: 7m55s -> 2m46s.

Expectations regenerated

The random stream changed (refill threshold, walker pools), so the exact grids of test_mazer_generate_17x14_order_0/1, test_mazer_generate_19x13, test_digger_maze_order_0, test_walker_generate_17x14/7x7/19x13 and bench_walker_17x14_504 are regenerated (printouts included). The walker's scalar reference follows the new stream. No statistical or property test was loosened.

🤖 Generated with Claude Code