perf: random Choice allocates twice per drawn element (200 013 allocations for 100 000 draws) #103

Generator.Choice, RandomState.Choice and random.Choice allocate twice per drawn element, so they cannot be as fast as numpy.random.Generator.choice, which draws the indices and gathers them in one vectorised pass.

Witness (allocation counts, exact under any load; clone at c7b9263, bench/ of the W3 refbench PR)

go test -run TestPerf -perf.record -cpu 1 (GOMAXPROCS=1), size = 100000 draws from a 1000-element array:

ChoiceN draws the same indices through choiceIndices and costs 8 allocations; the 200 000 extra are in the gather: Choice (random/generator.go:756-773) ends in takeAxis(a, idx, size, ax), and the count is 2 x size (the other 13 are fixed). NumPy: 1 allocation (the result).

To do

• Make the gather of Choice allocation-free per element (one result array); the allocation count of the three rows above must drop to a small constant. The row inventory and counts are recorded in docs/perf.json, so the PR that fixes it must record the decrease.
• Timing is deferred (1-minute load was 8 to 80 when this was filed); this issue is filed on the allocation witness alone, as the campaign rules allow.

🤖 Generated with Claude Code