Games
One of them called Math.sin. Across 200,000 inputs, Node 23 and Python 3.14 on the same M3 disagreed on sin for 4.40% of them and on tan for 41.16% — by one to three ULP. That is enough to put two clients a full world unit apart in 464 steps, 7.7 seconds of play. The same simulation using only + - *
Games
Keep the flat array and narrow it. Dropping tile ids from Int32 to Uint8 took a 4096x4096 terrain map from 64 MB to 16 MB and made random reads 3.3x faster, while hand-rolled RLE came out 1.23x larger than the plain array once both were gzipped. Measured across dense, sparse and layered maps at thre
Games
Stop running A* once per unit. On a 512x512 grid with 500 units heading to one goal, per-unit A* took 904 ms and a flow field took 5.17 ms — 175x — and the flow field was already ahead at six units. Measured across four map types, five heuristics, three open-set structures and the no-path worst case
Games
5,700. That is where an all-pairs nested loop stops fitting in a 16.67 ms frame on an M3 in Node 23. A uniform grid holds to 158,400 and a spatial hash to 123,200 — but the quadtree everyone reaches for breaks at 21,200, and updating it incrementally instead of rebuilding it made it nine times slowe
Games
On disk, 1.17x — almost nothing. In texture memory, 1.54x against sprites padded to a power of two, and 16x against a naive fixed-cell grid. Measured over 340 generated sprites with three packing algorithms, four sort orders and four padding widths. Shelf packing matched MaxRects at 1/1500th of the