One fireteam, one voice

Thirty-two bots move as one fireteam, a warp. To tally the squad's readings you don't need shared memory or a barrier: the fireteam sums itself in a single hardware step, every lane handing its value inward at once.

A warp is 32 threads (lanes) executing in lockstep. warp_sum reduces a value across all 32 lanes in one instruction, no shared memory, no barrier. Warp-synchronous reductions are the fastest way to combine within a block.