The check sums ~1400 Fourier modes in fp32 on the device and compares against the same table summed in f64, against an absolute 2e-6. That constant was read off desktop Metal, which lands at 1.98e-6 -- one percent of margin on a quantity whose size is the backend's accumulation order. SwiftShader accumulates differently and lands at 3.77e-6, so CI's software adapter fails a kernel that is doing nothing wrong. Neither number is improvable: Dawn folds a Kahan compensation away, and this floor is summation-limited. Bound it by the field instead, at 1e-3 of the perturbation amplitude -- 2.9e-5 at the default wavelength, eight times SwiftShader's measurement. The check loses no power, because what it is built to catch is a mis-indexed read into the packed mode table, which yields a field that still looks smooth and random but is wrong by O(amp), a thousand times over the new bound. This never passed in CI. Every run died earlier, at numbl's missing stdlib bundle, so the browser suite had not reached it since randnfun3 landed. Verified on both backends: SwiftShader reproduces 3.77e-6 locally and the suite passes; desktop Dawn stays at 1.98e-6.