#57 replaced naive summation with Kahan (compensated) summation in 8 hot loops: numx_stats_mean, numx_vec_dot, numx_vec_norm (L1/L2), numx_integrate_trap, numx_integrate_simpson, and signal.c's convolve/correlate/fir.
Kahan summation adds roughly 4 extra floating-point operations per accumulated element compared to naive summation. Correctness was verified (337/337 tests, sanitizers clean, exhaustive stress tests), but the per-call timing numbers in validation/results/ and the README benchmark table now reflect the pre-Kahan implementation and are stale.
Needed: re-run benchmarks on at least the platforms already validated for these functions, update the affected validation/results/*.md files and the README benchmark table.
#57 replaced naive summation with Kahan (compensated) summation in 8 hot loops: numx_stats_mean, numx_vec_dot, numx_vec_norm (L1/L2), numx_integrate_trap, numx_integrate_simpson, and signal.c's convolve/correlate/fir.
Kahan summation adds roughly 4 extra floating-point operations per accumulated element compared to naive summation. Correctness was verified (337/337 tests, sanitizers clean, exhaustive stress tests), but the per-call timing numbers in validation/results/ and the README benchmark table now reflect the pre-Kahan implementation and are stale.
Needed: re-run benchmarks on at least the platforms already validated for these functions, update the affected validation/results/*.md files and the README benchmark table.