Great article! I have two technical questions regarding the implementation and benchmarking methodology:

1. Batched readv() approach: Have you considered benchmarking a batched readv() implementation, where the iovec entries point to separate header buffers and directly to the corresponding offsets in the final contiguous pixel buffer? This could allow multiple frames to be read with a single system call, avoiding intermediate pixel-data copies while reducing syscall overhead. Given the IOV_MAX limit and the overhead of constructing the iovec array, how do you think this approach would compare against V2, particularly for smaller frame sizes where syscall overhead may be more significant?

2. Statistical significance of performance differences: For configurations where the measured performance differences are only a few percent, how confident can we be that these differences reflect actual performance improvements rather than benchmark noise? Have you considered increasing the number of independent runs, randomizing the execution order of the implementations, and using paired measurements under comparable cache conditions? Reporting confidence intervals for the relative performance differences could help establish whether the observed rankings are statistically robust.