normalize

Scale the matrix so a full-scale input can never sum past full scale.

False by default, which is FFmpeg's own behaviour for float output and therefore what this engine's reference recordings are made of: ReferencePcmTest compares the whole audio path against ffmpeg -ac 2, and normalising by default would put the engine a fixed 7 dB below its own oracle. It is also what a listener expects, because it is what mpv and VLC sound like.

The risk that leaves is real and worth naming: the coefficients can sum above full scale on a passage loud in several channels at once. The engine's own pipeline is float and does not clip there, so this only bites at the conversion into a device that takes integer samples.

True removes that risk by arithmetic rather than by luck: the whole matrix is divided by its largest row sum, so the balance between speakers is untouched and only the level moves. It costs about 7 dB on a 5.1 film. Choose it for an integer output you cannot afford to clip.