The signal should mainly be in the green area, but this implies that the yellow LEDs will light on peaks.

You will get the best signal to noise ration when you have the most signal (no surprise) but you don't want distortion caused by the input clipping, so the red LEDs should light rarely, if at all.

Any real-world signal will have an average or typical level, but will have occasional peaks at a higher level. So, while the user can make sure that his average  level is (say) 0dB, he can't control the peaks (without using a limiter or compressor, which will reduce the dynamics). Having the extra 26dB headroom gives you lots of space for peaks without distortion.