Normalization currently works on an entire file, regardless of what is shown in any single (visible) clip from the file.
So, if somewhere outside of the visible clip there is a louder section or transient, the entire file is normalized to that loudest point.
I propose that Normalization take into account what is visible in the clip to be normalized. (Each and every clip as separated from the whole file.)
And possibly, if some users prefer the existing behavior, that existing and proposed behaviors be an Option.
Example Use Case:
When recording rock vocals, I often cut up the take into phrases. I've found the best results (most often) when these phrases are normalized to the same level. (Some phrases might need +5 dB; others might need +10 dB etc.) Compression then works the best. (The same compression applied to different passages having different loudnesses result in different sounding presentations (compression saturation) -- often not desirable.)