pub fn normalize_words(words: &[TimedWord]) -> Vec<TimedWord>
Trims, drops empty/zero-length words, collapses inner whitespace and clamps overlaps so each word ends no later than the next starts.