pub fn estimate_tokens(text: &str) -> usize
Rough token estimate: word count × 1.3 + punctuation count.