Density is the signal
Characters per token tells you what kind of content you are looking at. English prose sits near four. Below three means code, tables, identifiers or non-Latin script — content that costs two or three times more per page than the prose around it.
A document that is more expensive than its length suggests almost always has a few dense sections doing it. Finding them takes seconds here and is otherwise invisible.
What to do about a dense section
If it is a table, convert it to CSV — usually a two-thirds reduction. If it is code, consider whether the comments are needed. If it is a base64 blob or a list of identifiers, ask whether the model needs it at all.
For retrieval corpora, dense sections are also the ones most likely to chunk badly, so profiling doubles as a chunking sanity check.