Caching less effective on large files

[danvk]

If you add or remove a line, you shift the generated IDs of all subsequent code samples. Maybe a content-based hash would work better for literate-ts.