feature processing improvements

* better per-document prompt truncation
* paging over documents >> truncation
* will also necessitate some mechanism for storing compound features. should probably just save features as JSON so I can save multiple attributes, lists, whatever.

The issue does not name files or tests. Start by locating the current per-document prompt truncation and feature-storage paths, then trace how documents are paged and features are persisted. Done means per-document truncation, paging before truncation, and compound features with multiple attributes or lists can be stored as JSON.

Written by the indexing model from the issue text.