Probative ingests ordinary documents.
Every extractor produces a normalised text projection.
Offsets into that projection must survive re-extraction.
