Flat, page-relative word list; AnalysisBlock.wordIndexes indexes into it. box is always normalized [0,1], origin bottom-left, y up, regardless of path.
OCR-path is genuinely per-word; TEXT-path is per visual LINE (box.w always 0, box.y the line's baseline) — coarser but honest, not faked.
One page's flat words plus its structural blocks.