Synthesis quote-verification drift — known issue in book ingest pipeline

1. Paraphrase drift (most common). The LLM remembers the gist of a quote, rewrites it in slightly different words, and presents it as verbatim. Example: Man's Search for Meaning — actual quote on p.33 was lowercase "everything can be taken from a man but one thing: the last of the human freedoms—to choose one's attitude..." Brain page has it as capitalized "Everything can be taken from a man but one thing: the last of the human freedoms — to choose..." with full quote marks added.

For future runs: the synthesis step should be followed by an explicit verification step, ideally a deterministic script that doesn't trust the synthesis LLM's grep. The verifier writes a media/books/<slug>.verification.md per book with the diff between intended quotes and verified quotes. The dream cycle reads the verification pages and can flag for re-synthesis if drift is above threshold.

The synthesis LLM is operating in synthesis-by-recall mode rather than synthesis-by-reading. It has the source text in ~/brain/inbox/zotero-books/<key>.md but defaults to drawing from parametric memory (what it already knows about the book) rather than reading the file. The quotes come out of memory, not from the file. Memory drifts; the file doesn't.