v1.6.0

Added

• Digest carries document provenance and refuses to store fiction as fact. (#84)

  digestDocument parsed frontmatter for salience and dropped type and tags. So an observation extracted from a style-transfer exercise, a cross-model debate, a dream journal or a self-roast was stored with the same declarative content type and the same provenance shape as one from a journal. On one live store a sentence the agent wrote imitating its owner — "he learned more about my music in 20 minutes than most people" — was extracted as a fact about the owner, promoted to a memory, and refined by a later dream into "uniquely talented … distinct learning capability". The trace back to the file was only possible because source_files existed; nothing said what kind of file.

  Observations now carry source_type (frontmatter type, lower-cased) and source_tags. classifyDocument decides from them whether the document's claims are the author's facts: a type of workshop / experiment / creative / fiction / draft / exercise, or a tag among experiments, style-transfer, impersonation, cross-model, debate, dreams, humor, satire, fiction, roleplay, makes it speculation, and every observation the document produces that would have been declarative is stored speculative instead — so dream never refines a fact out of it. Implemented as one store wrapper around the whole pipeline (withDocProvenance) rather than a parameter threaded through four steps and six write sites. treat_as: 'fact' | 'speculation' on DigestOptions, the digest tool and fozikio digest --treat-as override the decision in either direction.

  The extract prompt now keeps the speaker in the item ("GPT-4o argued that…", "Written in Virgil's voice: …") instead of normalising another voice's claim into a bare statement.

  SQLite gains two nullable columns via the existing addColumn migration; JSON and Firestore stores carry the fields as-is. Existing observations are untouched (source_type null).