[tenphi]

What

maintenance.retain.stop_words: names and terms the owner never wants retained.

{ "maintenance": { "retain": { "stop_words": ["Vulpine Mutual"] } } }

• Extraction is told to propose nothing about them, as data only.
• A deterministic check then holds every candidate that still names one, together with everything that depends on it through a relation. It looks at the text, subject, page, destination, exact support and speaker, and runs before verification, so a held candidate costs no verifier call.
• The hold has a new typed reason, stop_word (RetainHoldReason), at hold_stage: "validation", and writes nothing.
• The rule applies to extracted and supplied (provided) candidates, remember, and folder timeline curation.

Matching is whole-word and case-insensitive. Whitespace in a term also matches - and _, so Vulpine Mutual matches companies/vulpine-mutual but not VulpineMutual. The discourse frame is not checked, so a name in a quoted thread does not stop an unrelated fact. The speaker is checked because "nothing from them" is the usual reason for a stop word, and a fact an email states rarely names its sender.

Tests

• packages/core/src/write/retain-stop-words.test.ts: the matcher, the fields that count, the transitive relation cascade, frame exclusion, and that runRetain tells the extractor and never sends a stopped candidate to the verifier.
• packages/core/test/retain.test.ts: provided mode holds without writing or calling a model and replays. Extract mode keeps the clean candidate; the verifier stub would refuse everything if the stopped one reached it. remember honours the list too.

Locally, build, typecheck, lint, knip and check:repository pass. The full suite has 5067 passing; the 4 failures are ingest/OCR timeouts that fail the same way on main in this sandbox.

Changeset: minor for protocol and core, because a new hold reason widens the protocol enum.

🤖 Generated with Claude Code