llm.tagger: cosine shortlist of the vocabulary's theme cards against a
comment's already-indexed chunks (no re-embedding — text and vector come
from langchain_pg_embedding), one yes/no judgement per candidate on the
largest live host at temperature 0 with the scoring chunk as evidence,
up to max_tags accepted with the shortlist similarity as confidence.
State on the bib item (extra_json.llm_tags: vocab version, model,
content hash, tags with evidence chunk, the full shortlist with
verdicts) makes runs resumable; apply_tags replaces exactly the item's
llm:* tags and never touches hand tags. Store.merge_extra merges
top-level keys into extra_json. Every collaborator is injected so the
pure pieces are tested without Ollama or pgvector.
stack llm tag --docket D [--limit N] [--force] [--dry-run] [--top]
[--max-tags] [--min-confidence] [--verbose]; _tagging_runtime wires the
store, pgvector loader, card vectors, judge and model.
65f8d785a1