luna-vitals-extract

Extract health-series rows for ONE time-bucket into a review artifact. Inputs: a date range
<START>..<END> and the artifact output path <ARTIFACT>. Tracked metrics: migraine, skin_flare,
sleep_hours, hrv_ms, rhr_bpm.

1. Locate health-relevant notes in the bucket: use qmd over collection salus (and 20-Timeline/,
   daily notes) filtered to the date range. Collect the file paths.
2. Deterministic pass: for each structured file, run
   bun run <repo>/scripts/luna-vitals/cli.ts parse <file> --note-date <date-if-daily> --out <tmp/det-N.json>.
3. LLM pass: read the prose/timeline files. Extract ONLY explicit dated values (e.g. "14 Jun 2024 —
   migraine, bad day" → migraine on 2024-06-14; map qualitative→the metric's scale using the registry
   50-Vitals/_series.md). Do NOT infer from relative statements ("worse lately"). Record each as a row
   {metric,date,value,source} where source = ": ". Write these as one artifact
   {bucket:"<START>..<END>", rows:[...], conflicts:[]} to <tmp/llm.json> (same schema as the CLI).
4. Merge this bucket's artifacts, routing the deterministic-parse artifacts and the LLM artifact
   into their respective pools so deterministic values win on overlap:
   bun run <repo>/scripts/luna-vitals/cli.ts merge --det <tmp/det-*.json> --llm <tmp/llm.json> --out <ARTIFACT>.
5. Report the row count per metric and any conflicts. STOP — do not write 50-Vitals/. The operator
   reviews <ARTIFACT>, then the single-writer merge + write step lands it.