akcli-datasheet-factslisted

# akcli-datasheet-facts — audited facts extraction, the agent half of the loop You (the agent) are the extractor: you can read PDFs natively, and the facts store is designed around that. akcli owns the *discipline* — every fact is pinned to its source PDF by sha256 + page (+ verbatim quote), and `review facts verify` audits what you wrote. Your job is to fill the store honestly; akcli's job is to catch you if you don't. **The prize:** a heuristic finding ("verify Vref against the datasheet") becomes a `datasheet_backed` judgement with the exact page as evidence — and `review propose` can then emit an auto-applicable fix. No fact, no claim. ## Extraction doctrine (non-negotiable) - **Read the actual PDF.** Never fill a fact from memory, a search snippet, or a "typical value" you believe. If you cannot open the PDF, stop and say so. - **Page-accurate, quote-verbatim.** The `page` is where the number IS, not where the section starts. A `quote` is copied character-for-character from that page — `verify` runs a text match against the PDF (when `pdftotext` is installed) and a mismatch is a WARNING with your name on it. - **min/typ/max as printed.** Record the row as the table gives it; do not average, do not pick a bound and call it `value`. Conditions (temperature, load) go in `conditions`. - **Absence is honest.** A part whose datasheet does not state a number gets NO fact for it. Detectors fall back to their heuristics — never to a guess. - **One MPN, one file.