Quiz deck from records

The quiz asks only about what a record says. You choose the facts; the script
refuses any fact whose answer is not in words the record actually contains.

Steps

1. Read the records with the tools you already have (masc_board_post_get,
   keeper_memory_search, gh issue view). Save each one as a text file under
   one directory, exactly as the tool returned it. Do not retype or summarise.
2. Write candidates.json: a list of
   {"id","field","subject","answer","quote","record":{"uri","file"}}.
   • field must be one of author, claimant, status, merged_commit,
     cause, decision (the questioner has no template for anything else).
   • quote is copied from the saved file; answer is copied from quote.
   • uri names the record (e.g. masc://board/<post>#<comment>), file is the saved text.
   • Give each field at least two facts with different answers, or no question is asked.
3. Run python3 scripts/build_deck.py candidates.json <record-dir> <deck-id> deck.json.
   Exit 1 lists rejected facts; fix the quote or drop the fact, never the check.
4. Point both installations' snapshot_file at the same deck.json with the same
   source_id (deck). A new deck-id makes old questions ungradable on purpose.

What this does not prove

The record hash is of your saved copy, not of the Board's stored bytes. The
check establishes that the answer is a whole word of the quote and the quote is
in the saved record. It does not catch a negation elsewhere in the quote
("원인은 unmerged 가 아니라 merged"), and it does not show you picked the
important facts. Quiz scores split by the answerer's own label, not by the
authenticated requester.