Fresh supermarket trainer - the morning-huddle quiz tool: three questions, thirty seconds each, to find out whether last month's training survived

Build me a single-file HTML morning-huddle quiz tool for a
fresh-food supermarket trainer. Opens by double-click, works
offline, projects onto the huddle screen or TV.

Background: operational training evaporates within weeks, and
the morning huddle has five spare minutes. I want a "random
quiz" segment: one question on screen, pick someone to answer,
flip to the answer points, thirty seconds per question. Question
maintenance must be nearly free - edit in a spreadsheet, paste
into the tool.

Question bank:
- Question structure: role (produce / scale / cashier /
  everyone), prompt, answer points (2-4, one short line each),
  optional "common mistake" note
- Two entry paths: single add/edit/delete in the UI, or bulk
  paste from a spreadsheet (tab-separated: role<TAB>prompt<TAB>
  answer points, semicolons between points). Bulk via
  spreadsheet is the main path
- Data in localStorage, plain-text export/import for backup

Quiz mode (the core scene):
- Screen/TV friendly: huge prompt type, one question per view,
  readable from the back row
- Draw rules: filter by role (only produce staff today? draw
  from produce + everyone only); default 3 questions per
  session, adjustable
- Random but not dumb: no repeat within 14 days of last use;
  questions marked "missed" enter a miss pool and get drawn
  first next time
- Flip interaction: prompt first; click (or press space) to
  reveal answer points; then two buttons: "got it" and
  "missed" - "missed" logs the miss and tints the card red
- After each question, show its history: times drawn, times
  missed - making the chronic offenders visible

Miss board:
- Sorted by miss count: prompt / role / times drawn / times
  missed / last miss date
- Export as text - the high-miss questions ARE next month's
  training agenda

Hard constraints:
- Single file, no build, no external CDN or fonts; spacebar
  flips, arrow keys advance - for remote use while projected
- No accounts, no team-battle mode, no points leaderboard -
  the quiz exists for recall, not competition; points invite
  gaming, and gamed data is worthless
- No preloaded sample questions - empty start, first screen
  teaches the paste format
- No "employee mastery reports" - the tool knows how often a
  question was missed, not who missed it (answering is
  anonymous); don't fabricate conclusions the data can't
  support

Why it is written this way

The decay curve of training is brutal: a sliver left after a week, indistinguishable from never-trained after a month. The only cheap countermeasure is high-frequency micro-retrieval, and the huddle is a standing gathering that already exists. Every design choice here serves one goal - driving the cost of quizzing toward zero: bank maintained in a spreadsheet and pasted in, one click to draw, thirty seconds a question. If the ritual costs real effort, it won't survive two weeks.

The two "random but not dumb" rules keep the bank from going stale. The 14-day no-repeat window stops the same questions cycling - employees memorize answers instead of actions. Missed questions jump the queue, so weak spots automatically get the higher frequency. It's the same loop as cycle counting - "the less accurate the item, the more often it's counted" - with memory in place of inventory.

The single "missed" button is deliberately the ONLY data capture. A huddle segment is thirty seconds; recording any extra field (who answered, which shift) would kill the ritual. Anonymity buys data honesty - employees don't guard against it, and the trainer gets a clean miss distribution. The cost is that no per-person mastery analysis is possible, which is exactly why the "mastery report" is banned outright: the data doesn't contain that dimension, so any such report would be fiction.

No points, no battles, no sample questions - same logic as the other tool entries: turn quizzing into a contest and behavior shifts from recalling actions to gaming answers, corrupting the miss data with it. The one asset this tool must protect is that the miss counts are real.