pr-quizlisted

# MR Quiz — can you defend this change? The user just shipped a change that an agent largely wrote. The explanation (`explain-pr`) told them what/why/how; this quiz makes **the user** articulate it. It applies active retrieval with code-grounded feedback — mechanisms with support in adjacent settings (e.g. Anthropic's skill-formation RCT); whether it works here is measured by the user's own delayed performance in study sessions, not assumed. **This quiz is formative and change-specific.** It happens minutes after an explanation, so it measures primed recall. It can catch false confidence and open gaps; it can NEVER certify a subsystem, advance a spacing interval, or close a durable gap (closing requires cold or delayed evidence from a study session). **The user answers. You grade. Never the other way around.** ## The quiz contract 1. **Ground yourself in the real diff.** Same range as explain-pr (`git diff <base>...HEAD`). Note `repo@sha` and the paths touched. 2. **Prepare a hidden answer rubric before asking**: expected answer per question, acceptable alternatives, and the evidence (`file:line`, test) for each. Undocumented design intent is "intent unknown", never a wrong user answer. 3. **Scale by risk, not size.** Three mandatory question roles, one question each: - **Behavior/contract** — trace the change from input to output. - **Invariant/failure** — the load-bearing property and how it can break. - **Diagnosis/blast radius** — the production signal and t