PvZ exam recollection

Later Dave could miss the player's recorded 2-3 grade when asked about “your little quiz” or “my score”: the memory selector recognized only a narrower set of exam words and selected unrelated recent chapters instead. It now recognizes those questions and corresponding Chinese/Spanish wording. Unrelated combat conversation still uses recent context instead of forcing an exam callback.

A second regression uses a valid observed paper with no accompanying event list. Its marks survived notebook serialization, but three later-history filters discarded the result. The source-to-source and source-to-classic prompt paths now retain that recorded paper within the existing context budget. A recorded shopping bill likewise survives the source-to-source filter.

The grade remains an observation of the displayed paper. It does not prove a reward was collected, authorize another reward, reveal chosen answers or carry across an Adventure cycle. Restarts and rewound attempts retain their existing invalidation rules. There is no new UI, pause or provider call.

Verification on current production base 24bc4f065:

• Both new regressions fail on the prior code: missing grade and wrong chapters.
• All 472 NPC/route tests pass using the isolated local test launcher.
• The paraphrase regression verifies the final model prompt contains the actual 60/100 fixture grade in English, Chinese and Spanish. It also checks that unrelated combat chatter does not select the older paper.
• pnpm build and pnpm typecheck pass.

These checks establish which observed facts reach the model, not a guarantee that every provider will produce polished dialogue. A separate normal-input 2-3 campaign journey is exercising the native award/save handoff.