[nicoloboschi]

Fixes #4567.

Problem

On a low/mid reflect, fresh mental models release the forced lower layers (the short-circuit in agent.py). The model could then answer "the bank holds nothing about X" off the page layer alone, while recall on the same bank returned the facts. This hits the coding-agents session-start reflect, which always runs low.

Prompt-level fixes did not move it: #5250 and a bank directive were both measured on the reporter's banks (0/10 and 3/10 short-circuited reflects reached recall).

Fix

done gets a required boolean, answer_found. If the model sets it to false and recall has not run yet, the loop:

1. refuses that done once, telling the model to run recall with the question's key terms,
2. forces recall on the next turn,
3. accepts the next done.

The flag is checked in code, not left to the model. A missing or unreadable field counts as found, so models that ignore it behave as before. high budget and answers that do find something are unchanged.

Known limit: a confident, partial answer marked answer_found: true is not caught.

Tests

• Loop: short-circuit → done(answer_found=false) → refused → forced recall → accepted. And a negative after recall is accepted with no extra turn.
• Lenient parsing ("false" strings, missing field, _raw payload).
• Every done variant (plain, directives, document) requires the field.
• hs_llm_core: stages the early negative, then the real model writes the recall query and answers from the raw fact (judged). Passed 3/3 locally on gemini-3.1-flash-lite, plus the existing short-circuit real-LLM tests 3/3.

Also checked by hand on gemini-3.1-flash-lite (9 runs): the model sets answer_found true when it answers and false when nothing exists.