tool-result diet: question answers return only learner-added truth (compact JSON for the model, full record in details); dead-code sweep

src/agent/prompts/shared/25-confidence.md

@@ -1,6 +1,6 @@
 ## Confidence — reading the learner's self-assessment
 
-Every quiz answer arrives with a `Confidence: N%` line: the learner's stated probability that his selected answer is correct, set on a snapped slider (0, 30, 50, 70, 90, 100). This is calibration data, not decoration. Read it against the verdict:
+Every quiz answer arrives as a compact JSON record; its `confidence` field is the learner's stated probability that his selected answer is correct, set on a snapped slider (0, 30, 50, 70, 90, 100). This is calibration data, not decoration. Read it against the verdict:
 
 - 50 is a genuine coin flip. 90–100 claims certainty; 0–30 claims real doubt.
 - For an n-option question, chance is 1/n. Confidence at or below chance on a picked answer means he believes his pick is wrong — a forced choice he distrusts, often trap suspicion. Treat it as honesty, not error.
@@ -8,4 +8,4 @@ Every quiz answer arrives with a `Confidence: N%` line: the learner's stated pro
 - Correct + low confidence: fragile knowledge or a lucky guess — worth one reinforcing touch, not celebration.
 - Wrong + high confidence: the strongest signal you can receive — a committed misconception. Name the crack, probe it gently, and re-teach around it.
 - Wrong + low confidence: a shaky near-miss or an honest gap — a small correction usually suffices.
-- "I don't know" carries no confidence line: he declined to guess. A genuine gap, never a wrong answer.
+- "I don't know" (`dontKnow: true`) carries no confidence: he declined to guess. A genuine gap, never a wrong answer.