feat(quiz): immediate per-question feedback + exam-mode toggle

[jonasbrami]

What

The first slice of the feedback loop (roadmap #2). In the new default practice mode, committing a deterministic question (mcq / tf / mermaid) reveals it immediately — correct/incorrect, the right answer, and the explanation — and locks it. The verdict is computed client-side (the answer already ships in /state), reusing the existing results card, so there's no server round-trip.

A sidebar Mode toggle ("Reveal answers as I go") flips between practice and the classic batch/exam flow, sticky per browser via localStorage.

Production-over-recognition: seeing why you were wrong at the point of the mistake beats a scorecard at the end.

Behaviour

• Practice (default) — deterministic question reveals + locks on first commit; open questions stay interactive (no client-side answer to reveal — they grade at submit, as today).
• Exam — answering view behaves exactly as before; toggling preserves existing selections; submit-bar message reflects the active mode.

Scope

• Frontend only — no backend, schema, or endpoint change.
• Confidence rating (feat(cli)!: collapse to a single quizz take command #3) and teach-back (feat(engine): layered mermaid validator + debug logs #6) — the rest of the feedback loop — follow as their own PRs.

Invariants preserved

• textContent-only frontend (reveal reuses the existing result-card renderer).
• No new browser↔host channel; nothing posted until submit/publish (reveal is purely local).
• Backward-compatible: exam mode is the prior behaviour byte-for-byte.

Testing

• node --check on the bundle; full mcp/engine suite green (no Python touched).
• Real-browser check (Chrome DevTools): correct reveal (green + explanation) and incorrect reveal ("your pick" vs "correct answer" + explanation) for mcq/tf; open question stays interactive; the exam-mode toggle round-trips without losing answers and updates the submit-bar message.

🤖 Generated with Claude Code