fix(questions): close the paid-LLM abuse gap in question generation

[tanvishdesai]

Summary

POST /questions/generate was reachable by any teacher-role account with an arbitrary transcript string — no roomId, no dedicated rate limit, and no length caps — turning a "generate quiz questions" endpoint into a de facto free-form LLM proxy on the org's paid keys. The transcript was also spliced verbatim into the prompt with no framing against prompt injection, and the model's returned question/option/explanation text had no length cap.

• Require roomId and verify the caller owns that room (checkRoomOwnership), tying generation to an actual lesson instead of any transcript from any teacher account.
• Add a dedicated per-user rate limiter on /generate (40/15min, Redis-backed when available), separate from the general apiLimiter which is sized for high-volume student polling, not for gating an expensive external API call.
• Cap transcript input length (MAX_TRANSCRIPT_CHARS, env-overridable) and fence it with explicit <<<BEGIN/END SESSION CONTENT>>> delimiters plus an instruction to treat it as inert data, not commands — narrows the prompt-injection surface without changing legitimate generation quality.
• Truncate returned question/option/explanation text so a hijacked or malfunctioning response can't balloon into unbounded stored/rendered text.

Intentionally out of scope: the student→teacher self-role-escalation via PUT /auth/role — left as-is per explicit instruction.

Test plan

• npx jest — all 124 backend tests pass (6 new, covering the length guards, injection framing, and the roomId requirement)
• vite build — frontend builds clean
• Manual smoke test: generate questions from the room UI as a real teacher and confirm the existing flow still works end-to-end

🤖 Generated with Claude Code