feat: anti-gaming — timer, paste blocking, adaptive follow-up

[krishhimself]

Summary

The obvious attack on a comprehension quiz is to paste the question into a chatbot and
paste the answer back. Three measures raise the cost of that, and the flow is
restructured to generate → submit → followup.

Measured result

Same repo (psf/requests), same time budget, same flow:

Profile	Behaviour	Score
Pasted AI answer	committed with 63s of 75s left, could not defend its own wording	0/100
Genuine author	typed distinct answers, defended the follow-up	100/100

The follow-up quotes the candidate's wording back, which is what makes it un-fakeable:

"You stated that 'error states are surfaced through the standard exception mechanism'
without 'leaking implementation details' — when the pool experiences a connection
timeout versus pool exhaustion, what specific exceptions does your 'boundary' raise?"

The genuine run scored 10/10 — "Exceptional defense under pressure" on the follow-up row.

1. Per-question 75s clock

Starts when the question renders. Computed from wall-clock time, not accumulated
ticks — a backgrounded tab can't buy seconds by having its interval throttled. On
expiry the answer commits as-is (blank included) and the field locks; when every
question has expired the quiz submits itself. Time remaining at commit rides along with
each answer.

2. Paste blocked

paste and drop are both prevented. Drag-and-drop text walks straight past a
paste-only guard.

3. Adaptive follow-up, before grading

The answer least likely to have been typed by its author is selected by weighting typing
rate by length — (length/elapsed) × length — so a suspiciously fast essay outranks a
fast one-liner. Falls back to the first answered question, then the first question, so
the round always happens even with no timing data.

Grading is deferred until after this round on purpose. Grading at submit would let a
candidate bank a score and walk away from the round they can't pass.

API change

/submit no longer returns a score — it returns the follow-up question. /followup
returns the final result in the shape /submit previously returned, so ScoreResult
needed no changes. Attempt documents gain answers, followup, and an
awaiting_followup status between generated and graded.

What this does NOT do

The timer and paste block are client-side. They raise the cost of casual cheating;
they do not stop anyone calling the API directly with a forged seconds_left. The
follow-up round is the measure that actually holds, because it demands understanding at
response time regardless of transport. Server-issued timestamps at generation, with
elapsed time computed server-side, are the real fix and are not built here. This is
stated plainly in the README rather than footnoted — overstating it would be a security
claim that isn't true.

Operational note

The follow-up adds a third Gemini call per quiz. On the free tier's 20 requests/day/model
cap that means roughly 6 complete quizzes per day, down from 10.

Verification

pytest passes, vite build clean, documents persist through
generated → awaiting_followup → graded with answers and follow-up recorded, test data
cleared. Suspect-selection heuristic verified against four cases including no-timing-data
and all-blank fallbacks.

Not verified: the interactive browser flow. The Chrome extension would not connect,
so frontend checks were server-rendered. This PR adds the most timing-sensitive UI in the
project — countdowns, auto-submit on expiry, a two-phase form — and none of it has been
exercised by an actual click. Worth a manual pass before merge.

🤖 Generated with Claude Code

https://claude.ai/code/session_01PET9qKZXhgjEbZK7MReYQj