feat(exercises): show students their past AI task results, legibly (#581)

[guillermoscript]

Closes #581.

A student who came back to a finished lesson checkpoint or standalone exercise saw a bare score badge. The AI's feedback was never missing — it has always been written to lesson_checkpoint_attempts.evaluation and exercise_evaluations.ai_result. The loaders just never read those columns back, and every engine held its result in component state that a reload discarded. This was a display gap, not a storage gap.

This PR restores it, then fixes what an /impeccable critique found wrong with how it read.

Walkthrough

Same student, two surfaces: the standalone essay they passed, then the in-lesson checkpoint they did not. 390px.

Restoring the result

Mobile · 390px	Desktop · 1440px

• loadLessonCheckpoints now selects response and evaluation. Defensive parsers turn those two free-form jsonb columns back into the student's own answer plus the feedback, next-step hint and per-question explanations. Four writer shapes exist (ai / deterministic / fallback / external sync); anything unrecognised yields "nothing to show" rather than being trusted into the UI.
• The standalone exercise page needed no new query. Artifact and essay rows were already in the exercise_evaluations result set fetched for the audio/video history, and were simply never passed to their components.
• Per-question feedback rendered correctValue raw, so a multiple-choice miss read "Incorrect — 1" — the option index. It now maps to option text, and booleans to localized true/false.
• Coding challenges write no evaluation row at all. They are graded by their test runner and store a hardcoded completion, so there is nothing to replay. The card says so explicitly instead of rendering an empty result. That is a storage gap and out of scope here.

Making it legible

Verified live at 390×844 and 1440×900, light and dark, /en and /es, on cold loads.

Accessibility. Six WCAG AA text-contrast failures in light and four in dark, measured canvas-sampled over the real composited background. All now clear; 14/14 elements pass in both themes.

Element	Was	Now
PASSED badge, white on emerald-500	2.47	✅
/ 100, text-muted-foreground/60	2.31 light / 3.29 dark	✅
STRENGTHS / AREAS TO IMPROVE	3.65 / 3.20	✅
+ / - list markers	2.47 / 2.13	✅
AI FEEDBACK, text-primary	3.49 dark	✅

The AI Feedback label no longer uses text-primary at all: tenants override that token, so small text in it cannot be guaranteed to clear AA for any school. The icon keeps the accent — non-text UI only needs 3:1.

The artifact panel printed the score over the pass mark, twice. 62 / 70 reads as 89%, but 62 against a pass mark of 70 is a failure. That whole block was a near-copy of ExerciseResultSummary, so it now renders the shared card and the duplicate is gone.

A failed checkpoint showed a green tick, because its tone came from completed and a failed attempt is still completed. And the restored feedback sat behind a closed disclosure, so a returning student saw a score chip and nothing else. It now opens itself for the one state that asks something of them — a graded attempt they did not pass — and an explicit toggle still wins.

The retry field also sat between the answer and the grade for that answer, looking identical to the read-only answer above it. Order is now answer → result → retry, and the restored answer is borderless so it stops reading as an editable field.

A sticky-column regression I introduced. Putting the result card inside lg:sticky lg:top-6 took that container to 1078px against a 900px viewport. position: sticky does not pin an element taller than its scrollport, so the chat input scrolled away unreachable. The card is now a sibling above it (650px, sticks again) — which also puts the result first on mobile instead of 1157px down the page, behind instructions the student has already followed.

Measure and touch. Prose capped at 68ch (was 85–90). Header stacks below sm:, where the metadata line was being squeezed to 161px and two lines. Long unbroken input wraps. On phones the retry button is 36px tall (was 24) and inputs are 16px, so iOS Safari stops zooming the page on focus.

Verification

• npm run build exit 0 · npm run typecheck clean · 669 unit tests pass (53 files, 23 new covering the parsers and the auto-expand rule)
• npx impeccable detect clean on all touched files
• Contrast audited programmatically in-page, both themes, 0 failures
• All QA fixtures removed afterwards; lesson content restored byte-for-byte

Not in scope

The artifact page restores a grade while its own answer fields sit empty — checkpoints show "Your answer", the artifact engine has no equivalent. Noted in #581 as deserving its own issue.