The user requested remediation of every open issue and a browser review of the released student experience against the mockups. The baseline and individual claims are summarized in PR #98.

• Use the mockup type scale throughout the application (#96). Keep text-entry controls at 16px on phones to avoid Safari's automatic input zoom, and make primary student touch targets at least 44px high.
• The initial pass preserved the student layouts. The user subsequently approved the independent redesign in 92-student-ux.md, which supersedes that choice. The flag belongs beside the prompt; it must not narrow the answer options. Keyboard navigation starts each question at the top and works after clicking an answer. Text-entry controls retain their normal editing keys. The final next-arrow opens the submission review.
• The user approved #82: retain integrity events for 13 months; retain the audit log; anonymize students only on a manual request; do not automatically erase disabled accounts. Production deletion is a separate, explicit maintenance operation.
• The user has no Neon CLI access or alerting service configured. Prepare reproducible operational tooling and documentation for #79/#81, and report external configuration or production verification still required.

• Student keyboard, typed answers, save/submit races, exit and reload safety.
• Browser review of student home, classes, intro, paper, review and results at phone and desktop breakpoints, including narrow desktop layouts.
• #85 publish diagnostics and section/question jumps.
• #90 dashboard data and #93 roster/count parity.
• #95 remaining UI parity items and #96 shared typography.
• #78 headers, request limits and deployment checks.
• #80 mockup checks in CI.
• #82 maintenance tooling and retention documentation.
• #79 local restore drill/tooling and #81 opt-in monitor. Neon production PITR and delivered notifications remain pending external setup.
• Full local tests, build, lint and generated contract verification; PR #98 repeats the full checks in CI. Evidence and remaining production acceptance are recorded in PR #98 and ../setup/operations.md.

• Clicking a radio focuses an input, which the shortcut guard treats as a text field. Arrow keys then change its answer instead of navigating.
• The fill-blank Markdown renderer declares a new component during each render. Every keystroke replaces the input and loses focus; only the first character survives normal typing.
• Submitting while an autosave is outstanding does not wait for that save. A failed final save still allows submission. An old response can also clear dirty answers after switching attempts. Regression tests cover these races.

Run checks sequentially on this 14.5 GiB machine with no swap. Web tests use one worker; Go uses GOMAXPROCS=2 and -p=1. A local runner refuses to start below 2.5 GiB available and terminates its own process group below 2 GiB. Do not stop unrelated applications. ESLint needs more than a 512 MiB heap; its isolated ceiling is 1 GiB with the same host-memory guard.

_{Generated from docs/plan/91-issue-remediation.md. Edit that file in the repository; changes made on this page are overwritten.}