Audit follow-up: correctness fixes, usability upgrades, and cleanup

[Ben-K-Jordan]

Implements everything from the second full audit (code bug hunt, leftovers sweep, and a live browser walkthrough of the whole user journey).

Correctness

• Stable objective identity. Plan generators minted objective ids per block (obj_0, obj_1, ...), so plans with more than 6 objectives had colliding ids across blocks, cross-contaminating spaced-repetition mastery and keying feedback anchors to the wrong topic. Ids are now deterministic slugs of the objective title, identical wherever the objective appears.
• Error-repair decks and previews are scoped to the session's course (a CS repair session no longer injects chemistry errors).
• The final-attempt completion path guards its ACTIVE to COMPLETED transition, so racing it against "End Session" can no longer run post-completion effects (SM-2 advance, XP) twice.
• AI-generated prompt difficulty is sanitized to an integer 1..5 before hitting the Int column.

Usability (from the live walkthrough)

• Plan page: loads your existing plan on every visit (delete / reschedule / .ics always available), confirms and replaces same-course plans instead of silently stacking duplicates, and shows per-session focus lines, real dates with a Today marker, an exam-day row, and a "N days until exam" header.
• Session runner: empty materials search persists an explicit no-sources state and the review panel says "Nothing matching in your materials" instead of spinning forever; free-recall prompts reveal the model answer and key points at self-score time; confidence is only recorded when actually rated (calibration no longer reports noise from the untouched default); typed answers and error-log drafts survive a refresh.
• End screen: follow-up recommendations are one deterministic function of the run's results (no more changing across a refresh), and a "Schedule follow-ups" button inserts the recommended spaced sessions into the active plan, idempotently. "Start New Run" renamed to "Practice this session again".
• First run: signup auto-signs-in when email verification is off; a muted "Mock AI" badge shows when no real AI provider is configured; the app has a favicon; the timezone field gains a Detect button with a live clock preview.

Cleanup

• Dropped the orphaned notification-preferences route and model (migration also removes the dead display_name and leaderboard_visible columns; validated with migrate deploy from history plus an empty migrate diff).
• Deleted orphan reminder email templates, practice-set and evidence zod schemas plus their tests, the sw.js header block, @types/web-push, the asset-manifest orphan, and the phantom CRON_SECRET from setup, env example, compose, and README.
• Learn and achievements pages moved off the deprecated identity header; getOrCreateUserId is gone.

Verification

• tsc --noEmit clean
• 646 unit tests passed (36 new)
• 137 integration tests passed (11 new)
• Full e2e after production rebuild: 57 passed, 15 skipped (exact baseline)

---

Generated by Claude Code