feat: The Mastery Loop — self-graded understanding + library coverage (Pillar 3, Spec 1)

[New1Direction]

Summary

First slice of Pillar 3 — the Knowledge Game. Adds a learn → prove → see it grow loop, research-backed by the "AI-Augmented Textbook"/Learn Your Way paper (embedded assessment), adapted to RepoLens's zero-backend, BYO-key model: self-graded, no new AI calls, fully local.

Built spec-first → plan → subagent-driven execution. Spec: docs/superpowers/specs/2026-06-16-mastery-loop-design.md; plan: docs/superpowers/plans/2026-06-16-mastery-loop.md (both with rendered HTML).

What's in it

• mastery.js (new, pure, fully unit-tested) — deriveCheckResult, aggregateMastery, levels, UNDERSTOOD_THRESHOLD = 2/3 (2-of-3 passes — the exact fraction, not a rounded 0.67).
• store.js + store/idb.js — a new local IDB mastery store (DB v5→6, additive) with CRUD (getMastery / getAllMastery / setMastery), mirroring the decisions store. Round-trip tested with fake-indexeddb.
• Earn (deep-dive panel) — the old static "Test Yourself" block is now an interactive flip-card check: reveal → Got it / Shaky / Missed → auto-advance. On completion it computes the result, shows the level earned + Glows & Grows (never a raw %), and persists. Partial completion or a zero-question deep dive write nothing.
• See (library) — per-card level indicator (○ new / ◐ explored / ● understood, label on hover), an honest aggregate line, and a single-select level filter.

Reuse / invariants

• Earned from data already produced (the deep-dive Feynman questions) — no new AI calls, no background.js / message-contract changes (grep-confirmed in the final review). Mastery is local IDB only; BYO-key untouched.
• mastery.js is pure → all scoring/leveling/aggregation is unit-tested; persistence tested with fake-indexeddb; the store stays CRUD-only (no compute).

Test plan

• npx vitest run → 875 passed (93 files; +9 mastery.test.js, +3 store-mastery.test.js, +1 library-filters case)
• npx eslint . → 0 errors
• npm run check:html → 24 files clean
• node --check on the 7 touched JS files
• Final whole-branch review: readyToMerge, verified the loop end-to-end, the 2/3 boundary, partial/zero-question no-write, listener lifecycle, reduced-motion, and the onboarding-demo guard (the demo repo has no Feynman block, so it can't create a mastery record)
• Manual smoke (no DOM test env): run a deep dive → take the check → confirm level + Glows/Grows + "Saved"; close mid-way → nothing saved; library card shows the glyph + the aggregate + filter work

One choice to confirm

The aggregate line's denominator (Understood X of Y) is currently Y = repos you've checked (so X understood + N explored = Y), which is self-consistent and only appears once you've taken a check. If you'd prefer the coverage framing (Y = total library size, "12 of your 40 repos"), it's a one-line change — say the word.

Out of scope (per spec → future Pillar 3 work)

AI-graded MCQ, spaced-repetition resurfacing, the corkboard knowledge-graph (Spec 2), and mastery in the backup envelope.