Hatch handoff

Read AGENTS.md (conventions, commands) first, then content/AUTHORING.md before touching lessons.

State (all green at the time of writing)

Curriculum complete: 18 units, 266 lessons (u01-u18), strict npm run content passes, check-content has 0 errors / 0 warnings, verify-content runs every solution and invalid sample through the real checker with 0 problems.
App: lessons workspace (bento: lesson / CodeMirror editor / terminal), multi-file projects with file tabs, review sessions, playground, mastery map, achievements, settings, XP / levels / streaks, persistence with backups, mobile tabs, light and dark themes.
Quality gates: npm run typecheck, npm run lint, npm test (55 Vitest tests) all pass. Playwright: lesson-flow, features, and walkthrough (plays lessons with their reference solutions; multi-file aware).
Walkthroughs (final build, after the last app change): every lesson of all 18 units was played through the real UI (Chromium, JSPI available) with its reference solution and passed: u01-u09 134/134, u10-u11 + u17-u18 47/47, u12 8/8, u13-u16 77/77 (all 266). The lesson-flow and features specs (10 tests) pass too.

Area	Path
Python runner, checker, AST features, error shaping (single implementation)	`python/hatch/`
Worker that hosts Pyodide (module worker, Atomics-blocking `input()`)	`public/hatch-worker.js`, `src/runtime/`
Curriculum schema, graph, validators, execution verifier	`src/curriculum/`
Curriculum data	`content/` (spec: `content/AUTHORING.md`)
Pure engine: XP ledger, mastery, review, achievements	`src/engine/` (tested in `tests/`)
Store and persistence (`hatch.v1`, `hatch.drafts.v1`)	`src/store/`
Lesson/review orchestrator	`src/components/lesson/workspace.tsx`, `step-views.tsx`

Blue-accent, larger, sturdier UI (light = sky tint with white cards, dark = deep navy). Tokens live at the top of src/app/globals.css: --accent (solid blue for buttons and text, 5.4:1 on white), --accent-bright (#1cb0f6, fills only: progress bars, active step), --xp-bright (gold fills), green only for success. Nunito (rounded) replaced Geist/Fraunces; JetBrains Mono for code.
Root font size is 106.25% (mobile) and 112.5% (>= 768px); Tailwind text-xs/sm/base were enlarged in @theme and the old text-[0.6-0.7rem] micro sizes were bumped to >= 0.8rem. Buttons are chunky (2px border, 4px bottom edge, 3rem tall), panels have 2px borders and a flat ledge shadow, progress bars are 0.75-1rem thick. Default editor font size is now 17 (existing saved preferences keep their value).
Contrast for the small coloured labels was checked numerically (all >= 4.5:1 in light, >= 5.8:1 in dark). The Duolingo-style reference tokens were adapted, not copied: white text on the bright #1cb0f6 would fail contrast, so solid buttons use the darker #0a6eb4.

Checks are behavioural. AST rules apply to the entry file only; for multi-file projects the u11/u18 lessons use hidden fixtures (probe.py, mutants.py, h.py) that run the learner's own tests against broken implementations or parse the modules with ast.
Async lessons (u17) check the stdout of the learner's complete program. Pages in the fetcher project are simulated (SITE dict + asyncio.sleep): the lesson says so. Delays differ by at least 30 ms to keep the output order deterministic.
"Check" button = run the hidden check only; "Run" = run in the terminal, then check (also when the program errors, so steps that expect an error can pass).
Preferences: the old sound toggle and the "Full motion" option were removed because they did nothing. Old saved values are migrated by the zod schema (persistence.ts, covered by a test).
Lint ignores the vendored public/pyodide/** and src/generated/**.

Saved progress is a JSON document (hatch-progress, schemaVersion in src/engine/progress.ts). Reading goes through zod in src/store/persistence.ts; to change the shape, bump SCHEMA_VERSION and add a case to migrate(). A document from a newer app is never overwritten (the learner sees a notice), and a damaged one is copied to a backup key before anything is reset.
Every lesson has an integer version (YAML header, default 1) and progress records the version a learner completed. Bump it when a lesson changes in a way that should not silently count as the old one; course.version in content/course.yaml is the course-level counterpart.
Drafts (hatch.drafts.v1) are per step, pruned oldest first when over budget.

Browsers without JSPI (non-Chromium): live runs work through a rewrite of module-level asyncio.run, but hidden checks that call asyncio.run are not supported there. u17 only relies on stdout checks, but this has not been exercised in Firefox/Safari.
Walkthrough runs: HATCH_E2E_UNITS=u10,u11 npx playwright test walkthrough (about 3 minutes per 25 lessons). Make sure no stale next dev is on :2034 first (a hung one makes every test time out in page.goto). A failing lesson usually means a solution that passes the Node checker but behaves differently in Pyodide. Two real examples found this way: JSPI was never enabled because can_run_sync() was probed outside callPromising, and under JSPI asyncio.run turned 4.0 into 4 (fixed by boxing the result in bridge.fix_asyncio_run); and checklist ticks leaked between checklists of one lesson because they shared the scratchpad key.
Curriculum review: content was validated mechanically (prerequisite leaks, terminology, feature ownership, XP). A human read-through for tone, pacing and difficulty spikes, especially units 12-17, is still the best next investment.
Reasonable future features, none started: export/import of progress as a file, per-unit "skip test" placement quiz (deliberately absent: the app picks the order), keyboard shortcut cheat-sheet.

test-results/ and playwright-report/ are generated and safe to delete.
The repository has a single "Initial commit" and nothing committed since; no remote is configured.