phase: 02_level_content_and_validation

Phase `02_level_content_and_validation` — final verification **green**.

- **Gate:** `bash .agents/validate.sh` → ruff ✅ · pyright 0 errors ✅ · `undercooked-levels validate` OK (2 levels) ✅ · **668 passed** (incl. 10 Playwright e2e) ✅ · coverage **100%** ✅ · repo hygiene ✅
- **Criteria:** `validate` exit 0, one line per level; `list` = `01_first_kitchen`, `02_double_shift` in campaign order; live `:8000` `/healthz` 200, `/api/levels` (+detail) carries both levels with numeric `target_score`/`time_limit` and no grid/station leakage, `/api/config` correct, unknown id → 404, startup log `content loaded: 2 level(s), 4 recipe(s)`; `validate --dir tests/fixtures/levels/_bad/` → exit 1 with named `RECIPE_STATION_MISSING` / `SERVE_REACHABLE` / `TARGET_IMPLAUSIBLE`; A8 lock test proves a YAML-only level reaches loader+API+CLI.
- **Defect fixed (pre-existing path, phase-02-triggered):** `undercooked` with unusable content printed the report then a raw traceback and exited `1`. `src/undercooked/__main__.py` now catches `ContentError` → one closing line, exit `2` (like a bad setting); 2 new tests in `tests/unit/test_main.py` (incl. non-content failures still propagate).
- **Docs pass (full, project-wide):** README (content gate + exit `2`, config section), AGENTS.md §3, `docs/LEVELS.md` (startup refusal + exact CLI exit codes), `docs/DEPLOYMENT.md` (crash-loop diagnosis bullet); DESIGN/TESTING re-read, consistent.
- **Container re-verified (no phase-01 regression):** `podman build -f Containerfile` succeeds; the image ships `levels/` and serves `/healthz` 200, `/api/levels` (2 levels), `/` 200 as non-root.
- **Deviation to note:** the phase overview names `tests/fixtures/levels/bad/`; the actual broken-fixture dir is `tests/fixtures/levels/_bad/` (the `_` prefix is what keeps it out of normal discovery). Overview files are read-only to me — README/LEVELS.md document the real path.
- No git index/history commands run; all changes left unstaged for the harness. ntfy notification sent.

AGENTS.md

@@ -16,9 +16,10 @@
 - Execution is owned by the `phased-execution` skill (`run-task.sh`, `run-phase.sh`,
   `auto-phase.sh`). Each task runs in a fresh subprocess and must pass
   `.agents/validate.sh` before it moves to `complete/`. That gate fails fast in this
-  order: `ruff check .` → `pyright` → `undercooked-levels validate` (skipped with a
+  order: `ruff check .` → `pyright` → `undercooked-levels validate` (unconditional:
-  notice until phase 02 creates the loader) → the **full** `pytest` suite with the
+  the same loader + PLAN §8 rules `create_app` runs at startup, over
-  coverage floor → repo-hygiene checks. It runs from the repo root or from `.agents/`.
+  `$UNDERCOOKED_LEVELS_DIR`) → the **full** `pytest` suite with the coverage floor →
+  repo-hygiene checks. It runs from the repo root or from `.agents/`.
 - New work is authored by the `phase-authoring` skill as a phase directory — not
   implemented straight from chat.
 - **Never** modify `.agents/PLAN.md` or anything under `.agents/phases/complete/`.
@@ -37,6 +38,17 @@
   deviation from the default FastAPI stack.
 - **Data-driven content.** Levels/recipes/hazards come from YAML in `levels/` loaded
   through validated Pydantic models. No hard-coded level or recipe lists in Python.
+  `create_app()` loads the bundle once (`routes_content.load_validated_content`) and
+  **refuses to start** when content fails to load or violates a PLAN §8 rule — the
+  named-rule report goes to stderr, `ContentError` propagates and the `undercooked`
+  command exits `2` like it does for an impossible setting. The content API
+  projects through `extra="forbid"` response models: grids, station ids and spawn
+  coordinates never leave the server. `undercooked-levels` (`tools/levels.py`, its
+  author CLI) obeys the same rule: it prints whatever the loaded directory holds.
+  `tests/unit/test_content_not_hardcoded.py` is the lock — it adds a level YAML to a
+  copy of `levels/` and expects the loader, `/api/levels` and the CLI to carry it with
+  no Python edited, and fails if any `src/undercooked/**.py` or browser asset names a
+  shipped level or recipe id as a literal.
 - **Client: Canvas 2D + vanilla ES modules + this repo's own SVG art.** No bundler,
   no framework, no CDN, no analytics, no external fonts.
 - **Tests: pytest + Playwright.** One Playwright suite per user story; multi-player
@@ -52,7 +64,8 @@
 - [ ] **Full** project suite passes (`uv run pytest`), coverage gate met — including
       failures the task did not cause: fix them, do not skip, xfail, or delete tests.
 - [ ] `uv run ruff check .` and `uv run pyright` report zero errors.
-- [ ] `uv run undercooked-levels validate` passes if any level/content file changed.
+- [ ] `uv run undercooked-levels validate` passes — `.agents/validate.sh` runs it for
+      every task, whether or not any content file changed.
 - [ ] `README.md` and `AGENTS.md` updated when the change is user-visible
       (setup, commands, config, protocol, features, conventions). Every phase's final
       pass re-reviews both **in full, project-wide** and fixes inconsistencies,
@@ -115,8 +128,10 @@
 - **Protocol changes need a version bump**: editing a wire message means editing
   `src/undercooked/server/protocol.py`, `docs/PROTOCOL.md`, and the client in the same task,
   and bumping `PROTOCOL_VERSION` (PLAN §11).
-- **Determinism is a contract.** `sim/step.py` must produce identical state given the
+- **Determinism is a contract.** `sim/step.py` (phase 03) must produce identical state
-  same seed and input script — the replay tests in `tests/sim/` are the guard.
+  given the same seed and input script — the replay tests in `tests/sim/` are the guard
+  once the tick exists; until then `tests/sim/test_purity.py` holds the A5 boundary that
+  determinism depends on.
 - **ntfy is outbound-only.** Never use it to ask questions or wait for approval.
   If a task is ambiguous, make the best judgment consistent with PLAN.md, or stop
   and let the user rewrite the task.
@@ -134,9 +149,12 @@
 | Configuration (env vars → typed `Settings`), debugpy bootstrap | `src/undercooked/config.py`, `.env.example`, `src/undercooked/debug.py` |
 | Pure game rules (movement, stations, orders, scoring, hazards) | `src/undercooked/sim/` |
 | Level/recipe content | `levels/*.yaml` |
+| Content schema, ASCII-grid parser, content loader, playability validator (contract: `docs/LEVELS.md`) | `src/undercooked/sim/recipe.py`, `level.py`, `content.py`, `validation.py` |
 | Rooms, lobby lifecycle, session records | `src/undercooked/server/rooms.py`, `room.py` |
 | Wire messages + validation | `src/undercooked/server/protocol.py`, `docs/PROTOCOL.md` |
-| HTTP app (`/`, `/healthz`, `/static`), lifespan | `src/undercooked/server/app.py` |
+| HTTP app (`/`, `/healthz`, `/static`), lifespan, startup content gate | `src/undercooked/server/app.py` |
+| Content HTTP API (`/api/levels`, `/api/levels/{id}`, `/api/config`) + strict response models | `src/undercooked/server/routes_content.py` |
+| Author CLI `undercooked-levels` (`validate`, `list`, `show`, `new`, `axes`); workflow in `docs/LEVELS.md` | `src/undercooked/tools/levels.py` |
 | Security headers (CSP `default-src 'self'`, no-sniff, frame-ancestors) | `src/undercooked/server/headers.py` |
 | CLI entry point (`undercooked`): flags → `--help` → serve | `src/undercooked/__main__.py` |
 | Renderer, HUD, input, interpolation | `app/static/js/`, `app/static/assets/` |
@@ -146,8 +164,12 @@
 
 Tests mirror that: `tests/unit/` (pure, no-I/O logic incl. HTTP-route unit tests),
 `tests/integration/` (in-process app boot, live uvicorn + WebSocket clients), `tests/sim/`
-(headless determinism and replays), `tests/e2e/` (Playwright, named `test_usNN_*.py`,
+(the pure kernel without a server: grid geometry, entity model, the A5 purity guard,
-marked `e2e`). `tests/` is not a package; shared fixtures live in `tests/conftest.py`.
+the level validator — headless determinism and replays join it with the tick in
+phase 03), `tests/e2e/` (Playwright, named `test_usNN_*.py`, marked `e2e`). `tests/` is
+not a package; shared fixtures live in `tests/conftest.py`, and the autouse fixture there
+points `UNDERCOOKED_LEVELS_DIR` at `tests/fixtures/levels/` so no test silently reads the
+shipped `levels/`.
 
 - The browser suites use the **`playwright` sync API directly with fixtures written in
   `tests/e2e/conftest.py`** — `pytest-playwright` is deliberately not a dependency, so

README.md

@@ -10,21 +10,34 @@ email, no password: you create a room, share the 5-character code, play.
 
 ## Status
 
-Phase 01 (runnable service scaffold) is **complete**; the next pending phase is
+Phases **01** (runnable service scaffold) and **02** (level content, schema and
-`02_level_content_and_validation`. What runs today: the Python toolchain is locked
+validation) are **complete**; the next pending phase is
-(`uv.lock`), the pytest harness covers unit, integration and Playwright e2e layers,
+`03_simulation_movement_and_interaction`. What runs today:
-configuration is a single typed, validated `Settings` object (`DEBUGPY` support
+
-included), and `uv run undercooked` starts a FastAPI server that serves the client
+- **The service.** The Python toolchain is locked (`uv.lock`), configuration is a
-shell at `/`, the browser assets from `app/static/`, and `GET /healthz` — which answers
+  single typed, validated `Settings` object (`DEBUGPY` support included), and
-`{"status":"ok","version":"0.1.0","rooms":0}` (`rooms` stays `0` until phase 05 owns a
+  `uv run undercooked` starts a FastAPI server that serves the client shell at `/`,
-supervisor) — with the security headers (CSP `default-src 'self'`) on every response.
+  the browser assets from `app/static/`, and `GET /healthz` — which answers
-The rules are tests and not promises: the no-external-asset rule is
+  `{"status":"ok","version":"0.1.0","rooms":0}` (`rooms` stays `0` until phase 05 owns
-`tests/unit/test_no_external_assets.py`, the documentation set is guarded by
+  a supervisor) — with the security headers (CSP `default-src 'self'`) on every
-`tests/unit/test_docs_index.py`, `bash .agents/validate.sh` is the gate the phase
+  response.
-pipeline runs after every task, and `docs/` holds the DESIGN / TESTING / DEPLOYMENT
+- **The content.** The recipe book and the first two campaign levels live in
-documents. Lobby buttons answer "Not implemented until phase 05.": levels and their
+  `levels/`; `sim/level.py` parses them into typed models, `sim/validation.py` runs
-validator land in phase 02, the simulation in 03-04, rooms and the WebSocket protocol
+  PLAN §8's named rules, and `create_app` loads + validates that bundle at startup and
-in 05-06, the playable renderer in 07. Execute phases with the `phased-execution` skill
+  **refuses to boot** broken content — the named-rule report goes to stderr and
+  `undercooked` exits `2` instead of serving a half-working kitchen. `GET /api/levels`,
+  `GET /api/levels/{id}` and `GET /api/config` answer from it, and
+  `undercooked-levels` gives authors the same gate from the command line
+  (`docs/LEVELS.md` is the schema reference).
+- **The rules are tests, not promises.** The no-external-asset rule is
+  `tests/unit/test_no_external_assets.py`, "no level or recipe id in Python" is
+  `tests/unit/test_content_not_hardcoded.py`, the documentation set is guarded by
+  `tests/unit/test_docs_index.py`, and `bash .agents/validate.sh` is the gate the
+  phase pipeline runs after every task.
+
+What does not exist yet: the simulation (phases 03-04), rooms and the WebSocket
+protocol (05-06 — the lobby buttons answer "Not implemented until phase 05.") and the
+playable Canvas renderer (07). Execute phases with the `phased-execution` skill
 (`auto-phase.sh`); it always takes the lowest-numbered directory still in
 `.agents/phases/todo/`.
 
@@ -56,14 +69,16 @@ uv run undercooked                 # http://127.0.0.1:8000 (shell page + /health
 uv run undercooked --host 0.0.0.0 --port 8000   # e.g. to reach it from a phone
 uv run undercooked --reload        # dev reload
 uv run undercooked --levels-dir levels          # point at another content dir
+curl -fsS http://127.0.0.1:8000/api/levels      # the campaign cards the lobby will show
+curl -fsS http://127.0.0.1:8000/api/config      # client-safe settings + protocol version
 ```
 
 `python -m undercooked` is the same command (the container uses it). A flag
 always beats the environment, which beats the built-in default. There is no
 `/docs` or `/openapi.json`: FastAPI renders those pages from a public CDN, which
-this project forbids. `undercooked-levels validate` (check every level file loads
+this project forbids. The validation the server runs at startup (every level
-and is solvable) is declared in `pyproject.toml` but becomes runnable in phase 02,
+file loads, and no rule of PLAN §8 is broken) is a command too —
-when the content loader lands.
+`undercooked-levels`, see **Authoring content** below.
 
 Debug the server with a real debugger attached (dev only - the listener binds
 `0.0.0.0`, so never enable it on a public host):
@@ -72,6 +87,27 @@ Debug the server with a real debugger attached (dev only - the listener binds
 DEBUGPY=1 uv run undercooked       # opens a debugpy listener on :5678, then serves as usual
 ```
 
+Authoring content:
+
+```bash
+uv run undercooked-levels list                        # the campaign, in campaign order
+uv run undercooked-levels show 01_first_kitchen --grid  # resolved level + reachability overlay
+uv run undercooked-levels validate                    # the content gate; --json for scripts
+uv run undercooked-levels new 03_riverboat            # scaffold a level from the recipe book
+uv run undercooked-levels axes                        # escalation-axis table (phase 11's curve)
+```
+
+`undercooked-levels` is a thin shell over the loader and the PLAN §8 rules
+`create_app` runs at startup, so a green `validate` is a server that boots. It
+finds content through the settings (`--dir` > `UNDERCOOKED_LEVELS_DIR` >
+`levels`, plus `--recipes` > `UNDERCOOKED_RECIPES_PATH` > `<dir>/recipes.yaml`)
+and never from the current directory, so it works from any path. Exit codes:
+`0` ok, `1` a level breaks a rule, `2` nothing usable to work with (load
+failure, unknown id, refused overwrite, usage error). The schema, every field,
+the overlay trick and a "which rule did I just hit" table live in
+[docs/LEVELS.md](docs/LEVELS.md); `tests/unit/test_content_not_hardcoded.py`
+is what keeps levels and recipes out of the Python.
+
 Run the checks:
 
 ```bash
@@ -91,8 +127,9 @@ running first, and no test ever listens on a routable interface.
 
 `bash .agents/validate.sh` is the gate a phase task must pass before it moves to
 `.agents/phases/complete/`. It stops at the first failure, in this order:
-`ruff check .`, `pyright`, `undercooked-levels validate` (skipped with a notice until
+`ruff check .`, `pyright`, `undercooked-levels validate` (unconditional: the same
-phase 02 creates the loader), the **full** `pytest` suite with the 90 % coverage floor,
+loader and PLAN §8 rules `create_app` runs at startup, so a campaign that cannot
+validate cannot commit), the **full** `pytest` suite with the 90 % coverage floor,
 then repo-hygiene checks (nothing tracked that `.gitignore` excludes, no log drops or
 key material committed). Run it from the repo root or from `.agents/`.
 
@@ -113,7 +150,10 @@ nothing else in the codebase touches `os.environ`. Precedence: `undercooked` CLI
 flags, then exported environment variables, then a local `.env` (found from the
 current directory upwards; never commit it), then the built-in defaults. A blank value counts as
 unset. Invalid values (`UNDERCOOKED_TICK_HZ=99`) raise `ConfigError` at startup
-instead of half-configuring the server. `.env.example` and the table below list
+instead of half-configuring the server, and unusable content is refused the same
+way — an unreadable `UNDERCOOKED_LEVELS_DIR` or a level that breaks a PLAN §8
+rule prints the named-rule report and exits `2` (under `--reload` the reload
+child reports and restarts instead; docs/LEVELS.md). `.env.example` and the table below list
 every variable with its default; tests keep both in sync with `config.py`.
 
 | Variable | Default | Meaning |
@@ -125,6 +165,7 @@ every variable with its default; tests keep both in sync with `config.py`.
 | `UNDERCOOKED_MAX_PLAYERS` | `4` | Players per room (1-4) |
 | `UNDERCOOKED_MAX_ROOMS` | `200` | Concurrent rooms before new rooms are refused |
 | `UNDERCOOKED_LEVELS_DIR` | `levels` | Directory of level YAML files (`~` is expanded) |
+| `UNDERCOOKED_RECIPES_PATH` | *(derived)* | Recipe book file; empty = `recipes.yaml` inside `UNDERCOOKED_LEVELS_DIR` (here: `levels/recipes.yaml`) |
 | `UNDERCOOKED_STATIC_DIR` | *(derived)* | Directory served at `/static`; empty = `app/static/` beside the source (containers set it) |
 | `UNDERCOOKED_LOBBY_IDLE_TIMEOUT_S` | `900` | Close empty/idle lobbies |
 | `UNDERCOOKED_GUEST_IDLE_TIMEOUT_S` | `45` | Drop a disconnected player's seat after this |
@@ -142,17 +183,34 @@ src/undercooked/        server: FastAPI app, rooms, WebSocket protocol, simulati
 src/undercooked/__main__.py  the `undercooked` command (CLI > env > default)
 src/undercooked/config.py  typed env settings (the only `os.environ` reader)
 src/undercooked/debug.py   DEBUGPY-gated debugger bootstrap
-src/undercooked/server/app.py     app factory, `/healthz`, `/`, `/static`
+src/undercooked/sim/       pure simulation kernel: entities (`model`), grid geometry
+                           (`grid`), content schema + loader (`recipe`, `level`,
+                           `content` — the YAML contract is docs/LEVELS.md), the
+                           level validator (`validation` — PLAN §8 named rules;
+                           rule table in docs/LEVELS.md), and the
+                           A5 purity guard (`purity`, test-enforced). There is no
+                           `step()` yet: the tick arrives with phase 03
+src/undercooked/tools/levels.py   the `undercooked-levels` author CLI — validate, list,
+                        show [--grid], new, axes; same loader, same rules, no
+                        hard-coded content list (docs/LEVELS.md "Authoring workflow")
+src/undercooked/server/app.py     app factory, `/healthz`, `/`, `/static`, mounts the content API
+src/undercooked/server/routes_content.py  `/api/levels`, `/api/levels/{id}`, `/api/config`,
+                        plus the startup gate: broken content refuses startup
 src/undercooked/server/headers.py the security-header middleware
 .env.example            every configuration variable, with its default
-levels/                 data-driven level definitions (YAML) + recipes (empty until 02)
+levels/                 data-driven content (YAML): the recipe book (A8) and the
+                        first two campaign levels (01_first_kitchen, 02_double_shift)
 app/static/             browser client: the ES-module shell today; Canvas 2D renderer
                         and this repo's SVG art arrive with phase 07
-tests/                  unit/, integration/, sim/, e2e/ (one Playwright suite per story);
+tests/                  unit/, integration/, e2e/ (one Playwright suite per story) and
-                        fixtures/levels/ (test-isolated content dir); conftest.py
+                        sim/ (the pure kernel: grid, entities, purity guard, validation);
-docs/                   DESIGN.md, TESTING.md (how this project is verified),
+                        fixtures/levels/ is the content dir tests see (a recipe book and no
-                        DEPLOYMENT.md; PROTOCOL / LEVELS / SIMULATION / CLIENT land with
+                        levels; `_bad/` holds the deliberately-broken campaign the `_`
-                        their own phases (the index is PLAN.md §13)
+                        discovery rule ignores); conftest.py
+docs/                   DESIGN.md, LEVELS.md (the level & recipe YAML contract),
+                        TESTING.md (how this project is verified), DEPLOYMENT.md;
+                        PROTOCOL / SIMULATION / CLIENT land with their own phases
+                        (the index is PLAN.md §13)
 .agents/                PLAN.md, user_stories/, phases/, validate.sh — the phased-execution plan
 ```
 

docs/DEPLOYMENT.md

@@ -37,6 +37,11 @@ Rules that hold for every deployment of this project:
   template; the game has no secrets by design (anonymous, A6/A16).
 - Everything is in memory: a restart ends every round and erases every scoreboard. That is
   a documented product decision, not an incident.
+- **The campaign in the image is checked before the port opens.** `create_app` loads and
+  validates `UNDERCOOKED_LEVELS_DIR` at startup, so an image built around content that
+  breaks a PLAN §8 rule exits `2` with the named-rule report on stderr instead of serving
+  an empty lobby; with `restart: unless-stopped` that looks like a crash loop whose logs
+  are the diagnosis. `uv run undercooked-levels validate` reproduces it from the checkout.
 
 ## What this document will contain when it is filled in
 

docs/DESIGN.md

@@ -1,6 +1,6 @@
 # DESIGN.md — architecture and the decisions behind it
 
-**Status:** stub written by phase 01 · **Filled in by phase 09**, when the client, the
+**Status:** stub written by phase 01, module map refreshed by phase 02 · **Filled in by phase 09**, when the client, the
 protocol and the first two-player end-to-end story make the prose verifiable.
 
 ## Purpose
@@ -21,15 +21,21 @@ all. Everything around that kernel — `server/` (FastAPI app, WebSocket protoco
 supervisor, in-memory session records) and `app/static/` (vanilla ES modules, Canvas 2D)
 — exists only to move bytes between a player's hands and the simulation.
 
-## Module map (as it stands after phase 01)
+## Module map (as it stands after phase 02)
 
 | Path | Owns | Filled in by |
 |---|---|---|
 | `src/undercooked/config.py` | the only read of the environment → typed, validated `Settings` | phase 01 (done) |
 | `src/undercooked/debug.py` | `DEBUGPY`-gated debugger bootstrap (A18) | phase 01 (done) |
-| `src/undercooked/server/app.py` | app factory, `/healthz`, `/`, `/static`, startup logging | phase 01 (done) |
+| `src/undercooked/server/app.py` | app factory, `/healthz`, `/`, `/static`, startup logging, mounts the content router | phase 01 (done) |
+| `src/undercooked/server/routes_content.py` | `/api/levels`, `/api/levels/{id}`, `/api/config` behind `extra="forbid"` response models, plus the startup gate that refuses broken content | phase 02 (done) |
 | `src/undercooked/server/headers.py` | CSP + the four other security headers on every response | phase 01 (done) |
-| `src/undercooked/sim/` | game truth: model, `step()`, interactions, scoring, hazards | phase 03–04 |
+| `src/undercooked/sim/model.py`, `grid.py` | the entity vocabulary (`Vec2`, items, stations, players, orders, hazards, `KitchenState`) and tile geometry/walkability | phase 02 (done) |
+| `src/undercooked/sim/recipe.py`, `level.py`, `content.py` | the content contract (A8): recipe book, level schema + ASCII legend, discovery and campaign ordering | phase 02 (done); schema reference in `docs/LEVELS.md` |
+| `src/undercooked/sim/validation.py` | PLAN §8's named playability rules and the escalation axes | phase 02 (done) |
+| `src/undercooked/sim/purity.py` | the A5 guard: no I/O, asyncio, clock or FastAPI inside `sim/` | phase 02 (done) |
+| `src/undercooked/tools/levels.py` | the `undercooked-levels` author CLI — a shell over the loader and the rules, never a second implementation | phase 02 (done) |
+| `src/undercooked/sim/step.py` | game truth: the tick, interactions, scoring, hazards | phase 03–04 |
 | `src/undercooked/server/room.py`, `rooms.py` | one match, its FSM and tick task; the registry that owns all of them | phase 05–06 |
 | `src/undercooked/server/protocol.py` | versioned wire messages, whitelists, inbound validation | phase 06 |
 | `src/undercooked/server/sessions.py` | bounded in-memory progression records (A6: no database) | phase 05 |

docs/LEVELS.md

@@ -0,0 +1,499 @@
+# LEVELS.md — level & recipe YAML reference
+
+**Status:** current as of phase 02 (schema, parser, loader, validator, and the
+`undercooked-levels` author CLI). Phase 11 extends
+this document with the campaign-wide difficulty-curve notes; until then every
+field of every content model is documented here, and
+`tests/unit/test_levels_doc.py` keeps the two in lockstep.
+
+## Purpose
+
+This document defines the data-driven content contract for Undercooked: the
+recipe book, one YAML file per level, the ASCII-grid legend conventions, and
+every schema field with its type, default, and meaning. Content authors write
+files against this reference and the loader validates them at startup — the
+Python source never enumerates levels or recipes, and any document that fails
+these rules is a load error naming its file, never a silently-defaulted
+guess.
+
+The models live in `src/undercooked/sim/recipe.py` (book, recipes, steps,
+ingredients), `src/undercooked/sim/level.py` (levels, tiles, orders, hazards)
+and `src/undercooked/sim/content.py` (discovery and loading). All models are
+strict (`extra="forbid"`) and revalidate on assignment: an unknown key is an
+error, not a typo someone lives with.
+
+## Files on disk
+
+```text
+levels/
+  recipes.yaml            <- the recipe book (see "recipes.yaml" below)
+  01_first_kitchen.yaml   <- one level per file
+  02_double_shift.yaml
+  _draft.yaml             <- ignored: a leading underscore opts a file out
+  notes.md                <- ignored: only .yaml / .yml are content
+```
+
+`undercooked.sim.content.load_content(levels_dir, recipes_path)` is the whole
+loader. Its discovery rules:
+
+| Rule | Behaviour |
+|---|---|
+| Level discovery | every `*.yaml` / `*.yml` file under `levels_dir`, at any depth |
+| Drafts | files whose name — or any ancestor directory's name — starts with `_` are ignored |
+| The recipe book | `recipes.yaml` / `recipes.yml` inside `levels_dir` is never treated as a level; the book path is passed separately |
+| Ordering | the campaign order is `(order, id)` — never filename order |
+| `by_id` | derived from the ordered tuple; cannot disagree with it |
+| Duplicate ids | a load error listing every duplicated id |
+| Unknown pool recipes | a level whose `orders.pool` names a recipe the book does not define fails the load |
+| Error reporting | every file-level problem is aggregated into one `ContentError` naming each offending file |
+
+Every problem found in one run is aggregated into a single `ContentError` whose
+message is one line per offending file — no prefix, because the caller adds
+its own (`undercooked-levels` prefixes `undercooked-levels:`, `create_app`
+prefixes `undercooked: content failed to load from <dir>:`). It reads like:
+
+```text
+undercooked-levels: /…/levels/02_ghost.yaml: orders.pool names unknown recipes: ghost
+/…/levels/03_wreck.yaml: <root>: Value error, ragged grid: row 1 is 10 chars, want 9; grid character 'Z' has no tiles: legend entry
+```
+
+## recipes.yaml — the recipe book
+
+```yaml
+ingredients: [tomato, onion, plate]     # optional catalog of the campaign's items
+recipes:
+  soup:                                 # key must equal the recipe's id
+    id: soup
+    name: Tomato Soup
+    base_score: 40
+    ingredients:
+      - {kind: tomato, form: raw}
+      - {kind: onion, form: raw, count: 2}
+    steps:
+      - {kind: chop, station: chop_board, seconds: 0.9}
+      - {kind: cook, station: stove, seconds: 5.0}
+```
+
+### RecipeBook
+
+| Field | Type | Default | Meaning |
+|---|---|---|---|
+| `ingredients` | `list[ItemKind]` | `[]` | Catalog of item kinds the campaign uses. Informational in phase 02; item respawning uses it from phase 03 on. |
+| `recipes` | `map[id → Recipe]` | required, non-empty | The recipes. Each key must equal the recipe's own `id`. |
+
+Methods (deterministic, side-effect free):
+
+- `get(recipe_id)` → the `Recipe`; `KeyError` naming all known ids otherwise.
+- `ids()` → declaration order, as a tuple.
+- `required_station_kinds(recipe_id)` → `frozenset[StationKind]` of every
+  station the steps name. The `RECIPE_STATION_MISSING` rule consumes this;
+  nothing hard-codes "soup needs a stove".
+- `required_item_kinds(recipe_id)` → every ingredient kind plus `PLATE` when
+  `requires_plating` is true.
+- `min_seconds_to_complete(recipe_id)` → the theoretical minimum for a perfect
+  cook: the sum of all step `seconds` plus one flat handling allowance
+  (`HANDLING_S = 0.6` s per recipe, the constant in `sim/recipe.py`).
+  The `TARGET_IMPLAUSIBLE` rule uses this as the greedy per-order floor of
+  what a round can score (see "Validation rules").
+
+### Recipe
+
+| Field | Type | Default | Meaning |
+|---|---|---|---|
+| `id` | non-empty `str` | required | Stable identifier level order pools reference. |
+| `name` | non-empty `str` | required | Display name. |
+| `base_score` | `int ≥ 0` | required | Points for a perfect delivery before tips/combo (phase 04's quality, streak and time bonuses scale them). |
+| `ingredients` | `list[Ingredient]`, ≥ 1 | required | What the dish consumes. |
+| `steps` | `list[Step]`, ≥ 1 | required | The ordered procedure. |
+| `requires_plating` | `bool` | `true` | Whether a plain dish needs the plating step before serving. |
+| `tags` | `list[str]` | `[]` | Free-form labels for scoring/difficulty hints. |
+
+### Ingredient
+
+| Field | Type | Default | Meaning |
+|---|---|---|---|
+| `kind` | `ItemKind` | required | Which item (enum in `sim/model.py`; YAML uses the lowercase value). |
+| `form` | `ItemForm` | required | Form required at the moment of use: `raw`, `chopped`, `cooking`, `cooked`, `burnt`, `plated`, `dirty`. |
+| `count` | `int ≥ 1` | `1` | How many units the recipe consumes. |
+
+### Step
+
+| Field | Type | Default | Meaning |
+|---|---|---|---|
+| `kind` | `pick` \| `chop` \| `cook` \| `wash` \| `plate` \| `serve` | required | The verb. A closed vocabulary — anything else is a load error. Trashing is an interaction, not a recipe step, so it has no verb here. |
+| `seconds` | `float ≥ 0` | `0.0` | How long the action takes when performed. `pick`/`serve` are instantaneous (`0.0`). |
+| `station` | `StationKind` | required | Where the verb happens (`spawn_shelf`, `chop_board`, `stove`, `prep`, `serve`, `sink`, `trash`). |
+
+## A level file
+
+```yaml
+id: 01_first_kitchen            # convention: the file's stem (see Working conventions)
+name: First Kitchen
+order: 1                        # campaign position
+difficulty: 1                   # 1..10; feeds DIFFICULTY_ORDER + TRAVEL_TOO_SHORT, tuned in phase 11
+time_limit: 150                 # seconds of round time, > 0
+target_score: 45                # win threshold
+blurb: Two cooks, one pot.      # optional flavour text
+grid:                           # ASCII map, all rows equal length
+  - "#########"
+  - "#S..C...#"
+  - "#.......#"
+  - "#..V....#"
+  - "#########"
+tiles:                          # every grid char maps to exactly one TileSpec
+  ".": {kind: floor}
+  "#": {kind: wall}
+  "S": {kind: floor, station: spawn_shelf, item: tomato}
+  "C": {kind: floor, station: chop_board}
+  "V": {kind: floor, station: serve}
+spawns: [[1, 1]]                # player spawns as [x, y] grid coords
+orders:
+  pool: [soup]
+  concurrent: 1
+  spawn_interval: [16, 20]
+  deadline: 60
+hazards: []                     # optional, see HazardSpec
+station_limits: {plates: 2, pots: 1}   # optional, default 99 / 99
+```
+
+### LevelDef
+
+| Field | Type | Default | Meaning |
+|---|---|---|---|
+| `id` | non-empty `str` | required | Unique across the campaign; duplicate ids fail `load_content`. |
+| `name` | non-empty `str` | required | Display name. |
+| `order` | `int ≥ 1` | required | Campaign position. Levels are sorted by `(order, id)`. |
+| `difficulty` | `int` 1–10 | `1` | Author-declared difficulty. Feeds `DIFFICULTY_ORDER` (the campaign may not declare an easier level after a harder one) and switches on `TRAVEL_TOO_SHORT` from 3 upward; phase 11 tunes the curve against it. |
+| `time_limit` | `float > 0` | required | Round length in seconds. |
+| `target_score` | `int ≥ 1` | required | Score needed to win the round. |
+| `grid` | `list[str]`, ≥ 1 row | required | The ASCII map. Every row the same length, at most 64×64 (`MAX_GRID`), every character present in `tiles`. |
+| `tiles` | `map[char → TileSpec]`, ≥ 1 | required | The legend. Every legend character must be used by the grid — unused entries are typos. |
+| `spawns` | `list[[x, y]]`, ≥ 1 | required | Player spawn coordinates, each inside the grid. |
+| `orders` | `OrdersSpec` | required | The order engine configuration for this level. |
+| `hazards` | `list[HazardSpec]` | `[]` | Placed and periodic hazards. |
+| `station_limits` | `StationLimits` | `99 / 99` | Concurrent capacity of shared station resources. |
+| `blurb` | `str` | `""` | Flavour text for the level card. |
+| `escalation_notes` | `str` | `""` | Free-form note to the phase 11 difficulty-curve pass. |
+
+Schema-level rules (enforced right here, at load):
+
+- **rectangular grid** — differing row lengths are a load error naming the row;
+- **legend completeness both ways** — no grid character without a legend
+  entry, no legend character the grid never uses;
+- **spawn inside the grid**, at least one spawn;
+- **exactly one serve legend character** in the tiles map — the validator
+  rejects levels where the win condition cannot be reached by definition
+  (the validator's flood-fill rules below go further and reject pockets no
+  cook can walk into);
+- numeric sanity on every counter, interval, and limit listed above.
+
+### TileSpec
+
+| Field | Type | Default | Meaning |
+|---|---|---|---|
+| `kind` | `TileKind` (`floor` \| `counter` \| `wall` \| `void`) | required | Walkability class; only `floor` is walkable (`WALKABLE` in `sim/model.py`). |
+| `station` | `StationKind` or `null` | `null` | The station installed on this tile, if any. |
+| `item` | `ItemKind` or `null` | `null` | What this tile restocks — legal **only** on a `spawn_shelf` station tile, and **mandatory** there. |
+| `id_prefix` | non-empty `str` or `null` | `null` | Replaces the station-kind stem in generated ids (see "Station ids"). Allowed only on station tiles. |
+
+Stations sit on tiles of any `kind`; the schema allows it so hazards and
+decorations can mount stations, but the `STATION_REACHABLE` rule demands a
+walkable (`floor`) tile for every station players must touch.
+
+### OrdersSpec
+
+| Field | Type | Default | Meaning |
+|---|---|---|---|
+| `pool` | `list[str]`, ≥ 1 | required | Recipe ids this level draws orders from. Every entry must exist in the recipe book (`load_content` checks; the validator re-checks hand-built levels). |
+| `concurrent` | `int ≥ 1` | required | Maximum simultaneously active orders. |
+| `spawn_interval` | `[min, max]` seconds, both > 0 | required | Uniform range the round RNG draws from. `min ≤ max` is validated — reversed intervals are the classic misconfiguration. |
+| `deadline` | `float > 0` | required | Seconds an order may live before it burns. |
+| `penalty` | `int ≥ 0` | `10` | Points lost when an order burns out. |
+
+### HazardSpec
+
+One shape for placed and periodic hazards. Placement rules by `type`:
+
+| `type` | Placement | Meaning |
+|---|---|---|
+| `fire` | `station` **or** `x`/`y` (exactly one) | Ignites that station or that cell. |
+| `spill` | `station` **or** `x`/`y` (exactly one) | Grease puddle at that location. |
+| `moving_obstacle` | neither; `path` + `period_s` instead | Travels the path back and forth on a loop. |
+| `rush` | nowhere (level-wide) | A dining-rush surge; needs a positive `intensity`. |
+
+| Field | Type | Default | Meaning |
+|---|---|---|---|
+| `type` | `HazardKind` (`fire` \| `spill` \| `moving_obstacle` \| `rush`) | required | Which hazard. |
+| `station` | `str` or `null` | `null` | Station id to attach to (e.g. `stove_2`). Mutually exclusive with `x`/`y`. |
+| `x` | `int ≥ 0` or `null` | `null` | Cell x. Must be given together with `y`. |
+| `y` | `int ≥ 0` or `null` | `null` | Cell y. Must be given together with `x`. |
+| `chance_per_s` | `float ≥ 0` or `null` | `null` | Periodic re-arm probability per second; `null` = placed once at round start. |
+| `respawn_s` | `float > 0` or `null` | `null` | Delay before a cleared hazard can re-arm. |
+| `path` | `list[[x, y]]`, ≥ 2 points, or `null` | `null` | Route for `moving_obstacle` only. |
+| `period_s` | `float > 0` or `null` | `null` | Loop period for `moving_obstacle`; required when it has a `path`. |
+| `intensity` | `float` (must be > 0 when `type: rush`) | `1.0` | Magnitude (`rush` surge strength; severity multiplier elsewhere). |
+
+### StationLimits
+
+| Field | Type | Default | Meaning |
+|---|---|---|---|
+| `plates` | `int ≥ 0` | `99` | Clean plates available at once. `0` loads and is reported by `STATION_LIMIT_ZERO` — the rule, not a schema error, so the report names it. |
+| `pots` | `int ≥ 0` | `99` | Pots the stoves share. |
+
+## Station ids and the built `Level`
+
+`build_grid(spec)` walks the grid in reading order (row by row, left to
+right) and assigns each station tile an id: `{id_prefix or station-kind
+value}_{n}`, where `n` counts **per station kind** across the whole level.
+Two legend characters of the same kind therefore share one counter — a level
+with `T` (stove) and `t` (stove, `id_prefix: burner`) produces `stove_0` and
+`burner_1`. The same walk yields the `Station` list and `build_grid`'s
+`Grid.station_at` (cell index → station id), and `Grid.from_rows` re-checks
+geometry one last time so the grid cannot disagree with the spec.
+
+`Level.build(spec)` pairs the validated `LevelDef` with what the parser
+produces — `grid` (the `sim/grid.py` `Grid`), `stations`, `spawns` (as
+`(x, y)` tuples) and `station_items` (spawn-shelf id → restocked `ItemKind`)
+— and delegates `id`, `order`, `difficulty`, `time_limit`, `target_score`,
+`orders`, `hazards`, `station_limits`, `blurb` and `escalation_notes` to the
+spec, so simulation code reads one object. The `ContentBundle` returned by
+`load_content` holds the `recipe_book`, the campaign-ordered `levels`, and
+the derived `by_id` map.
+
+## Validation rules
+
+`sim/validation.py` proves a built `Level` is playable before anyone joins it —
+the PLAN §8 content gate. `validate_level(level, book, prev=None)` runs the
+per-level rules (`prev` is the campaign predecessor and enables the escalation
+check); `validate_levels(levels, book)` adds the bundle-wide rules over an
+`(order, id)`-ordered list and `validate_bundle(bundle)` runs that over a
+loaded bundle. Each hit is a `LevelViolation` naming a stable **rule id** with
+a human message carrying coordinates and a machine-readable `detail` mapping.
+`ValidationReport.render()` prints a deterministic, ANSI-free,
+level-grouped report, and **warnings never fail a level** — `report.ok` reads
+violations only.
+
+| Rule | Fires when |
+|---|---|
+| `GRID_SHAPE` | the built `Grid` disagrees with the spec's ASCII row count or width (hand-built levels; `LevelDef` validation already rejects ragged grids) |
+| `SPAWN_ON_FLOOR` | a spawn is not on a walkable floor tile |
+| `SPAWN_COUNT` | the level has no player spawns |
+| `STATION_REACHABLE` | a station cell cannot be walked to from any spawn |
+| `SERVE_REACHABLE` | the same for the serve window specifically — no reachable window means no round can ever be won |
+| `SPAWNS_MUTUALLY_REACHABLE` | a spawn cannot be walked to from the first spawn |
+| `RECIPE_STATION_MISSING` | a pool recipe needs a station kind the level does not have within the reachable area |
+| `RECIPE_INGREDIENT_UNAVAILABLE` | a pool recipe needs a `raw` ingredient no reachable spawn shelf produces (spawn shelves are currently the only item source the content model expresses) |
+| `ORDER_POOL_UNKNOWN_RECIPE` | `orders.pool` names a recipe the recipe book does not define |
+| `TARGET_IMPLAUSIBLE` | `target_score` exceeds the plausibility bound below — above 45% of the bound it is a **warning**, above 60% an error |
+| `TIME_LIMIT_INVALID` | `time_limit` is under 30 s |
+| `HAZARD_REF_UNKNOWN` | a hazard references a station id that does not exist |
+| `HAZARD_ON_UNREACHABLE_TILE` | a hazard tile or `moving_obstacle` path point is off-grid, not walkable, or cut off from the cooks |
+| `STATION_LIMIT_ZERO` | `station_limits.plates` or `.pots` is below 1 |
+| `DUPLICATE_ID` | two levels in the bundle share an `id` (bundle rule) |
+| `DIFFICULTY_ORDER` | a level raises fewer than two escalation axes over its predecessor, or the campaign's declared `difficulty` drops between consecutive levels (bundle rule) |
+| `TRAVEL_TOO_SHORT` | from difficulty 3 up, the mean spawn→serve path is under 3 tiles — sprint-to-serve layouts are not campaigns |
+
+**Target plausibility** (`theoretical_max_score`) is the greedy bound
+`floor(time_limit / mean_complete) × mean_order_value` over the level's order
+pool, where `mean_complete` is the recipe book's `min_seconds_to_complete()`
+average plus a travel bound of `2 × mean spawn→serve tiles ÷ 3.5 tiles/s` —
+one walk to the window and one back at max cook speed (PLAN §5). The bound is
+deliberately generous: it rejects targets a perfect team could never hit, not
+targets a sloppy one misses. Levels with an invalid time limit or no known
+pool recipes return no bound — those conditions have their own rules.
+
+**Escalation axes** (`escalation_axes`, every axis oriented *higher = harder*):
+
+| Axis | Measure |
+|---|---|
+| `recipe_complexity` | number of distinct recipes in the order pool |
+| `order_pressure` | `orders.concurrent` ÷ mean spawn interval |
+| `time_budget` | `target_score` ÷ `time_limit` |
+| `station_scarcity` | 1 ÷ station count |
+| `layout_hostility` | blocking tiles ÷ total tiles |
+| `hazard_density` | number of hazards |
+| `travel_distance` | mean BFS tiles from a spawn to the nearest serve cell |
+
+Reachability is a BFS flood fill over walkable tiles from every spawn.
+`moving_obstacle` tiles are deliberately *not* subtracted from the reachable
+set — movers move, so only their path points must be valid floor. When the
+serve window is unreachable the travel metric falls back to Manhattan
+distance so the numbers stay defined; the reachability rules report the break
+itself. `undercooked-levels validate` prints this report for the levels
+directory and exits `1` on any violation — see **Authoring workflow** below. The
+same report stops the game server: `create_app` runs that exact pair at startup,
+so `uv run undercooked` prints it, refuses to build an app around it and exits
+`2` rather than open a lobby nobody can cook in.
+
+## Working conventions
+
+- `id` should equal the file's stem; the loader keys everything by `id`, and
+  the bundle rule `DUPLICATE_ID` refuses a bundle that reuses one.
+- Two characters may share a `TileSpec` meaning only by writing it twice;
+  characters are single characters — a multi-character legend key can never
+  match a grid cell, so the load fails with `legend character 'VV' is never
+  used by the grid`, while a grid character with no entry at all fails with
+  `grid character 'Z' has no tiles: legend entry`.
+- Spawn shelves are the only item source at schema level; deeper restocking
+  semantics belong to phase 03's simulation work.
+- Everything strict: when a field is added to a model, `docs/LEVELS.md` must
+  gain a row in the same change — `tests/unit/test_levels_doc.py` fails if
+  the doc table and `model_fields` ever disagree.
+
+## Authoring workflow
+
+`undercooked-levels` is the author's half of this contract, and a thin shell
+over the code above: `validate` is exactly `load_content` + `validate_bundle`
+(the same pair `create_app` runs at startup), so green here means a server that
+boots there. It ships with the project (`[project.scripts]` in
+`pyproject.toml`) and never enumerates content itself — it prints whatever the
+directory it was pointed at contains. `tests/unit/test_content_not_hardcoded.py`
+holds the CLI, the loader and the content API to that rule.
+
+```bash
+uv run undercooked-levels new 03_riverboat      # 1. scaffold a level, validate that file
+uv run undercooked-levels list                   # 2. the campaign, in campaign order
+uv run undercooked-levels show 03_riverboat --grid   # 3. read the resolved level + overlay
+uv run undercooked-levels validate               # 4. the gate .agents/validate.sh runs
+uv run undercooked-levels axes                   # 5. the escalation table phase 11 tunes
+```
+
+Every subcommand takes `--dir PATH` (levels) and `--recipes PATH` (the book),
+which beat `UNDERCOOKED_LEVELS_DIR` / `UNDERCOOKED_RECIPES_PATH`, which beat the
+defaults `levels` and `<levels>/recipes.yaml`. Path resolution is settings-driven,
+never cwd-driven: a *relative* path with nothing under the current directory is
+anchored to this checkout, so the commands above work from `/tmp` exactly as
+they do from the repo root.
+
+1. **`new ID [--order N]`** writes `<levels>/ID.yaml` — a valid, minimal kitchen
+   *generated from the recipe book*: one spawn shelf per raw ingredient of the
+   book's first dish, one station per kind that dish's steps name, and two
+   spawns on the floor row far enough from the window to clear
+   `TRAVEL_TOO_SHORT`. `--order` defaults to one past the highest in the
+   directory; `difficulty` starts one above the campaign's hardest. It refuses
+   to overwrite an existing level, and prints the single-level validation
+   report of what it wrote.
+2. **Edit** the grid, the tunables and the prose. `id` should equal the file's
+   stem. Fill `escalation_notes` with the axes you moved: the campaign rule
+   (`DIFFICULTY_ORDER`) wants at least two axes up versus the level that
+   precedes yours by `order`, and a `difficulty` that never drops.
+3. **`list` and `show ID [--grid]`** read the level back the way the server sees
+   it — campaign order and resolved geometry, not the text you typed. `show`
+   prints the tunables, the station and hazard counts, how much of the floor is
+   reachable, and each pool dish resolved against the book (or
+   `MISSING from the recipe book`).
+4. **`validate`** prints one status line per level and then
+   `ValidationReport.render()`. Warnings print and stay green — a warning is a
+   tuning question, a violation is a broken level:
+
+   ```text
+   99_broken_room               4 error(s)               11x6, 4 stations
+   4 error(s), 0 warning(s)
+   level 99_broken_room
+     ERROR RECIPE_STATION_MISSING: recipe 'soup' needs a reachable prep station; the level has none
+     ERROR RECIPE_STATION_MISSING: recipe 'soup' needs a reachable stove station; the level has none
+     ERROR SERVE_REACHABLE: serve station serve_0 at (8,1) is not reachable from any spawn
+     ERROR TARGET_IMPLAUSIBLE: target_score 999 is above 60% of the theoretical maximum 160; no one could ever reach it
+   ```
+
+5. **`axes`** prints the seven escalation axes per level in campaign order
+   (higher = harder), which is how you check a level really got harder instead
+   of only *feeling* harder:
+
+   ```text
+   LEVEL                        ORDER   recipe_complexity      order_pressure         time_budget    station_scarcity    layout_hostility      hazard_density     travel_distance
+   01_first_kitchen                 1              1.0000              0.0556              0.3000              0.1429              0.3519              0.0000              7.5000
+   02_double_shift                  2              2.0000              0.1429              0.6154              0.1250              0.4786              0.0000              5.5000
+   ```
+
+### The overlay debugging trick
+
+`STATION_REACHABLE`, `SERVE_REACHABLE` and `HAZARD_ON_UNREACHABLE_TILE` all
+come down to one question: which walkable floor can the cooks actually get to?
+`show --grid` re-draws the authored map as terrain so you see the answer
+instead of deducing it from a BFS you cannot see:
+
+```bash
+uv run undercooked-levels show 99_broken_room --grid --dir tests/fixtures/levels/_bad
+```
+
+```text
+map (as authored)         <- the file, exactly as written
+  ###########
+  #T..C#..V.#
+  #....#....#
+  #.#########
+  #S........#
+  ###########
+
+reachability  . reachable   ? walkable but cut off   # blocking
+  ###########
+  #....#????#
+  #....#????#
+  #.#########
+  #.........#
+  ###########
+
+  18/26 walkable tiles are reachable from the spawns
+```
+
+The `?` band is the sealed servery: walkable floor, and the wall at column 6
+cuts it off from the spawn on row 4. Fix the wall, not the target score. Note
+that a moving obstacle's `path` tiles stay `.` even when a mover currently
+blocks them — movers move, so the validator only insists the path itself is
+valid floor.
+
+### Scripting it
+
+`validate --json` is the machine-readable form of the same report — stable key
+names, sorted hits, so a diff of two runs is only a diff of real changes:
+
+```bash
+uv run undercooked-levels validate --json | python -m json.tool
+```
+
+```json
+{
+  "ok": false,
+  "counts": {"errors": 4, "warnings": 0},
+  "errors": [
+    {
+      "level": "99_broken_room",
+      "rule": "SERVE_REACHABLE",
+      "message": "serve station serve_0 at (8,1) is not reachable from any spawn",
+      "detail": {"station_id": "serve_0", "x": 8, "y": 1}
+    }
+  ]
+}
+```
+
+Exit codes are the whole scripting surface: `0` everything is fine, `1` the
+content loaded and breaks at least one rule, `2` the tool could not do its job
+(content would not load, an unknown level id, a file `new` refuses to overwrite,
+or an argparse usage error). `.agents/validate.sh` runs `validate` unconditionally,
+so a campaign that cannot validate cannot commit even when no test file changed.
+
+### Which rule did I just hit?
+
+The full semantics are in the **Validation rules** table above; this is the
+"what do I type next" view of the same ids.
+
+| Rule | Usually means |
+|---|---|
+| `GRID_SHAPE` | the row count/width you wrote and the built grid disagree — check for a ragged row or a stray character |
+| `SPAWN_ON_FLOOR` | a `spawns:` coordinate lands on wall/counter — move it onto a `.` |
+| `SPAWN_COUNT` | `spawns:` is empty — a kitchen nobody stands in |
+| `STATION_REACHABLE` | a fixture sits behind a wall — `show --grid` and open a path |
+| `SERVE_REACHABLE` | the window is unreachable, so the round cannot be won — same fix, and the level is unwinnable until then |
+| `SPAWNS_MUTUALLY_REACHABLE` | the cooks start on islands — connect the spawn areas |
+| `RECIPE_STATION_MISSING` | a pool dish needs a station the level lacks (or cannot reach) — add the fixture or trim the pool |
+| `RECIPE_INGREDIENT_UNAVAILABLE` | a pool dish eats an item no reachable spawn shelf produces — add a shelf for it |
+| `ORDER_POOL_UNKNOWN_RECIPE` | `orders.pool` names a dish `recipes.yaml` does not define — check the spelling |
+| `TARGET_IMPLAUSIBLE` | `target_score` is out of reach of a perfect team — lower the target, lengthen `time_limit`, or raise the dish values |
+| `TIME_LIMIT_INVALID` | `time_limit` under 30 s — a round that is not a round |
+| `HAZARD_REF_UNKNOWN` | a hazard's `station:` id is not in the level — re-read the station ids the loader built |
+| `HAZARD_ON_UNREACHABLE_TILE` | a hazard sits in a wall or a sealed pocket — move it where the cooks are |
+| `STATION_LIMIT_ZERO` | `station_limits.plates`/`.pots` is 0, so no dish can ever be plated |
+| `DUPLICATE_ID` | two files declare the same `id` — one of them is a copy-paste |
+| `DIFFICULTY_ORDER` | the level moves fewer than two escalation axes up, or `difficulty` dropped — compare with `axes` |
+| `TRAVEL_TOO_SHORT` | from difficulty 3, spawn-to-serve averages under 3 tiles — spread the kitchen out |

docs/TESTING.md

@@ -1,6 +1,6 @@
 # TESTING.md — how this project is verified
 
-**Status:** current as of phase 01 · **Filled in by phase 14**, which re-runs every layer
+**Status:** current as of phase 02 · **Filled in by phase 14**, which re-runs every layer
 below against the finished game and records the release-time suite inventory.
 
 ## Purpose
@@ -20,7 +20,7 @@ layer live in `tests/conftest.py`, and layer-specific fixtures live in that laye
 | Directory | Contains | I/O allowed |
 |---|---|---|
 | `tests/unit/` | pure logic with no I/O: config parsing, header values, CLI precedence, protocol/validator helpers, and the doc/content guards that keep README, `.env.example`, `PLAN.md` and the code agreeing | none (in-process `TestClient` requests are fine) |
-| `tests/sim/` | headless determinism and replay tests against the pure kernel (`tests/sim/test_replay_*.py`, golden state hashes) | none — the sim must stay callable without a server |
+| `tests/sim/` | tests of the pure kernel that need no server: grid geometry, the entity model, the A5 import-purity guard, and the level validator (phase 02) — plus the headless determinism/replay suites with golden state hashes (`tests/sim/test_replay_*.py`) that arrive with the tick in phase 03 | none — the sim must stay callable without a server |
 | `tests/integration/` | multi-module, in-process tests: app boot through `create_app()`, HTTP routes, static serving, room/protocol behaviour with real WebSocket clients (room + protocol integration is PLAN §9 layer 4, and lives here) | in-process app, loopback sockets |
 | `tests/e2e/` | one Playwright suite per user story, named `test_usNN_*.py`; multiplayer stories drive **two browser contexts** against one room | real browser, real socket |
 | `tests/fixtures/` | fixture data only. `tests/fixtures/levels/` is the content dir every test sees — an autouse session fixture points `UNDERCOOKED_LEVELS_DIR` at it so no test can silently read the shipped `levels/` | none |
@@ -59,6 +59,16 @@ uv run pyright                                 # types
 bash .agents/validate.sh                       # the pipeline gate, in that order + hygiene
 ```
 
+The gate runs one check that is not a pytest command: `uv run undercooked-levels
+validate` between pyright and pytest (docs/LEVELS.md **Authoring workflow**). It is
+unconditional — a campaign that breaks a PLAN §8 rule stops the pipeline even when no
+test file changed — and it is the same `load_content` + `validate_bundle` pair
+`create_app()` runs at startup, so it can disagree with the server about nothing.
+Content tests stay hermetic the same way: the autouse session fixture points
+`UNDERCOOKED_LEVELS_DIR` at `tests/fixtures/levels/` (a recipe book and zero levels),
+and a test that wants real levels builds one under `tmp_path` or names `levels/`
+explicitly. `tests/unit/test_content_not_hardcoded.py` is the A8 lock on top of that.
+
 Coverage is configured in `pyproject.toml`: `source = ["src/undercooked"]`, branch
 coverage on, `--cov-fail-under=90` (AGENTS.md A20 asks for > 90 % on new/changed code; the
 floor is the mechanical half of that).

levels/01_first_kitchen.yaml

@@ -0,0 +1,45 @@
+# levels/01_first_kitchen.yaml — campaign level 1, the tutorial kitchen.
+# Schema: docs/LEVELS.md · validated at startup and by `undercooked-levels
+# validate`. Every '#' is wall, every station stands on walkable floor.
+
+id: 01_first_kitchen
+name: First Kitchen
+order: 1
+difficulty: 1
+time_limit: 150
+target_score: 45
+blurb: >-
+  Your first shift: grab a tomato and an onion from the shelves, chop them,
+  cook the soup, plate it, and run it to the serve window before the clock
+  runs out. One dish, one of everything, nowhere to bump into each other.
+grid:
+  - "############"
+  - "#T..C...S..#"
+  - "#..........#"
+  - "#.P......X.#"
+  - "#..........#"
+  - "#..K....O..#"
+  - "#..........#"
+  - "#..........#"
+  - "############"
+tiles:
+  "#": {kind: wall}
+  ".": {kind: floor}
+  "T": {kind: floor, station: spawn_shelf, item: tomato}
+  "O": {kind: floor, station: spawn_shelf, item: onion}
+  "C": {kind: floor, station: chop_board}
+  "S": {kind: floor, station: stove}
+  "P": {kind: floor, station: prep}
+  "X": {kind: floor, station: serve}
+  "K": {kind: floor, station: sink}
+spawns: [[2, 7], [9, 7]]
+orders:
+  pool: [soup]
+  concurrent: 1
+  spawn_interval: [16, 20]
+  deadline: 60
+hazards: []
+escalation_notes: >-
+  Tutorial baseline (PLAN §8 campaign plan): one recipe, one order at a time,
+  slowest spawn interval, longest time budget, open floor, no hazards. Every
+  later level moves at least two axes up from these numbers.

levels/02_double_shift.yaml

@@ -0,0 +1,49 @@
+# levels/02_double_shift.yaml — campaign level 2: two dishes at once.
+# Schema: docs/LEVELS.md · validated at startup and by `undercooked-levels
+# validate`. The counter bands pinch the kitchen into three rooms joined by
+# one narrow walk-through.
+
+id: 02_double_shift
+name: Double Shift
+order: 2
+difficulty: 2
+time_limit: 130
+target_score: 80
+blurb: >-
+  Salads join the soup and the orders overlap now. Two counter bands pinch the
+  floor, so someone is always walking the long way through the gap — plan who
+  crosses and who stays on their side.
+grid:
+  - "#############"
+  - "#T....C....O#"
+  - "#.....S.....#"
+  - "####=...=####"
+  - "#....P.X....#"
+  - "#...........#"
+  - "####=...=####"
+  - "#..K.....L..#"
+  - "#############"
+tiles:
+  "#": {kind: wall}
+  ".": {kind: floor}
+  "=": {kind: counter}
+  "T": {kind: floor, station: spawn_shelf, item: tomato}
+  "O": {kind: floor, station: spawn_shelf, item: onion}
+  "L": {kind: floor, station: spawn_shelf, item: lettuce}
+  "C": {kind: floor, station: chop_board}
+  "S": {kind: floor, station: stove}
+  "P": {kind: floor, station: prep}
+  "X": {kind: floor, station: serve}
+  "K": {kind: floor, station: sink}
+spawns: [[2, 2], [10, 5]]
+orders:
+  pool: [soup, salad]
+  concurrent: 2
+  spawn_interval: [12, 16]
+  deadline: 52
+hazards: []
+escalation_notes: >-
+  Moves four escalation axes vs 01_first_kitchen (PLAN §8 requires two):
+  recipe_complexity 1 -> 2 (salad joins), order_pressure 0.056 -> 0.143
+  (2 concurrent at shorter intervals), time_budget 0.30 -> 0.62 target/s,
+  and layout_hostility 0.35 -> 0.48 blocking tiles with the counter pinch.

levels/recipes.yaml

@@ -0,0 +1,80 @@
+# levels/recipes.yaml — the campaign recipe book (PLAN §5/§8, A8).
+#
+# Every dish the game knows lives here; no Python enumerates recipes. The
+# schema is `src/undercooked/sim/recipe.py` and the reference is
+# docs/LEVELS.md — every model is strict (`extra="forbid"`), so a typo here
+# fails the load loudly instead of silently defaulting.
+#
+# Step lists mirror PLAN §5's station pipelines: pick from the spawn shelf,
+# chop (~0.9 s), cook on the stove, plate at the prep station, serve at the
+# window. `seconds` feeds only the validator's greedy target-plausibility
+# bound; the simulation (phase 04) re-enacts these steps per tick.
+
+# Catalog of the item kinds the campaign's dishes consume (documentation of
+# intent; spawn shelves are what actually put items on the floor).
+ingredients: [tomato, onion, lettuce, dough, patty, rice, fish, plate]
+
+recipes:
+  # tomato + onion -> chop -> pot on the stove -> plate
+  soup:
+    id: soup
+    name: Tomato Soup
+    base_score: 40
+    ingredients:
+      - {kind: tomato, form: raw}
+      - {kind: onion, form: raw}
+    steps:
+      - {kind: pick, station: spawn_shelf}
+      - {kind: chop, station: chop_board, seconds: 0.9}
+      - {kind: cook, station: stove, seconds: 5.0}
+      - {kind: plate, station: prep, seconds: 0.4}
+      - {kind: serve, station: serve}
+    tags: [hot, soup]
+
+  # lettuce + tomato -> chop -> plate; the fast, no-fire dish
+  salad:
+    id: salad
+    name: Garden Salad
+    base_score: 30
+    ingredients:
+      - {kind: lettuce, form: raw}
+      - {kind: tomato, form: raw}
+    steps:
+      - {kind: pick, station: spawn_shelf}
+      - {kind: chop, station: chop_board, seconds: 0.9}
+      - {kind: plate, station: prep, seconds: 0.4}
+      - {kind: serve, station: serve}
+    tags: [cold, quick]
+
+  # rice + onion straight onto the stove (PLAN §5: raw rice needs no board)
+  fried_rice:
+    id: fried_rice
+    name: Fried Rice
+    base_score: 50
+    ingredients:
+      - {kind: rice, form: raw}
+      - {kind: onion, form: raw}
+    steps:
+      - {kind: pick, station: spawn_shelf}
+      - {kind: cook, station: stove, seconds: 6.0}
+      - {kind: plate, station: prep, seconds: 0.4}
+      - {kind: serve, station: serve}
+    tags: [hot, grain]
+
+  # patty + dough + lettuce -> chop + pan on the stove -> plate. Defined now;
+  # later levels (phase 11) put it in an order pool.
+  burger:
+    id: burger
+    name: Chef's Burger
+    base_score: 70
+    ingredients:
+      - {kind: patty, form: raw}
+      - {kind: dough, form: raw}
+      - {kind: lettuce, form: raw}
+    steps:
+      - {kind: pick, station: spawn_shelf}
+      - {kind: chop, station: chop_board, seconds: 0.9}
+      - {kind: cook, station: stove, seconds: 5.0}
+      - {kind: plate, station: prep, seconds: 0.5}
+      - {kind: serve, station: serve}
+    tags: [hot, stacked]

src/undercooked/__main__.py

@@ -9,7 +9,8 @@ exactly the same configuration as this process, and so an invalid ``--port`` is
 rejected by the same validator as an invalid `UNDERCOOKED_PORT`.
 
 Exit codes: ``0`` after a clean shutdown, ``2`` for a configuration error
-(matching argparse's own usage-error code), anything else comes from uvicorn.
+(matching argparse's own usage-error code) *or* level content the startup gate
+refused to build an app around, anything else comes from uvicorn.
 """
 
 from __future__ import annotations
@@ -21,6 +22,7 @@ from pathlib import Path
 from typing import TYPE_CHECKING, Any
 
 from undercooked.config import ConfigError, Settings, apply_cli_overrides
+from undercooked.sim.recipe import ContentError
 
 if TYPE_CHECKING:
     from collections.abc import Mapping, Sequence
@@ -148,7 +150,19 @@ def main(argv: Sequence[str] | None = None) -> int:
         print(f"undercooked: {exc}", file=sys.stderr)
         return EXIT_CONFIG_ERROR
 
-    _serve(options)
+    try:
+        _serve(options)
+    except ContentError:
+        # Building the app is what reads the content, and `create_app` has already
+        # written the full named-rule report to stderr before raising. A developer
+        # gets that report plus one closing line, not a traceback: content nobody
+        # can cook is as much a startup error as a port nobody can bind.
+        print(
+            "undercooked: refusing to start — the report above names every broken "
+            "rule (`undercooked-levels validate` re-checks them)",
+            file=sys.stderr,
+        )
+        return EXIT_CONFIG_ERROR
     return EXIT_OK
 
 

src/undercooked/config.py

@@ -57,6 +57,7 @@ ENV_KEYS: Mapping[str, str] = {
     "max_players": f"{ENV_PREFIX}MAX_PLAYERS",
     "max_rooms": f"{ENV_PREFIX}MAX_ROOMS",
     "levels_dir": f"{ENV_PREFIX}LEVELS_DIR",
+    "recipes_path": f"{ENV_PREFIX}RECIPES_PATH",
     "static_dir": f"{ENV_PREFIX}STATIC_DIR",
     "lobby_idle_timeout_s": f"{ENV_PREFIX}LOBBY_IDLE_TIMEOUT_S",
     "guest_idle_timeout_s": f"{ENV_PREFIX}GUEST_IDLE_TIMEOUT_S",
@@ -76,6 +77,9 @@ DEFAULT_SNAPSHOT_HZ = 20
 DEFAULT_MAX_PLAYERS = 4
 DEFAULT_MAX_ROOMS = 200
 DEFAULT_LEVELS_DIR = Path("levels")
+# `None` means "the recipe book beside the levels directory" (see
+# `server.routes_content.resolve_recipes_path`), i.e. `levels/recipes.yaml`.
+DEFAULT_RECIPES_PATH: Path | None = None
 # `None` means "derive it from the installed package layout" (see
 # `server.app.STATIC_DIR`); containers set an explicit path instead.
 DEFAULT_STATIC_DIR: Path | None = None
@@ -119,6 +123,8 @@ class Settings:
         max_players: seats per room (PLAN 2: 1-4).
         max_rooms: process-wide room cap before new rooms are refused.
         levels_dir: directory of level YAML files (A8); ``~`` is expanded.
+        recipes_path: the recipe book file (A8); ``None`` resolves to
+            ``levels_dir/recipes.yaml``, and ``~`` is expanded when set.
         static_dir: directory served at ``/static``; ``None`` derives it from the
             package layout (``<repo>/app/static``), which is wrong inside a
             container image, so those set it explicitly.
@@ -140,6 +146,7 @@ class Settings:
     max_players: int = DEFAULT_MAX_PLAYERS
     max_rooms: int = DEFAULT_MAX_ROOMS
     levels_dir: Path = DEFAULT_LEVELS_DIR
+    recipes_path: Path | None = DEFAULT_RECIPES_PATH
     static_dir: Path | None = DEFAULT_STATIC_DIR
     lobby_idle_timeout_s: float = DEFAULT_LOBBY_IDLE_TIMEOUT_S
     guest_idle_timeout_s: float = DEFAULT_GUEST_IDLE_TIMEOUT_S
@@ -286,6 +293,7 @@ def load_settings(env: Mapping[str, str] | None = None) -> Settings:
         max_players=max_players,
         max_rooms=_int(env, f"{prefix}MAX_ROOMS", DEFAULT_MAX_ROOMS, MIN_ROOMS),
         levels_dir=_path(env, f"{prefix}LEVELS_DIR", DEFAULT_LEVELS_DIR),
+        recipes_path=_optional_path(env, f"{prefix}RECIPES_PATH"),
         static_dir=_optional_path(env, f"{prefix}STATIC_DIR"),
         lobby_idle_timeout_s=_float(
             env, f"{prefix}LOBBY_IDLE_TIMEOUT_S", DEFAULT_LOBBY_IDLE_TIMEOUT_S, MIN_TIMEOUT_S

src/undercooked/server/app.py

@@ -5,9 +5,13 @@ WebSocket endpoint (`/ws/{code}`, phase 06) and the level manifest (phase 02)
 all register on the app built here. Nothing game-like exists yet; the contract
 is only:
 
-* ``GET /healthz`` — liveness probe for containers and tests,
+* ``GET /healthz``  — liveness probe for containers and tests,
-* ``GET /``        — the client shell from ``app/static/index.html``,
+* ``GET /``         — the client shell from ``app/static/index.html``,
-* ``/static/*``    — the client's ES modules, CSS and art (never a CDN, A12),
+* ``/static/*``     — the client's ES modules, CSS and art (never a CDN, A12),
+* ``/api/levels``, ``/api/levels/{id}``, ``/api/config`` — the content API
+  (:mod:`undercooked.server.routes_content`), fed by the bundle loaded once
+  here: content that will not load or fails PLAN §8 validation refuses
+  startup with the report on stderr instead of serving a half-working lobby,
 * security headers on every response (:mod:`undercooked.server.headers`).
 
 ``create_app`` takes a :class:`~undercooked.config.Settings` as plain data, so a
@@ -31,6 +35,7 @@ from undercooked import __version__
 from undercooked.config import ConfigError, Settings, get_settings
 from undercooked.debug import maybe_start_debugpy
 from undercooked.server.headers import SecurityHeadersMiddleware
+from undercooked.server.routes_content import content_router, load_validated_content
 
 __all__ = [
     "STATIC_DIR",
@@ -87,6 +92,9 @@ def create_app(settings: Settings | None = None) -> FastAPI:
     Raises:
         ConfigError: the resolved static directory does not exist — better to
             fail at startup than to serve 500s for every asset.
+        ContentError: the level/recipe content will not load, or a level fails
+            a PLAN §8 validation rule — the full report goes to stderr and no
+            app is built (a broken kitchen must not open as a lobby).
     """
     resolved = settings if settings is not None else get_settings()
     static_dir = resolve_static_dir(resolved)
@@ -95,6 +103,7 @@ def create_app(settings: Settings | None = None) -> FastAPI:
             f"static directory {static_dir} does not exist "
             "(set UNDERCOOKED_STATIC_DIR to the directory holding index.html)"
         )
+    content = load_validated_content(resolved)
 
     @asynccontextmanager
     async def lifespan(_app: FastAPI) -> AsyncIterator[None]:
@@ -115,8 +124,10 @@ def create_app(settings: Settings | None = None) -> FastAPI:
         openapi_url=None,
     )
     application.state.settings = resolved
+    application.state.content = content  # loaded + validated once, above
     application.state.rooms = None  # phase 05 installs the RoomSupervisor here
     application.add_middleware(SecurityHeadersMiddleware)
+    application.include_router(content_router)
 
     @application.get("/healthz")
     async def healthz() -> dict[str, object]:

src/undercooked/server/routes_content.py

@@ -0,0 +1,213 @@
+"""The content HTTP surface: `/api/levels`, `/api/levels/{id}`, `/api/config`.
+
+The client's level cards and lobby settings come from here — never from a
+bundled JSON copy — so content edits need no client change (A8). Everything
+is projected through the strict response models below, which exist to
+*forgive* nothing: `extra="forbid"` means a field that leaks into a response
+(a grid row, a station id, an internal spec) breaks `model_validate` in
+`tests/integration/test_content_api.py` instead of quietly reaching the
+browser. Layout, legend, station ids and spawn coordinates are internal and
+never leave the server (PLAN §8, US-07).
+
+`load_validated_content` is the startup half of the same contract:
+`create_app` calls it once, stores the bundle on `app.state.content`, and a
+content bug — a file that will not load or a level that violates a named
+validation rule — prints the full report on stderr and raises
+`ContentError`, so the app refuses to start rather than serve a half-working
+lobby around a broken kitchen.
+"""
+
+from __future__ import annotations
+
+import logging
+import sys
+from pathlib import Path
+from typing import Final, NoReturn
+
+from fastapi import APIRouter, HTTPException, Request
+from pydantic import BaseModel, ConfigDict
+
+from undercooked.config import Settings
+from undercooked.sim.content import ContentBundle, load_content
+from undercooked.sim.level import Level
+from undercooked.sim.model import HazardKind
+from undercooked.sim.recipe import ContentError, RecipeBook
+from undercooked.sim.validation import ValidationReport, validate_bundle
+
+__all__ = [
+    "PROTOCOL_VERSION",
+    "RECIPES_FILENAME",
+    "ClientConfig",
+    "HazardSummary",
+    "LevelDetail",
+    "LevelSummary",
+    "content_router",
+    "load_validated_content",
+    "resolve_recipes_path",
+]
+
+logger = logging.getLogger(__name__)
+
+#: Wire protocol major this server speaks (PLAN §7/A14). Phase 06 moves this
+#: constant to `server/protocol.py` and owns its bump policy; until then the
+#: client-safe value is published here so `/api/config` can promise it.
+PROTOCOL_VERSION: Final = 1
+
+#: Default file name of the recipe book, resolved inside `levels_dir`.
+RECIPES_FILENAME: Final = "recipes.yaml"
+
+
+class LevelSummary(BaseModel):
+    """One level card: everything a lobby may show, nothing it may not."""
+
+    model_config = ConfigDict(extra="forbid")
+
+    id: str
+    name: str
+    order: int
+    difficulty: int
+    blurb: str
+    target_score: int
+    time_limit: float
+    recipes: list[str]
+
+
+class HazardSummary(BaseModel):
+    """Only *that* a level carries a hazard kind — never where it sits."""
+
+    model_config = ConfigDict(extra="forbid")
+
+    type: HazardKind
+
+
+class LevelDetail(LevelSummary):
+    """A level card plus the hazard kinds it uses; layout stays server-side."""
+
+    hazards: list[HazardSummary]
+
+
+class ClientConfig(BaseModel):
+    """The client-safe settings subset — and nothing else (task 04)."""
+
+    model_config = ConfigDict(extra="forbid")
+
+    max_players: int
+    tick_hz: int
+    snapshot_hz: int
+    protocol_version: int
+
+
+def _level_summary(level: Level, book: RecipeBook) -> LevelSummary:
+    """Project one built level onto the public card shape."""
+    return LevelSummary(
+        id=level.id,
+        name=level.name,
+        order=level.order,
+        difficulty=level.difficulty,
+        blurb=level.blurb,
+        target_score=level.target_score,
+        time_limit=level.time_limit,
+        recipes=[book.get(recipe_id).name for recipe_id in level.orders.pool],
+    )
+
+
+def _level_detail(level: Level, book: RecipeBook) -> LevelDetail:
+    """The card plus hazard kinds — no positions, no station ids."""
+    return LevelDetail(
+        **_level_summary(level, book).model_dump(),
+        hazards=[HazardSummary(type=hazard.type) for hazard in level.hazards],
+    )
+
+
+def resolve_recipes_path(settings: Settings) -> Path:
+    """The recipe book: the explicit override, else `recipes.yaml` in `levels_dir`."""
+    if settings.recipes_path is not None:
+        return settings.recipes_path
+    return settings.levels_dir / RECIPES_FILENAME
+
+
+def _render_report(report: ValidationReport) -> str:
+    """The rendered validation report (errors *and* warnings, deterministic)."""
+    return report.render()
+
+
+def _refuse(message: str) -> NoReturn:
+    """Print the full report on stderr and refuse to start.
+
+    `create_app` callers (uvicorn, `python -m undercooked`, tests) see the
+    `ContentError`; the stderr half is what a human operator reads first, so
+    the named-rule report must be on screen even when the traceback is long.
+    """
+    print(message, file=sys.stderr)
+    raise ContentError(message)
+
+
+def load_validated_content(settings: Settings) -> ContentBundle:
+    """Load and validate the whole campaign; any failure refuses startup.
+
+    Loading is file + schema errors aggregated by `load_content`; validation
+    is PLAN §8's named-rule gate (`validate_bundle`). Errors print their full
+    report to stderr and raise `ContentError`; warnings are logged level by
+    level and never stop the server — a tight target is a warning, a broken
+    kitchen is not.
+    """
+    try:
+        bundle = load_content(settings.levels_dir, resolve_recipes_path(settings))
+    except ContentError as exc:
+        _refuse(f"undercooked: content failed to load from {settings.levels_dir}:\n{exc}")
+    report = validate_bundle(bundle)
+    for warning in report.warnings:
+        logger.warning(
+            "level %s: %s: %s", warning.level_id, warning.rule, warning.message
+        )
+    if not report.ok:
+        _refuse(f"undercooked: level content failed validation:\n{_render_report(report)}")
+    logger.info(
+        "content loaded: %d level(s), %d recipe(s) from %s",
+        len(bundle.levels),
+        len(bundle.recipe_book.recipes),
+        settings.levels_dir,
+    )
+    return bundle
+
+
+def _bundle(request: Request) -> ContentBundle:
+    """The bundle `create_app` refused to build a server without."""
+    bundle: ContentBundle = request.app.state.content
+    return bundle
+
+
+content_router = APIRouter(tags=["content"])
+
+
+@content_router.get("/api/levels", response_model=list[LevelSummary])
+async def list_levels(request: Request) -> list[LevelSummary]:
+    """Every shipped level, in campaign order, as strict level cards."""
+    bundle = _bundle(request)
+    return [_level_summary(level, bundle.recipe_book) for level in bundle.levels]
+
+
+@content_router.get("/api/levels/{level_id}", response_model=LevelDetail)
+async def get_level(request: Request, level_id: str) -> LevelDetail:
+    """One level card plus its hazard kinds; 404 for an unknown id.
+
+    The id comes from the client and is only ever a lookup into the loaded
+    bundle — nothing about it is trusted or echoed beyond the 404 decision.
+    """
+    bundle = _bundle(request)
+    level = bundle.by_id.get(level_id)
+    if level is None:
+        raise HTTPException(status_code=404, detail="unknown level")
+    return _level_detail(level, bundle.recipe_book)
+
+
+@content_router.get("/api/config", response_model=ClientConfig)
+async def get_config(request: Request) -> ClientConfig:
+    """The four client-safe settings; no secret-, operator- or origin-named key."""
+    settings: Settings = request.app.state.settings
+    return ClientConfig(
+        max_players=settings.max_players,
+        tick_hz=settings.tick_hz,
+        snapshot_hz=settings.snapshot_hz,
+        protocol_version=PROTOCOL_VERSION,
+    )

src/undercooked/sim/content.py

@@ -0,0 +1,174 @@
+"""The content loader: a directory of YAML files -> one validated `ContentBundle`.
+
+Purity contract (PLAN.md A5, enforced by `undercooked.sim.purity`): standard
+library plus `pydantic`/`PyYAML` and `undercooked.sim` imports only — no
+asyncio, no sockets, no wall-clock, no FastAPI. Reading the level directory
+is the load-time half of the content pipeline (A8): the shipped `levels/`
+tree is the only source of levels and recipes, and nothing in Python
+enumerates either.
+
+Discovery rules (documented in `docs/LEVELS.md`):
+
+* every `*.yaml` / `*.yml` file under `levels_dir`, at any depth, is a level;
+* files whose name — or any ancestor directory's name — starts with `_` are
+  ignored (fixture overrides, drafts);
+* `recipes.yaml` / `recipes.yml` is ignored because it is the recipe book,
+  which is passed to `load_content` as its own path;
+* levels are campaign-ordered by `(order, id)` — ties break on id, never on
+  file-system order;
+* every problem in the run — unreadable file, bad YAML, schema violation,
+  duplicate level id, an order pool naming an unknown recipe — is collected
+  and raised as one `ContentError`, not one per run.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Mapping
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Final
+
+from pydantic import ValidationError
+
+from undercooked.sim.grid import GridError
+from undercooked.sim.level import Level, LevelDef
+from undercooked.sim.recipe import (
+    ContentError,
+    RecipeBook,
+    describe_validation_error,
+    load_recipe_book,
+    read_yaml_mapping,
+)
+
+__all__ = [
+    "LEVEL_FILE_SUFFIXES",
+    "ContentBundle",
+    "ContentError",
+    "load_content",
+    "load_level_file",
+]
+
+#: File extensions a level file may use (lower-cased comparison).
+LEVEL_FILE_SUFFIXES: Final = frozenset({".yaml", ".yml"})
+
+#: Drafts and fixture overrides opt out of loading by starting their name with this.
+_IGNORED_PREFIX: Final = "_"
+
+#: The recipe book lives beside the levels but is not itself a level file.
+_RECIPE_BASENAMES: Final = frozenset({"recipes.yaml", "recipes.yml"})
+
+
+@dataclass(frozen=True)
+class ContentBundle:
+    """Everything the game reads about content: the recipe book and the levels.
+
+    `levels` is in campaign order; `by_id` is derived from it and can never
+    disagree. Duplicate level ids are refused here as well as in
+    `load_content`, so a hand-built bundle cannot smuggle one through.
+    """
+
+    recipe_book: RecipeBook
+    levels: tuple[Level, ...]
+    by_id: Mapping[str, Level] = field(init=False)
+
+    def __post_init__(self) -> None:
+        counts: dict[str, int] = {}
+        for level in self.levels:
+            counts[level.id] = counts.get(level.id, 0) + 1
+        duplicates = sorted(level_id for level_id, seen in counts.items() if seen > 1)
+        if duplicates:
+            msg = "duplicate level ids: " + ", ".join(duplicates)
+            raise ContentError(msg)
+        object.__setattr__(self, "by_id", {level.id: level for level in self.levels})
+
+
+def load_level_file(path: Path | str) -> Level:
+    """Parse, validate, and build the single level defined by one YAML file.
+
+    Raises:
+        ContentError: the file is unreadable, is not valid YAML, violates the
+            `LevelDef` schema (every field problem is named), or names tiles
+            that cannot build a grid.
+    """
+    file = Path(path)
+    data = read_yaml_mapping(file)
+    try:
+        spec = LevelDef.model_validate(data)
+    except ValidationError as exc:
+        raise ContentError(describe_validation_error(exc, str(file))) from exc
+    try:
+        return Level.build(spec)
+    except GridError as exc:
+        raise ContentError(f"{file}: {exc}") from exc
+
+
+def _level_files(levels_dir: Path) -> tuple[Path, ...]:
+    """Every candidate level file under `levels_dir`, in stable sorted order."""
+    if not levels_dir.is_dir():
+        msg = f"levels directory {levels_dir} does not exist"
+        raise ContentError(msg)
+    candidates = (
+        path
+        for path in levels_dir.rglob("*")
+        if path.is_file()
+        and path.suffix.lower() in LEVEL_FILE_SUFFIXES
+        and not _is_draft(levels_dir, path)
+        and path.name not in _RECIPE_BASENAMES
+    )
+    return tuple(sorted(candidates, key=str))
+
+
+def _is_draft(levels_dir: Path, path: Path) -> bool:
+    """True when the file or any ancestor directory opts out with a `_` prefix."""
+    return any(
+        part.startswith(_IGNORED_PREFIX)
+        for part in path.relative_to(levels_dir).parts
+    )
+
+
+def _unknown_pool_recipes(spec: LevelDef, book: RecipeBook) -> list[str]:
+    """Order-pool ids the recipe book does not define (sorted, deduplicated)."""
+    return sorted({recipe_id for recipe_id in spec.orders.pool if recipe_id not in book.recipes})
+
+
+def load_content(levels_dir: Path | str, recipes_path: Path | str) -> ContentBundle:
+    """Load the recipe book and every level file into a campaign-ordered bundle.
+
+    Args:
+        levels_dir: directory holding `*.yaml` level files (searched
+            recursively); missing is a `ContentError`, empty is a legal empty
+            campaign (until phase 02 task 04 seeds it).
+        recipes_path: the recipes YAML (`levels/recipes.yaml` by convention),
+            passed explicitly so tests and tools can point elsewhere.
+
+    Raises:
+        ContentError: every failure found across the whole run — reading,
+            parsing, schema, unknown order-pool recipes, and duplicate level
+            ids — joined into one message, one per problem.
+    """
+    book = load_recipe_book(recipes_path)
+    levels: list[Level] = []
+    errors: list[str] = []
+    for path in _level_files(Path(levels_dir)):
+        try:
+            level = load_level_file(path)
+        except ContentError as exc:
+            errors.append(str(exc))
+            continue
+        unknown = _unknown_pool_recipes(level.spec, book)
+        if unknown:
+            errors.append(f"{path}: orders.pool names unknown recipes: {', '.join(unknown)}")
+            continue
+        levels.append(level)
+    seen_ids: set[str] = set()
+    duplicates: set[str] = set()
+    for level in levels:
+        if level.id in seen_ids:
+            duplicates.add(level.id)
+        seen_ids.add(level.id)
+    if duplicates:
+        errors.append(f"duplicate level ids: {', '.join(sorted(duplicates))}")
+    if errors:
+        raise ContentError("\n".join(errors))
+    ordered = sorted(levels, key=lambda level: (level.order, level.id))
+    return ContentBundle(recipe_book=book, levels=tuple(ordered))

src/undercooked/sim/grid.py

@@ -0,0 +1,146 @@
+"""Grid geometry and walkability for the pure simulation kernel (PLAN §5/§8).
+
+Purity contract (PLAN.md A5, enforced by `undercooked.sim.purity`): standard
+library and `undercooked.sim` imports only — no I/O, no asyncio, no clock.
+Cells are indexed row-major as ``y * width + x``; tiles are continuous-world
+footprints, while player positions stay `Vec2` float values.
+"""
+
+from __future__ import annotations
+
+from collections import deque
+from collections.abc import Mapping, Sequence
+from dataclasses import dataclass, field
+from typing import Final
+
+from undercooked.sim.model import WALKABLE, TileKind, Vec2
+
+__all__ = ["MAX_GRID", "Grid", "GridError"]
+
+#: Maximum grid width/height in tiles; larger maps are a validation error.
+MAX_GRID: Final = 64
+
+# Deterministic neighbour orders, as (dx, dy).
+_DIRS4: Final = ((0, -1), (-1, 0), (1, 0), (0, 1))  # N, W, E, S
+_DIRS8: Final = ((0, -1), (-1, -1), (1, -1), (-1, 0), (1, 0), (-1, 1), (1, 1), (0, 1))
+
+
+class GridError(ValueError):
+    """Raised for malformed grid content or out-of-range cell access."""
+
+
+@dataclass(frozen=True)
+class Grid:
+    """Rectangular tile map: `tiles` is row-major, `station_at` maps cell -> station id."""
+
+    width: int
+    height: int
+    tiles: tuple[TileKind, ...]
+    station_at: Mapping[int, str] = field(default_factory=dict)
+
+    def __post_init__(self) -> None:
+        if not 0 < self.width <= MAX_GRID or not 0 < self.height <= MAX_GRID:
+            msg = f"grid must be 1..{MAX_GRID} tiles per side, got {self.width}x{self.height}"
+            raise GridError(msg)
+        if len(self.tiles) != self.width * self.height:
+            msg = f"expected {self.width * self.height} tiles, got {len(self.tiles)}"
+            raise GridError(msg)
+
+    @classmethod
+    def from_rows(cls, rows: Sequence[str], legend: Mapping[str, TileKind]) -> Grid:
+        """Parse an ASCII map (PLAN §8 `grid:` + `tiles:` legend).
+
+        Raises:
+            GridError: no rows, ragged rows, unknown legend character, or a
+                size of 0 / beyond `MAX_GRID`.
+        """
+        if len(rows) == 0:
+            msg = "grid has no rows"
+            raise GridError(msg)
+        width = len(rows[0])
+        tiles: list[TileKind] = []
+        for y, row in enumerate(rows):
+            if len(row) != width:
+                msg = f"ragged grid: row {y} is {len(row)} chars, expected {width}"
+                raise GridError(msg)
+            for x, char in enumerate(row):
+                kind = legend.get(char)
+                if kind is None:
+                    msg = f"unknown legend char {char!r} at ({x},{y})"
+                    raise GridError(msg)
+                tiles.append(kind)
+        return cls(width=width, height=len(rows), tiles=tuple(tiles))
+
+    def in_bounds(self, x: int, y: int) -> bool:
+        return 0 <= x < self.width and 0 <= y < self.height
+
+    def tile_at(self, x: int, y: int) -> TileKind | None:
+        """Tile at (x, y), or None when the coordinate is outside the map."""
+        if not self.in_bounds(x, y):
+            return None
+        return self.tiles[y * self.width + x]
+
+    def is_walkable(self, x: int, y: int) -> bool:
+        """True only for an in-bounds tile the cooks can stand on."""
+        tile = self.tile_at(x, y)
+        return tile is not None and tile in WALKABLE
+
+    def walkable_cells(self) -> tuple[int, ...]:
+        """Walkable cell indices in deterministic row-major order."""
+        return tuple(index for index, tile in enumerate(self.tiles) if tile in WALKABLE)
+
+    def blocking_mask(self) -> tuple[bool, ...]:
+        """Per-cell True where cooks cannot stand (wall, counter, void, off-map)."""
+        return tuple(tile not in WALKABLE for tile in self.tiles)
+
+    def neighbours(self, cell: int, four_way: bool = True) -> tuple[int, ...]:
+        """In-bounds 4- (default) or 8-connected neighbours of `cell`.
+
+        Geometry only — walkability is the caller's filter (see `bfs_reachable`).
+        Raises GridError if `cell` is not a valid index.
+        """
+        if not 0 <= cell < len(self.tiles):
+            msg = f"cell {cell} out of range for a {len(self.tiles)}-cell grid"
+            raise GridError(msg)
+        x, y = cell % self.width, cell // self.width
+        directions = _DIRS4 if four_way else _DIRS8
+        return tuple(
+            (y + dy) * self.width + (x + dx)
+            for dx, dy in directions
+            if self.in_bounds(x + dx, y + dy)
+        )
+
+    def bfs_reachable(
+        self, start: int, extra_blocked: frozenset[int] = frozenset()
+    ) -> frozenset[int]:
+        """Every walkable cell 4-connected to `start`, including `start` itself.
+
+        `extra_blocked` treats otherwise-walkable cells as impassable (mover
+        exclusion in the phase-02 validator). A start that is not walkable (or
+        is itself extra-blocked) reaches nothing; a start outside the map is a
+        GridError.
+        """
+        if not 0 <= start < len(self.tiles):
+            msg = f"cell {start} out of range for a {len(self.tiles)}-cell grid"
+            raise GridError(msg)
+        if start in extra_blocked or self.tiles[start] not in WALKABLE:
+            return frozenset()
+        seen = {start}
+        queue: deque[int] = deque([start])
+        while queue:
+            cell = queue.popleft()
+            for neighbour in self.neighbours(cell):
+                if (
+                    neighbour in seen
+                    or neighbour in extra_blocked
+                    or self.tiles[neighbour] not in WALKABLE
+                ):
+                    continue
+                seen.add(neighbour)
+                queue.append(neighbour)
+        return frozenset(seen)
+
+    @staticmethod
+    def manhattan(a: Vec2, b: Vec2) -> float:
+        """Tile distance (PLAN §8 escalation axis 7)."""
+        return abs(a.x - b.x) + abs(a.y - b.y)

src/undercooked/sim/level.py

@@ -0,0 +1,354 @@
+"""Level content: the YAML schema, the ASCII-grid legend, and the built `Level`.
+
+Purity contract (PLAN.md A5, enforced by `undercooked.sim.purity`): standard
+library plus `pydantic`/`PyYAML` and `undercooked.sim` imports only — no
+asyncio, no sockets, no wall-clock, no FastAPI. These models mirror PLAN §8:
+a level file *is* a `LevelDef`, and `build_grid`/`Level.build` turn it into
+the grid geometry the simulation and the validator walk. The schema reference
+authors write against is `docs/LEVELS.md`.
+
+Legend conventions (all enforced by the validators below):
+
+* every character used by the grid needs a `tiles:` entry, and every legend
+  entry must be used at least once — dead legend chars are a content bug;
+* stations stand on walkable `floor` tiles (a station on wall/counter/void is
+  unreachable; task 03's rules reject it);
+* exactly one legend character per level maps to the `serve` station;
+* station ids are deterministic — `f"{kind.value}_{n}"` (or the legend's
+  `id_prefix` instead of the kind) with `n` counting up in reading order
+  (row-major) across every tile of that kind, so hazards can reference
+  `stove_2` from the file.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Iterator, Mapping
+from dataclasses import dataclass, field
+
+from pydantic import Field, model_validator
+
+from undercooked.sim.grid import MAX_GRID, Grid
+from undercooked.sim.model import HazardKind, ItemKind, Station, StationKind, TileKind
+from undercooked.sim.recipe import StrictModel
+
+__all__ = [
+    "HazardSpec",
+    "Level",
+    "LevelDef",
+    "OrdersSpec",
+    "StationLimits",
+    "TileSpec",
+    "build_grid",
+    "build_station_items",
+    "build_stations",
+]
+
+
+class TileSpec(StrictModel):
+    """One `tiles:` legend entry: what a grid character *is*.
+
+    `kind` is the terrain; `station` (when set) puts a fixture on that tile;
+    `item` is what a spawn shelf replenishes; `id_prefix` replaces the station
+    kind in generated station ids for this character.
+    """
+
+    kind: TileKind
+    station: StationKind | None = None
+    item: ItemKind | None = None
+    id_prefix: str | None = Field(default=None, min_length=1)
+
+    @model_validator(mode="after")
+    def _station_rules(self) -> TileSpec:
+        problems: list[str] = []
+        if self.item is not None and self.station is not StationKind.SPAWN_SHELF:
+            problems.append("only a spawn_shelf tile may name an item")
+        if self.station is StationKind.SPAWN_SHELF and self.item is None:
+            problems.append("a spawn_shelf tile must name the item it spawns")
+        if self.station is None and self.id_prefix is not None:
+            problems.append("id_prefix belongs to a station tile")
+        if problems:
+            raise ValueError("; ".join(problems))
+        return self
+
+
+class OrdersSpec(StrictModel):
+    """How orders arrive: what may be ordered and how fast (PLAN §8 `orders:`)."""
+
+    pool: list[str] = Field(min_length=1)
+    concurrent: int = Field(ge=1)
+    spawn_interval: tuple[float, float]
+    deadline: float = Field(gt=0)
+    penalty: int = Field(default=10, ge=0)
+
+    @model_validator(mode="after")
+    def _spawn_interval_ordered(self) -> OrdersSpec:
+        low, high = self.spawn_interval
+        problems: list[str] = []
+        if low <= 0:
+            problems.append(f"spawn_interval must start above 0 s, got {low}")
+        if low > high:
+            problems.append(f"spawn_interval start {low} s is greater than its end {high} s")
+        if problems:
+            raise ValueError("; ".join(problems))
+        return self
+
+
+class HazardSpec(StrictModel):
+    """One authored hazard (PLAN §8 `hazards:`), placed or path-walking.
+
+    Placement is `station` XOR `x`/`y`: never both, and the pair must come
+    together. `fire` and `spill` additionally need a location;
+    `moving_obstacle` needs a `path` of at least 2 points and a `period_s`;
+    `rush` is global and needs a positive `intensity`.
+    """
+
+    type: HazardKind
+    x: int | None = None
+    y: int | None = None
+    station: str | None = None
+    chance_per_s: float | None = Field(default=None, ge=0.0)
+    respawn_s: float | None = Field(default=None, gt=0.0)
+    path: list[tuple[int, int]] | None = None
+    period_s: float | None = Field(default=None, gt=0.0)
+    intensity: float = 1.0
+
+    @model_validator(mode="after")
+    def _placement_rules(self) -> HazardSpec:
+        problems: list[str] = []
+        has_coords = self.x is not None and self.y is not None
+        if (self.x is None) != (self.y is None):
+            problems.append("x and y must be given together")
+        if self.station is not None and has_coords:
+            problems.append("a hazard takes either station or x/y, not both")
+        if self.type in (HazardKind.FIRE, HazardKind.SPILL) and not (
+            self.station is not None or has_coords
+        ):
+            problems.append(f"a {self.type.value} hazard needs a station or x/y coordinates")
+        if self.type is HazardKind.MOVING_OBSTACLE:
+            if self.path is None or len(self.path) < 2:
+                problems.append("a moving_obstacle hazard needs a path of at least 2 points")
+            if self.period_s is None:
+                problems.append("a moving_obstacle hazard needs period_s")
+        if self.type is HazardKind.RUSH and self.intensity <= 0:
+            problems.append("a rush hazard needs a positive intensity")
+        if problems:
+            raise ValueError("; ".join(problems))
+        return self
+
+
+class StationLimits(StrictModel):
+    """How many of each shared prop may be out of the sink at once.
+
+    Defaults are effectively unlimited (99); a campaign level tightens them
+    (PLAN §8 escalation axis 4). Zero is schema-legal so the validator
+    (task 03's `STATION_LIMIT_ZERO`) reports it as a named rule, not a crash.
+    """
+
+    plates: int = Field(default=99, ge=0)
+    pots: int = Field(default=99, ge=0)
+
+
+class LevelDef(StrictModel):
+    """One level file, exactly as authored (PLAN §8), before geometry is built.
+
+    Validation here is what makes a content file *well-formed*: rectangular
+    grid within `MAX_GRID`, complete legend both directions, at least one
+    spawn inside the map, one serve legend char, sane numbers. Playability
+    (reachability, solvability, target plausibility) is `sim/validation.py`.
+    """
+
+    id: str = Field(min_length=1)
+    name: str = Field(min_length=1)
+    order: int = Field(ge=1)
+    time_limit: float = Field(gt=0)
+    target_score: int = Field(gt=0)
+    grid: list[str] = Field(min_length=1)
+    tiles: dict[str, TileSpec] = Field(min_length=1)
+    spawns: list[tuple[int, int]]
+    orders: OrdersSpec
+    hazards: list[HazardSpec] = Field(default_factory=list)
+    station_limits: StationLimits = Field(default_factory=StationLimits)
+    difficulty: int = Field(default=1, ge=1, le=10)
+    blurb: str = ""
+    escalation_notes: str = ""
+
+    @model_validator(mode="after")
+    def _grid_is_a_complete_rectangle(self) -> LevelDef:
+        problems: list[str] = []
+        width = len(self.grid[0])
+        if width == 0:
+            problems.append("grid rows must not be empty")
+        elif width > MAX_GRID or len(self.grid) > MAX_GRID:
+            problems.append(
+                f"grid must be at most {MAX_GRID}x{MAX_GRID}, got {width}x{len(self.grid)}"
+            )
+        for row_index, row in enumerate(self.grid):
+            if len(row) != width:
+                problems.append(f"ragged grid: row {row_index} is {len(row)} chars, want {width}")
+        used = set("".join(self.grid))
+        problems.extend(
+            f"grid character {char!r} has no tiles: legend entry"
+            for char in sorted(used - set(self.tiles))
+        )
+        problems.extend(
+            f"legend character {char!r} is never used by the grid"
+            for char in sorted(set(self.tiles) - used)
+        )
+        if problems:
+            raise ValueError("; ".join(problems))
+        return self
+
+    @model_validator(mode="after")
+    def _spawns_are_inside(self) -> LevelDef:
+        problems: list[str] = []
+        if not self.spawns:
+            problems.append("a level needs at least one spawn")
+        width = len(self.grid[0])
+        height = len(self.grid)
+        problems.extend(
+            f"spawn ({x},{y}) is outside the {width}x{height} grid"
+            for x, y in self.spawns
+            if not (0 <= x < width and 0 <= y < height)
+        )
+        if problems:
+            raise ValueError("; ".join(problems))
+        return self
+
+    @model_validator(mode="after")
+    def _exactly_one_serve(self) -> LevelDef:
+        serve_chars = sorted(
+            char for char, spec in self.tiles.items() if spec.station is StationKind.SERVE
+        )
+        if len(serve_chars) != 1:
+            msg = (
+                f"exactly one legend character must map to the serve station, "
+                f"found {len(serve_chars)}: {serve_chars}"
+            )
+            raise ValueError(msg)
+        return self
+
+
+def _station_tiles(level: LevelDef) -> Iterator[tuple[int, Station, TileSpec]]:
+    """Yield (cell index, Station, TileSpec) for every station tile in reading order.
+
+    The per-kind id counter runs across the whole level in reading order, so
+    ids are deterministic for a given file and stable under legend reordering.
+    """
+    width = len(level.grid[0])
+    counters: dict[StationKind, int] = dict.fromkeys(StationKind, 0)
+    for y, row in enumerate(level.grid):
+        for x, char in enumerate(row):
+            spec = level.tiles[char]
+            if spec.station is None:
+                continue
+            kind = spec.station
+            number = counters[kind]
+            counters[kind] += 1
+            prefix = spec.id_prefix if spec.id_prefix is not None else kind.value
+            station = Station(id=f"{prefix}_{number}", kind=kind, x=float(x), y=float(y))
+            yield y * width + x, station, spec
+
+
+def build_grid(level: LevelDef) -> Grid:
+    """Legend + ASCII rows -> `Grid`, with `station_at` for every station cell.
+
+    The `LevelDef` validators already guarantee rectangularity and a complete
+    legend, so `Grid.from_rows` cannot fail here for a constructed model; it
+    stays the geometry's single parser (`GridError` remains possible by hand).
+    """
+    legend = {char: spec.kind for char, spec in level.tiles.items()}
+    parsed = Grid.from_rows(level.grid, legend)
+    station_at = {cell: station.id for cell, station, _ in _station_tiles(level)}
+    return Grid(
+        width=parsed.width,
+        height=parsed.height,
+        tiles=parsed.tiles,
+        station_at=station_at,
+    )
+
+
+def build_stations(level: LevelDef) -> tuple[Station, ...]:
+    """Every station the legend places, in reading order, with deterministic ids."""
+    return tuple(station for _, station, _ in _station_tiles(level))
+
+
+def build_station_items(level: LevelDef) -> dict[str, ItemKind]:
+    """station id -> the item its spawn shelf replenishes (shelves only)."""
+    return {
+        station.id: spec.item
+        for _, station, spec in _station_tiles(level)
+        if spec.item is not None
+    }
+
+
+@dataclass(frozen=True)
+class Level:
+    """A `LevelDef` with its geometry built: grid, stations, spawns.
+
+    `Level.build(spec)` is the only way content should become a `Level` — it
+    runs `build_grid`/`build_stations` so ids, `station_at`, and
+    `station_items` (spawn shelf -> item) agree with the file. The delegated
+    properties keep callers reading `level.target_score` instead of
+    `level.spec.target_score`.
+    """
+
+    spec: LevelDef
+    grid: Grid
+    stations: tuple[Station, ...]
+    spawns: tuple[tuple[int, int], ...]
+    station_items: Mapping[str, ItemKind] = field(default_factory=dict)
+
+    @classmethod
+    def build(cls, spec: LevelDef) -> Level:
+        """Build the geometry from a validated definition."""
+        return cls(
+            spec=spec,
+            grid=build_grid(spec),
+            stations=build_stations(spec),
+            spawns=tuple(spec.spawns),
+            station_items=build_station_items(spec),
+        )
+
+    @property
+    def id(self) -> str:
+        return self.spec.id
+
+    @property
+    def name(self) -> str:
+        return self.spec.name
+
+    @property
+    def order(self) -> int:
+        return self.spec.order
+
+    @property
+    def difficulty(self) -> int:
+        return self.spec.difficulty
+
+    @property
+    def time_limit(self) -> float:
+        return self.spec.time_limit
+
+    @property
+    def target_score(self) -> int:
+        return self.spec.target_score
+
+    @property
+    def orders(self) -> OrdersSpec:
+        return self.spec.orders
+
+    @property
+    def hazards(self) -> list[HazardSpec]:
+        return self.spec.hazards
+
+    @property
+    def station_limits(self) -> StationLimits:
+        return self.spec.station_limits
+
+    @property
+    def blurb(self) -> str:
+        return self.spec.blurb
+
+    @property
+    def escalation_notes(self) -> str:
+        return self.spec.escalation_notes

src/undercooked/sim/model.py

@@ -0,0 +1,408 @@
+"""Core simulation data model — the vocabulary every later layer shares (PLAN §5).
+
+Purity contract (PLAN.md A5, enforced by `undercooked.sim.purity`): this module
+imports nothing outside the standard library and `undercooked.sim`. No FastAPI,
+no asyncio, no sockets, no wall-clock, no I/O, and no RNG outside the round
+seed. Every entity is a frozen dataclass; updates go through `with_*` helpers
+(or `dataclasses.replace`) so state is never mutated in place and a replay of
+the same seed + input script reproduces identical state bit-for-bit.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import math
+from collections.abc import Mapping
+from dataclasses import dataclass, field, fields, is_dataclass, replace
+from enum import StrEnum
+from types import MappingProxyType
+from typing import TYPE_CHECKING, Final
+
+if TYPE_CHECKING:  # imported for types only; avoids a runtime cycle with grid.py
+    from undercooked.sim.grid import Grid
+
+__all__ = [
+    "OPPOSITE",
+    "WALKABLE",
+    "Event",
+    "EventKind",
+    "Facing",
+    "Hazard",
+    "HazardKind",
+    "Item",
+    "ItemForm",
+    "ItemKind",
+    "KitchenState",
+    "Order",
+    "Player",
+    "PlayerState",
+    "RoundResult",
+    "Station",
+    "StationKind",
+    "StationState",
+    "TileKind",
+    "Vec2",
+]
+
+
+@dataclass(frozen=True)
+class Vec2:
+    """A continuous point/offset in tile units (positions are not cell-snapped)."""
+
+    x: float = 0.0
+    y: float = 0.0
+
+    def __add__(self, other: Vec2) -> Vec2:
+        return Vec2(self.x + other.x, self.y + other.y)
+
+    def __sub__(self, other: Vec2) -> Vec2:
+        return Vec2(self.x - other.x, self.y - other.y)
+
+    def scale(self, factor: float) -> Vec2:
+        return Vec2(self.x * factor, self.y * factor)
+
+    @property
+    def length(self) -> float:
+        return math.hypot(self.x, self.y)
+
+    def normalized(self) -> Vec2:
+        """Unit vector in the same direction; the zero vector normalizes to itself."""
+        length = self.length
+        if length == 0.0:
+            return Vec2(0.0, 0.0)
+        return self.scale(1.0 / length)
+
+    def quantized(self, digits: int = 2) -> Vec2:
+        """Rounded copy — the wire encoder's canonical form (PLAN §4 snapshots)."""
+        return Vec2(round(self.x, digits), round(self.y, digits))
+
+    @property
+    def KEY(self) -> tuple[int, int]:  # noqa: N802 - fixed spatial-key vocabulary
+        """Spatial hash key at 1/64-tile resolution (bucketing, dedup)."""
+        return (round(self.x * 64), round(self.y * 64))
+
+
+class Facing(StrEnum):
+    """The four directions; row 0 is the top of the map, so UP decreases y."""
+
+    UP = "up"
+    RIGHT = "right"
+    DOWN = "down"
+    LEFT = "left"
+
+    def delta(self) -> Vec2:
+        """Unit step for this direction in tile space."""
+        return _FACING_DELTA[self]
+
+
+_FACING_DELTA: Final[Mapping[Facing, Vec2]] = MappingProxyType(
+    {
+        Facing.UP: Vec2(0.0, -1.0),
+        Facing.RIGHT: Vec2(1.0, 0.0),
+        Facing.DOWN: Vec2(0.0, 1.0),
+        Facing.LEFT: Vec2(-1.0, 0.0),
+    }
+)
+
+OPPOSITE: Final[Mapping[Facing, Facing]] = MappingProxyType(
+    {
+        Facing.UP: Facing.DOWN,
+        Facing.DOWN: Facing.UP,
+        Facing.LEFT: Facing.RIGHT,
+        Facing.RIGHT: Facing.LEFT,
+    }
+)
+
+
+class TileKind(StrEnum):
+    """Terrain under a grid cell (ASCII legend characters arrive with task 02)."""
+
+    FLOOR = "floor"
+    COUNTER = "counter"
+    WALL = "wall"
+    VOID = "void"
+
+
+WALKABLE: Final[frozenset[TileKind]] = frozenset({TileKind.FLOOR})
+
+
+class StationKind(StrEnum):
+    """Kitchen fixture types (PLAN §5 station pipelines)."""
+
+    SPAWN_SHELF = "spawn_shelf"
+    CHOP_BOARD = "chop_board"
+    STOVE = "stove"
+    PREP = "prep"
+    SERVE = "serve"
+    SINK = "sink"
+    TRASH = "trash"
+
+
+class StationState(StrEnum):
+    """Lifecycle of a station."""
+
+    IDLE = "idle"
+    WORKING = "working"
+    READY = "ready"
+    BUSY = "busy"
+    DIRTY = "dirty"
+    BLOCKED = "blocked"
+
+
+class ItemKind(StrEnum):
+    """Ingredient/dish types; recipe content decides which exist per level."""
+
+    TOMATO = "tomato"
+    ONION = "onion"
+    LETTUCE = "lettuce"
+    DOUGH = "dough"
+    PATTY = "patty"
+    RICE = "rice"
+    FISH = "fish"
+    PLATE = "plate"
+
+
+class ItemForm(StrEnum):
+    """Preparation stage of an item."""
+
+    RAW = "raw"
+    CHOPPED = "chopped"
+    COOKING = "cooking"
+    COOKED = "cooked"
+    BURNT = "burnt"
+    PLATED = "plated"
+    DIRTY = "dirty"
+
+
+@dataclass(frozen=True)
+class Item:
+    """A carried or station-held ingredient/dish with value semantics."""
+
+    kind: ItemKind
+    form: ItemForm
+    quality: float = 1.0  # 1.0 -> 0.0 as it burns (PLAN §5)
+    ttl: float = 0.0  # seconds until the item rots; 0.0 = never
+    dish_id: int | None = None  # recipe id when this item is a plated dish
+
+    def is_trash(self) -> bool:
+        """Burnt items are unrecoverable and belong in the trash."""
+        return self.form is ItemForm.BURNT
+
+    def with_form(self, form: ItemForm) -> Item:
+        """A new Item in `form`; never mutates this one (value semantics)."""
+        return replace(self, form=form)
+
+
+@dataclass(frozen=True)
+class Station:
+    """One kitchen fixture at (x, y) with its own pipeline state."""
+
+    id: str
+    kind: StationKind
+    x: float = 0.0
+    y: float = 0.0
+    contents: tuple[Item, ...] = ()
+    state: StationState = StationState.IDLE
+    progress: float = 0.0  # 0.0 -> 1.0 while WORKING
+    duration: float = 0.0  # total seconds the current job takes
+    cooldown: float = 0.0  # seconds before the station can be used again
+    worker_id: str | None = None  # player currently working this station
+    dirty: bool = False  # needs the sink before reuse
+
+
+class PlayerState(StrEnum):
+    """What a player's cook is doing this tick."""
+
+    IDLE = "idle"
+    MOVING = "moving"
+    USING = "using"
+    STUNNED = "stunned"
+    DISCONNECTED = "disconnected"
+
+
+@dataclass(frozen=True)
+class Player:
+    """One seat in the kitchen; `id` is the session token, `seat` is stable."""
+
+    id: str
+    pos: Vec2 = Vec2()
+    vel: Vec2 = Vec2()
+    facing: Facing = Facing.DOWN
+    carrying: Item | None = None
+    state: PlayerState = PlayerState.IDLE
+    stun: float = 0.0
+    score: int = 0
+    dishes_served: int = 0
+    connected: bool = True
+    seat: int = 0
+
+
+@dataclass(frozen=True)
+class Order:
+    """A live customer order counting down to `deadline`."""
+
+    id: int
+    recipe_id: str
+    time_left: float = 0.0
+    deadline: float = 0.0
+    value: int = 0
+
+
+class HazardKind(StrEnum):
+    """Hazard families (PLAN §8 escalation axis 6)."""
+
+    FIRE = "fire"
+    SPILL = "spill"
+    MOVING_OBSTACLE = "moving_obstacle"
+    RUSH = "rush"
+
+
+@dataclass(frozen=True)
+class Hazard:
+    """One active hazard instance on the floor or attached to a station."""
+
+    id: str
+    kind: HazardKind
+    pos: Vec2 = Vec2()
+    life: float = 0.0  # seconds remaining; 0.0 = permanent until removed
+    path: tuple[Vec2, ...] = ()  # waypoints for MOVING_OBSTACLE
+    period_s: float = 0.0  # traversal period for MOVING_OBSTACLE
+    station_id: str | None = None  # station a FIRE burns on
+    chance_per_s: float = 0.0  # spawn/respawn probability for the spawner
+
+
+class RoundResult(StrEnum):
+    """Terminal outcome of a round."""
+
+    WON = "won"
+    LOST = "lost"
+    ABORTED = "aborted"
+
+
+_FLOAT_HASH_DIGITS: Final = 6
+
+
+def _canonical(value: object) -> str:
+    """Deterministic repr used by `KitchenState.state_hash`.
+
+    Mapping keys are sorted and floats are quantized to `_FLOAT_HASH_DIGITS`
+    (folding -0.0 into 0.0), so cosmetic differences cannot change the hash
+    while any real divergence does. Raises TypeError on anything unhashable-by-
+    contract so a new field can never be silently skipped.
+    """
+    if isinstance(value, bool):
+        out = f"b:{int(value)}"
+    elif isinstance(value, float):
+        rounded = round(value, _FLOAT_HASH_DIGITS) + 0.0  # +0.0 folds -0.0 to 0.0
+        out = f"f:{rounded:.{_FLOAT_HASH_DIGITS}f}"
+    elif isinstance(value, int):
+        out = f"i:{value}"
+    elif isinstance(value, str):  # covers every StrEnum in this module
+        out = f"s:{value}"
+    elif value is None:
+        out = "n:"
+    elif isinstance(value, Mapping):
+        items = sorted(value.items(), key=lambda kv: (_canonical(kv[0]), _canonical(kv[1])))
+        out = "m:{" + ",".join(f"{_canonical(k)}={_canonical(v)}" for k, v in items) + "}"
+    elif isinstance(value, (tuple, list)):
+        out = "t:[" + ",".join(_canonical(item) for item in value) + "]"
+    elif is_dataclass(value) and not isinstance(value, type):
+        parts = ",".join(
+            f"{field.name}={_canonical(getattr(value, field.name))}"
+            for field in sorted(fields(value), key=lambda f: f.name)
+        )
+        out = f"d:{type(value).__name__}({parts})"
+    else:
+        msg = f"cannot canonicalise {type(value).__name__} for state_hash"
+        raise TypeError(msg)
+    return out
+
+
+class EventKind(StrEnum):
+    """Transient FX/log event kinds emitted by `step()` (PLAN §4 data flow)."""
+
+    CHOPPED = "chopped"
+    SERVED = "served"
+    BURNT = "burnt"
+    ORDER_MISSED = "order_missed"
+    FIRE_STARTED = "fire_started"
+    DENIED = "denied"
+    SCORE = "score"
+    STUNNED = "stunned"
+    WASHED = "washed"
+    TRASHED = "trashed"
+    TICK_SLOW = "tick_slow"
+
+
+@dataclass(frozen=True)
+class Event:
+    """A one-shot event attached to a position; FX/log only, never game truth."""
+
+    kind: str
+    x: float = 0.0
+    y: float = 0.0
+    payload: Mapping[str, object] = field(default_factory=dict)
+
+
+@dataclass(frozen=True)
+class KitchenState:
+    """Everything that is true about one round at one tick (PLAN §5).
+
+    Frozen: phases 03/04 advance the round by *returning* a new state from
+    `step()`; the `with_*` helpers keep every mutation in one obvious place.
+    """
+
+    level_id: str
+    seed: int
+    grid: Grid
+    tick: int = 0
+    stations: tuple[Station, ...] = ()
+    players: tuple[Player, ...] = ()
+    orders: tuple[Order, ...] = ()
+    hazards: tuple[Hazard, ...] = ()
+    score: int = 0
+    target_score: int = 0
+    time_left: float = 0.0
+    streak: int = 0
+    result: RoundResult | None = None
+    counters: Mapping[str, int] = field(default_factory=dict)
+    rng_state: int = 0
+
+    def state_hash(self) -> str:
+        """sha256 over the canonical form of every field (A21 replay guard).
+
+        Cosmetic differences — mapping insertion order, float repr noise,
+        -0.0 — never change the hash; any scored change does.
+        """
+        return hashlib.sha256(_canonical(self).encode("utf-8")).hexdigest()
+
+    def with_tick(self, tick: int) -> KitchenState:
+        return replace(self, tick=tick)
+
+    def with_time_left(self, seconds: float) -> KitchenState:
+        return replace(self, time_left=seconds)
+
+    def with_score(self, score: int, streak: int | None = None) -> KitchenState:
+        """New score, and optionally a new streak, in one atomic replacement."""
+        if streak is None:
+            return replace(self, score=score)
+        return replace(self, score=score, streak=streak)
+
+    def with_result(self, result: RoundResult | None) -> KitchenState:
+        return replace(self, result=result)
+
+    def with_stations(self, stations: tuple[Station, ...]) -> KitchenState:
+        return replace(self, stations=stations)
+
+    def with_players(self, players: tuple[Player, ...]) -> KitchenState:
+        return replace(self, players=players)
+
+    def with_orders(self, orders: tuple[Order, ...]) -> KitchenState:
+        return replace(self, orders=orders)
+
+    def with_hazards(self, hazards: tuple[Hazard, ...]) -> KitchenState:
+        return replace(self, hazards=hazards)
+
+    def with_counters(self, **updates: int) -> KitchenState:
+        """Merge `updates` into a copy of the counters; the original is untouched."""
+        return replace(self, counters={**self.counters, **updates})

src/undercooked/sim/purity.py

@@ -0,0 +1,57 @@
+"""A5 purity guard: `sim/` imports nothing that touches the outside world.
+
+`assert_pure_imports()` statically parses every module in this package (AST,
+so the pytest process's own `asyncio`/`socket` imports cannot poison the
+result) and fails on any import rooted in `FORBIDDEN_ROOTS`. This module's own
+file reads are the guard's single exception — they only ever run from tests.
+"""
+
+from __future__ import annotations
+
+import ast
+from pathlib import Path
+from typing import Final
+
+__all__ = ["FORBIDDEN_ROOTS", "PurityError", "assert_pure_imports", "find_violations"]
+
+#: Root package names no `sim` module may import (PLAN.md A5).
+FORBIDDEN_ROOTS: Final[frozenset[str]] = frozenset(
+    {"asyncio", "fastapi", "socket", "time", "uvicorn"}
+)
+
+SIM_DIR: Final[Path] = Path(__file__).resolve().parent
+
+
+class PurityError(RuntimeError):
+    """Raised when a `sim` module breaks the purity contract."""
+
+
+def _forbidden_roots_in(tree: ast.Module) -> set[str]:
+    """Forbidden import roots appearing anywhere in the module's import statements."""
+    found: set[str] = set()
+    for node in ast.walk(tree):
+        if isinstance(node, ast.Import):
+            found.update(alias.name.split(".")[0] for alias in node.names)
+        elif isinstance(node, ast.ImportFrom) and node.level == 0 and node.module:
+            found.add(node.module.split(".")[0])
+    return found & FORBIDDEN_ROOTS
+
+
+def find_violations(package_dir: Path = SIM_DIR) -> list[str]:
+    """`'<file>: imports <root>'` lines for every forbidden import in the package."""
+    violations: list[str] = []
+    for path in sorted(package_dir.rglob("*.py")):
+        tree = ast.parse(path.read_text(encoding="utf-8"), filename=str(path))
+        violations.extend(
+            f"{path.relative_to(package_dir)}: imports {root}"
+            for root in sorted(_forbidden_roots_in(tree))
+        )
+    return violations
+
+
+def assert_pure_imports(package_dir: Path = SIM_DIR) -> None:
+    """Raise `PurityError` listing every forbidden import found in the package."""
+    violations = find_violations(package_dir)
+    if violations:
+        msg = "sim/ breaks the A5 purity contract: " + "; ".join(violations)
+        raise PurityError(msg)

src/undercooked/sim/recipe.py

@@ -0,0 +1,204 @@
+"""Recipe content: ingredients, steps, and the recipe book (PLAN §5/§8, A8).
+
+Purity contract (PLAN.md A5, enforced by `undercooked.sim.purity`): standard
+library plus `pydantic`/`PyYAML` and `undercooked.sim` imports only — no
+asyncio, no sockets, no wall-clock, no FastAPI. Parsing content files is this
+pipeline's whole job: it happens once at load time and never during `step()`.
+
+Every model here is strict — unknown keys are errors (`extra="forbid"`) and
+assignment re-validates — so a typo in `levels/recipes.yaml` fails loudly at
+startup instead of silently defaulting. `load_recipe_book` turns the YAML file
+into a validated `RecipeBook`; `min_seconds_to_complete` is the greedy solver
+bound used by the target-plausibility validator (task 03) and phase 11's curve
+work. The schema reference players of this file write against is
+`docs/LEVELS.md`.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Final, Literal
+
+import yaml
+from pydantic import BaseModel, ConfigDict, Field, ValidationError, model_validator
+
+from undercooked.sim.model import ItemForm, ItemKind, StationKind
+
+__all__ = [
+    "HANDLING_S",
+    "ContentError",
+    "Ingredient",
+    "Recipe",
+    "RecipeBook",
+    "describe_validation_error",
+    "load_recipe_book",
+    "read_yaml_mapping",
+]
+
+#: Seconds of handling (pick up, put down, plate transfer) added once to every
+#: dish by `RecipeBook.min_seconds_to_complete` — the solver's floor for work
+#: the authored step list does not itself time.
+HANDLING_S: Final = 0.6
+
+
+class ContentError(ValueError):
+    """Content that cannot load: missing file, bad YAML, schema violation, duplicates.
+
+    Raised by the whole content pipeline (`recipe.py`, `level.py`,
+    `content.py`). Its message lists *every* problem found in the run, never
+    just the first, because an author fixing content one error per cycle is a
+    slow author.
+    """
+
+
+class StrictModel(BaseModel):
+    """Base for every content model: no unknown keys, revalidate on assignment."""
+
+    model_config = ConfigDict(extra="forbid", validate_assignment=True)
+
+
+class Ingredient(StrictModel):
+    """One ingredient a recipe consumes: which item, in which form, how many."""
+
+    kind: ItemKind
+    form: ItemForm
+    count: int = Field(default=1, ge=1)
+
+
+class Step(StrictModel):
+    """One authored action of a recipe (PLAN §5 station pipelines).
+
+    `kind` is the interaction, `station` where it happens, and `seconds` the
+    work time it costs (0.0 for instant interactions like `pick`/`serve`).
+    """
+
+    kind: Literal["pick", "chop", "cook", "plate", "serve", "wash"]
+    station: StationKind
+    seconds: float = Field(default=0.0, ge=0.0)
+
+
+class Recipe(StrictModel):
+    """A dish: what it eats, how it is made, and what it is worth."""
+
+    id: str = Field(min_length=1)
+    name: str = Field(min_length=1)
+    base_score: int = Field(ge=0)
+    ingredients: list[Ingredient] = Field(min_length=1)
+    steps: list[Step] = Field(min_length=1)
+    requires_plating: bool = True
+    tags: list[str] = Field(default_factory=list)
+
+
+class RecipeBook(StrictModel):
+    """The parsed `levels/recipes.yaml`: every recipe, keyed by id.
+
+    The mapping key and `Recipe.id` must agree — the key is what level order
+    pools reference, and a silent mismatch would be a level nobody can cook.
+    `ingredients` is the optional top-level catalog of item kinds the content
+    uses; it documents intent and lets tooling catch unknown kinds early.
+    """
+
+    recipes: dict[str, Recipe]
+    ingredients: list[ItemKind] = Field(default_factory=list)
+
+    @model_validator(mode="after")
+    def _keys_match_ids(self) -> RecipeBook:
+        mismatched = sorted(
+            f"{key!r} names recipe {recipe.id!r}"
+            for key, recipe in self.recipes.items()
+            if recipe.id != key
+        )
+        if mismatched:
+            msg = "recipes mapping keys must equal each recipe's id: " + "; ".join(mismatched)
+            raise ValueError(msg)
+        return self
+
+    def get(self, recipe_id: str) -> Recipe:
+        """The recipe with this id.
+
+        Raises:
+            KeyError: the id is not in the book (level pools must only name
+                recipes that exist; this is where that becomes loud).
+        """
+        try:
+            return self.recipes[recipe_id]
+        except KeyError:
+            known = ", ".join(sorted(self.recipes))
+            msg = f"unknown recipe {recipe_id!r} (known: {known})"
+            raise KeyError(msg) from None
+
+    def ids(self) -> tuple[str, ...]:
+        """Every recipe id, in file (declaration) order."""
+        return tuple(self.recipes)
+
+    def required_station_kinds(self, recipe_id: str) -> frozenset[StationKind]:
+        """Every station kind the recipe's steps must touch, deduplicated."""
+        return frozenset(step.station for step in self.get(recipe_id).steps)
+
+    def required_item_kinds(self, recipe_id: str) -> frozenset[ItemKind]:
+        """Every item kind the dish consumes: its ingredients, plus `PLATE`
+        when `requires_plating` — a plated dish is not completable without one.
+        """
+        recipe = self.get(recipe_id)
+        kinds = frozenset(ingredient.kind for ingredient in recipe.ingredients)
+        if recipe.requires_plating:
+            kinds |= {ItemKind.PLATE}
+        return kinds
+
+    def min_seconds_to_complete(self, recipe_id: str) -> float:
+        """Greedy lower bound on cooking the dish: summed step seconds + `HANDLING_S`.
+
+        Never a promise (travel is the validator's job, task 03); it is the
+        per-order floor that keeps `target_score` plausibility checks honest.
+        """
+        recipe = self.get(recipe_id)
+        return sum(step.seconds for step in recipe.steps) + HANDLING_S
+
+
+def describe_validation_error(exc: ValidationError, source: str) -> str:
+    """Flatten a pydantic error into one `<source>: <field.path>: <message>` per problem."""
+    parts: list[str] = []
+    for problem in exc.errors():
+        location = ".".join(str(part) for part in problem["loc"]) or "<root>"
+        parts.append(f"{source}: {location}: {problem['msg']}")
+    return "; ".join(parts)
+
+
+def read_yaml_mapping(path: Path | str) -> dict[str, object]:
+    """Read a content file as a mapping; every failure is a `ContentError`.
+
+    An empty file reads as `{}` so the model reports its own
+    "field required" errors against the same source path.
+    """
+    file = Path(path)
+    try:
+        text = file.read_text(encoding="utf-8")
+    except OSError as exc:
+        reason = exc.strerror or str(exc)
+        msg = f"{file}: cannot read content file: {reason}"
+        raise ContentError(msg) from exc
+    try:
+        data = yaml.safe_load(text)
+    except yaml.YAMLError as exc:
+        msg = f"{file}: invalid YAML: {exc}"
+        raise ContentError(msg) from exc
+    if data is None:
+        return {}
+    if not isinstance(data, dict):
+        msg = f"{file}: expected a YAML mapping at the top level, got {type(data).__name__}"
+        raise ContentError(msg)
+    return data
+
+
+def load_recipe_book(path: Path | str) -> RecipeBook:
+    """Load and validate `levels/recipes.yaml` (or any replacement book).
+
+    Raises:
+        ContentError: the file is missing, is not valid YAML, is not a mapping,
+            or violates the models above — with every field problem named.
+    """
+    data = read_yaml_mapping(path)
+    try:
+        return RecipeBook.model_validate(data)
+    except ValidationError as exc:
+        raise ContentError(describe_validation_error(exc, str(path))) from exc

src/undercooked/sim/validation.py

@@ -0,0 +1,713 @@
+"""Level validation: prove a level is playable before anyone plays it (PLAN §8).
+
+Purity contract (PLAN.md A5, enforced by `undercooked.sim.purity`): standard
+library and `undercooked.sim` imports only — no I/O, no asyncio, no clock.
+The validator consumes built `Level`s plus the `RecipeBook` and reports a
+structured list of **named rules**, so `undercooked-levels validate` (task 05)
+can print exactly which rule a level failed instead of a stack trace.
+
+Design notes:
+
+* every error names its `RuleId` constant and carries a human message with
+  coordinates; `detail` holds the machine-readable companion of the same fact;
+* **warnings never fail a level** — `ValidationReport.ok` reads violations
+  only, and the tight-target band is a warning;
+* reachability is a flood fill (BFS) over walkable tiles from every spawn;
+  `moving_obstacle` hazards are excluded from the blocked set because movers
+  are dynamic — only their path tiles must be valid floor;
+* spawn shelves are currently the only item source the content model can
+  express, so "a station that yields it" means exactly a reachable shelf; if
+  a future station yields ingredients, extend `reachable` production here and
+  no rule id changes.
+"""
+
+from __future__ import annotations
+
+from collections import deque
+from collections.abc import Mapping, Sequence
+from dataclasses import dataclass, field
+from math import floor
+from typing import Final
+
+from undercooked.sim.content import ContentBundle
+from undercooked.sim.grid import Grid
+from undercooked.sim.level import Level
+from undercooked.sim.model import ItemForm, StationKind
+from undercooked.sim.recipe import RecipeBook
+
+__all__ = [
+    "AXIS_EPS",
+    "AXIS_NAMES",
+    "MIN_ESCALATED_AXES",
+    "MIN_TIME_LIMIT_S",
+    "MIN_TRAVEL_TILES",
+    "TARGET_REJECT_FRACTION",
+    "TARGET_WARN_FRACTION",
+    "TRAVEL_MIN_DIFFICULTY",
+    "TRAVEL_SPEED_TILES_PER_S",
+    "TRIP_TRAVERSALS",
+    "LevelViolation",
+    "RuleId",
+    "ValidationReport",
+    "escalation_axes",
+    "reachable_cells",
+    "theoretical_max_score",
+    "validate_bundle",
+    "validate_level",
+    "validate_levels",
+]
+
+#: Shortest shippable round; anything below is a misconfiguration, not a sprint.
+MIN_TIME_LIMIT_S: Final = 30.0
+
+#: `TRAVEL_TOO_SHORT` applies from this declared difficulty upward…
+TRAVEL_MIN_DIFFICULTY: Final = 3
+
+#: …and requires the mean spawn→serve path to be at least this many tiles.
+MIN_TRAVEL_TILES: Final = 3.0
+
+#: Max cook speed in tiles/s (PLAN §5) — converts travel tiles to seconds.
+TRAVEL_SPEED_TILES_PER_S: Final = 3.5
+
+#: One order costs a walk to the serve window and the walk back.
+TRIP_TRAVERSALS: Final = 2.0
+
+#: Above this fraction of the theoretical maximum, a target is rejected…
+TARGET_REJECT_FRACTION: Final = 0.6
+
+#: …and above this (but under the reject line) it warns: tight but fair.
+TARGET_WARN_FRACTION: Final = 0.45
+
+#: A campaign level must move at least this many escalation axes vs its predecessor.
+MIN_ESCALATED_AXES: Final = 2
+
+#: Float noise below this never counts as "an axis moved".
+AXIS_EPS: Final = 1e-9
+
+#: The escalation-axis names in `escalation_axes`' key order — the column order
+#: `undercooked-levels axes` prints, so the table and the rule share one list.
+AXIS_NAMES: Final = (
+    "recipe_complexity",
+    "order_pressure",
+    "time_budget",
+    "station_scarcity",
+    "layout_hostility",
+    "hazard_density",
+    "travel_distance",
+)
+
+
+class RuleId:
+    """Named validation rules (PLAN §8's gate, one constant per rule).
+
+    Values are stable documentation strings — `render()` prints them and the
+    `undercooked-levels` contract reports on them — so never renumber or
+    reuse one; add a new constant instead.
+    """
+
+    GRID_SHAPE: Final = "GRID_SHAPE"
+    SPAWN_ON_FLOOR: Final = "SPAWN_ON_FLOOR"
+    SPAWN_COUNT: Final = "SPAWN_COUNT"
+    STATION_REACHABLE: Final = "STATION_REACHABLE"
+    SERVE_REACHABLE: Final = "SERVE_REACHABLE"
+    SPAWNS_MUTUALLY_REACHABLE: Final = "SPAWNS_MUTUALLY_REACHABLE"
+    RECIPE_STATION_MISSING: Final = "RECIPE_STATION_MISSING"
+    RECIPE_INGREDIENT_UNAVAILABLE: Final = "RECIPE_INGREDIENT_UNAVAILABLE"
+    ORDER_POOL_UNKNOWN_RECIPE: Final = "ORDER_POOL_UNKNOWN_RECIPE"
+    TARGET_IMPLAUSIBLE: Final = "TARGET_IMPLAUSIBLE"
+    TIME_LIMIT_INVALID: Final = "TIME_LIMIT_INVALID"
+    HAZARD_REF_UNKNOWN: Final = "HAZARD_REF_UNKNOWN"
+    HAZARD_ON_UNREACHABLE_TILE: Final = "HAZARD_ON_UNREACHABLE_TILE"
+    STATION_LIMIT_ZERO: Final = "STATION_LIMIT_ZERO"
+    DUPLICATE_ID: Final = "DUPLICATE_ID"
+    DIFFICULTY_ORDER: Final = "DIFFICULTY_ORDER"
+    TRAVEL_TOO_SHORT: Final = "TRAVEL_TOO_SHORT"
+
+    #: Every rule id, in the order the checks run (the report renders sorted).
+    ALL: Final = (
+        GRID_SHAPE,
+        SPAWN_ON_FLOOR,
+        SPAWN_COUNT,
+        STATION_REACHABLE,
+        SERVE_REACHABLE,
+        SPAWNS_MUTUALLY_REACHABLE,
+        RECIPE_STATION_MISSING,
+        RECIPE_INGREDIENT_UNAVAILABLE,
+        ORDER_POOL_UNKNOWN_RECIPE,
+        TARGET_IMPLAUSIBLE,
+        TIME_LIMIT_INVALID,
+        HAZARD_REF_UNKNOWN,
+        HAZARD_ON_UNREACHABLE_TILE,
+        STATION_LIMIT_ZERO,
+        DUPLICATE_ID,
+        DIFFICULTY_ORDER,
+        TRAVEL_TOO_SHORT,
+    )
+
+
+@dataclass(frozen=True)
+class LevelViolation:
+    """One rule hit: a stable `rule` id, the level it refers to, and prose.
+
+    `message` is the human half (coordinates included where the rule has a
+    position); `detail` is the machine half for tooling (task 05's CLI).
+    """
+
+    rule: str  # one of the RuleId constants
+    level_id: str
+    message: str
+    detail: Mapping[str, object] = field(default_factory=dict)
+
+
+@dataclass(frozen=True)
+class ValidationReport:
+    """Errors and warnings for one level or a whole bundle.
+
+    Warnings never fail: `ok` reads `violations` only. `render()` is
+    deterministic (grouped by level, sorted by rule and message, ANSI-free)
+    so CLI output and tests can pin it.
+    """
+
+    violations: tuple[LevelViolation, ...] = ()
+    warnings: tuple[LevelViolation, ...] = ()
+
+    @property
+    def ok(self) -> bool:
+        """True when nothing violated a rule (warnings are allowed)."""
+        return not self.violations
+
+    def errors_for(self, level_id: str) -> tuple[LevelViolation, ...]:
+        """The violations belonging to one level id, in render order."""
+        return tuple(v for v in self._sorted(self.violations) if v.level_id == level_id)
+
+    def render(self) -> str:
+        """Plain text grouped by level: errors first, then warnings."""
+        if self.ok and not self.warnings:
+            return "content OK: no violations"
+        lines = [
+            f"{len(self.violations)} error(s), {len(self.warnings)} warning(s)",
+        ]
+        error_groups: dict[str, list[LevelViolation]] = {}
+        for violation in self._sorted(self.violations):
+            error_groups.setdefault(violation.level_id, []).append(violation)
+        warning_groups: dict[str, list[LevelViolation]] = {}
+        for warning in self._sorted(self.warnings):
+            warning_groups.setdefault(warning.level_id, []).append(warning)
+        for level_id, items in error_groups.items():
+            lines.append(f"level {level_id}")
+            lines.extend(f"  ERROR {v.rule}: {v.message}" for v in items)
+        for level_id, items in warning_groups.items():
+            lines.append(f"level {level_id} (warnings)")
+            lines.extend(f"  WARN {w.rule}: {w.message}" for w in items)
+        return "\n".join(lines)
+
+    @staticmethod
+    def _sorted(items: Sequence[LevelViolation]) -> tuple[LevelViolation, ...]:
+        return tuple(sorted(items, key=lambda item: (item.rule, item.message)))
+
+
+def _add(
+    into: list[LevelViolation],
+    rule: str,
+    level: Level,
+    message: str,
+    **detail: object,
+) -> None:
+    into.append(LevelViolation(rule=rule, level_id=level.id, message=message, detail=dict(detail)))
+
+
+# --- geometry helpers -----------------------------------------------------------
+
+
+def _cell(grid: Grid, x: int, y: int) -> int:
+    return y * grid.width + x
+
+
+def reachable_cells(level: Level) -> frozenset[int]:
+    """Every walkable cell reachable (4-connected) from *any* spawn.
+
+    Movers are dynamic and never block this set; only terrain does. Public
+    because `undercooked-levels show --grid` renders exactly this set as the
+    reachability overlay authors debug a broken layout with.
+    """
+    grid = level.grid
+    cells: set[int] = set()
+    for x, y in level.spawns:
+        if grid.is_walkable(x, y):
+            cells |= grid.bfs_reachable(_cell(grid, x, y))
+    return frozenset(cells)
+
+
+def _serve_cells(level: Level) -> frozenset[int]:
+    grid = level.grid
+    return frozenset(
+        _cell(grid, int(station.x), int(station.y))
+        for station in level.stations
+        if station.kind is StationKind.SERVE
+        and grid.in_bounds(int(station.x), int(station.y))
+    )
+
+
+def _bfs_distance(grid: Grid, start: int, targets: frozenset[int]) -> int | None:
+    """Shortest walkable 4-path length from `start` to any target, or None.
+
+    `start` must be a valid cell index (walkable or not); unwalkable targets
+    are never "reached", which is what lets the caller fall back to a
+    straight-line estimate for broken geometry.
+    """
+    if start in targets:
+        return 0
+    depths: dict[int, int] = {start: 0}
+    queue: deque[int] = deque([start])
+    while queue:
+        cell = queue.popleft()
+        depth = depths[cell]
+        for neighbour in grid.neighbours(cell):
+            if neighbour in depths or not grid.is_walkable(
+                neighbour % grid.width, neighbour // grid.width
+            ):
+                continue
+            if neighbour in targets:
+                return depth + 1
+            depths[neighbour] = depth + 1
+            queue.append(neighbour)
+    return None
+
+
+def _travel_tiles(level: Level) -> tuple[float, ...]:
+    """Per-spawn tile distance to the nearest serve cell.
+
+    BFS over walkable tiles; when the serve window is unreachable from a spawn
+    (the reachability rules report that separately) the straight-line
+    Manhattan distance is used so the metric stays defined. Empty when the
+    level has no serve cell or no spawns.
+    """
+    grid = level.grid
+    targets = _serve_cells(level)
+    if not targets or not level.spawns:
+        return ()
+    distances: list[float] = []
+    for x, y in level.spawns:
+        found: int | None = None
+        if grid.is_walkable(x, y):
+            found = _bfs_distance(grid, _cell(grid, x, y), targets)
+        if found is None:
+            found = min(
+                abs(x - cell % grid.width) + abs(y - cell // grid.width) for cell in targets
+            )
+        distances.append(float(found))
+    return tuple(distances)
+
+
+def _mean_travel_tiles(level: Level) -> float:
+    tiles = _travel_tiles(level)
+    return sum(tiles) / len(tiles) if tiles else 0.0
+
+
+# --- escalation axes (PLAN §8, higher = harder) ----------------------------------
+
+
+def escalation_axes(level: Level) -> dict[str, float]:
+    """The seven escalation axes as numbers, each oriented *higher = harder*.
+
+    Purely geometric/author-declared content: no recipe book, so
+    `recipe_complexity` counts distinct pool entries, not chain length —
+    enough for the "≥ 2 axes moved" comparison and the phase-11 curve pass.
+    """
+    orders = level.orders
+    low, high = orders.spawn_interval
+    mean_interval = (low + high) / 2.0
+    blocking = sum(level.grid.blocking_mask())
+    # Key order is `AXIS_NAMES`; a test pins the two against each other.
+    return {
+        "recipe_complexity": float(len(set(orders.pool))),
+        "order_pressure": orders.concurrent / mean_interval,
+        "time_budget": level.target_score / level.time_limit,
+        "station_scarcity": 1.0 / max(len(level.stations), 1),
+        "layout_hostility": blocking / (level.grid.width * level.grid.height),
+        "hazard_density": float(len(level.hazards)),
+        "travel_distance": _mean_travel_tiles(level),
+    }
+
+
+def theoretical_max_score(level: Level, book: RecipeBook) -> float | None:
+    """Greedy plausibility bound for the level's target (PLAN §8).
+
+    `floor(time_limit / mean_complete) * mean_order_value` where
+    `mean_complete` is the recipe book's per-dish floor averaged over the
+    order pool plus a `travel_bound` of `TRIP_TRAVERSALS` traversals of the
+    mean spawn→serve distance at max cook speed (walk to the window, walk
+    back). Returns None when no bound exists — unknown pool recipes or a time
+    limit below `MIN_TIME_LIMIT_S` — because the dedicated rules report those.
+    """
+    if level.time_limit < MIN_TIME_LIMIT_S:
+        return None
+    pool = [recipe_id for recipe_id in level.orders.pool if recipe_id in book.recipes]
+    if not pool:
+        return None
+    travel_bound = TRIP_TRAVERSALS * _mean_travel_tiles(level) / TRAVEL_SPEED_TILES_PER_S
+    mean_complete = sum(book.min_seconds_to_complete(r) for r in pool) / len(pool) + travel_bound
+    max_orders = floor(level.time_limit / mean_complete)
+    mean_value = sum(book.get(recipe_id).base_score for recipe_id in pool) / len(pool)
+    return float(max_orders) * mean_value
+
+
+# --- per-rule checks --------------------------------------------------------------
+
+
+def _check_grid_shape(level: Level, into: list[LevelViolation]) -> None:
+    """GRID_SHAPE: the built geometry must agree with the spec's ASCII rows."""
+    rows, grid = level.spec.grid, level.grid
+    problems: list[str] = []
+    if len(rows) != grid.height:
+        problems.append(f"spec has {len(rows)} row(s), geometry has {grid.height}")
+    width = len(rows[0]) if rows else 0
+    if width != grid.width:
+        problems.append(f"spec rows are {width} char(s) wide, geometry is {grid.width}")
+    if problems:
+        _add(into, RuleId.GRID_SHAPE, level, "; ".join(problems))
+
+
+def _check_spawns(level: Level, into: list[LevelViolation]) -> None:
+    grid = level.grid
+    if not level.spawns:
+        _add(into, RuleId.SPAWN_COUNT, level, "level has no player spawns")
+        return
+    for x, y in level.spawns:
+        if not grid.is_walkable(x, y):
+            _add(
+                into,
+                RuleId.SPAWN_ON_FLOOR,
+                level,
+                f"spawn ({x},{y}) is not on a walkable floor tile",
+                x=x,
+                y=y,
+            )
+    if len(level.spawns) < 2 or not all(grid.is_walkable(x, y) for x, y in level.spawns):
+        return  # SPAWN_ON_FLOOR already owns a broken spawn
+    anchor_x, anchor_y = level.spawns[0]
+    from_anchor = grid.bfs_reachable(_cell(grid, anchor_x, anchor_y))
+    for x, y in level.spawns[1:]:
+        if _cell(grid, x, y) not in from_anchor:
+            _add(
+                into,
+                RuleId.SPAWNS_MUTUALLY_REACHABLE,
+                level,
+                f"spawn ({x},{y}) cannot be reached from spawn ({anchor_x},{anchor_y})",
+                x=x,
+                y=y,
+            )
+
+
+def _check_stations(
+    level: Level, reachable: frozenset[int], into: list[LevelViolation]
+) -> tuple[set[str], set[str]]:
+    """Every station cell must sit in the spawn-reachable set.
+
+    Returns what the recipes may rely on: the values of the reachable station
+    *kinds*, and the values of the items reachable spawn shelves produce.
+    """
+    grid = level.grid
+    reachable_kinds: set[str] = set()
+    shelf_items: set[str] = set()
+    for station in level.stations:
+        x, y = int(station.x), int(station.y)
+        if _cell(grid, x, y) not in reachable:
+            rule = (
+                RuleId.SERVE_REACHABLE
+                if station.kind is StationKind.SERVE
+                else RuleId.STATION_REACHABLE
+            )
+            _add(
+                into,
+                rule,
+                level,
+                f"{station.kind.value} station {station.id} at ({x},{y}) is not reachable "
+                f"from any spawn",
+                x=x,
+                y=y,
+                station_id=station.id,
+            )
+            continue
+        reachable_kinds.add(station.kind.value)
+        item = level.station_items.get(station.id)
+        if item is not None:
+            shelf_items.add(item.value)
+    return reachable_kinds, shelf_items
+
+
+def _check_order_pool(
+    level: Level,
+    book: RecipeBook,
+    reachable_kinds: set[str],
+    shelf_items: set[str],
+    into: list[LevelViolation],
+) -> None:
+    """Pool recipes exist, are cookable on reachable stations, from reachable goods."""
+    for recipe_id in sorted({r for r in level.orders.pool if r not in book.recipes}):
+        _add(
+            into,
+            RuleId.ORDER_POOL_UNKNOWN_RECIPE,
+            level,
+            f"orders.pool names recipe {recipe_id!r}, which the recipe book does not define",
+            recipe=recipe_id,
+        )
+    for recipe_id in dict.fromkeys(level.orders.pool):
+        if recipe_id not in book.recipes:
+            continue
+        for kind_name in sorted(
+            kind.value
+            for kind in book.required_station_kinds(recipe_id)
+            if kind.value not in reachable_kinds
+        ):
+            _add(
+                into,
+                RuleId.RECIPE_STATION_MISSING,
+                level,
+                f"recipe {recipe_id!r} needs a reachable {kind_name} station; the level has none",
+                recipe=recipe_id,
+                station=kind_name,
+            )
+        recipe = book.get(recipe_id)
+        raw_kinds = {
+            ingredient.kind.value
+            for ingredient in recipe.ingredients
+            if ingredient.form is ItemForm.RAW
+        }
+        for item_name in sorted(raw_kinds - shelf_items):
+            _add(
+                into,
+                RuleId.RECIPE_INGREDIENT_UNAVAILABLE,
+                level,
+                f"recipe {recipe_id!r} needs raw {item_name} but no reachable spawn shelf "
+                f"produces it",
+                recipe=recipe_id,
+                item=item_name,
+            )
+
+
+def _check_target(
+    level: Level,
+    book: RecipeBook,
+    errors: list[LevelViolation],
+    warnings: list[LevelViolation],
+) -> None:
+    cap = theoretical_max_score(level, book)
+    if cap is None:
+        return
+    if level.target_score > TARGET_REJECT_FRACTION * cap:
+        _add(
+            errors,
+            RuleId.TARGET_IMPLAUSIBLE,
+            level,
+            f"target_score {level.target_score} is above "
+            f"{TARGET_REJECT_FRACTION:.0%} of the theoretical maximum {cap:.0f}; "
+            f"no one could ever reach it",
+            target=level.target_score,
+            theoretical_max=cap,
+        )
+    elif level.target_score > TARGET_WARN_FRACTION * cap:
+        _add(
+            warnings,
+            RuleId.TARGET_IMPLAUSIBLE,
+            level,
+            f"target_score {level.target_score} is above "
+            f"{TARGET_WARN_FRACTION:.0%} of the theoretical maximum {cap:.0f}; "
+            f"winnable but tight",
+            target=level.target_score,
+            theoretical_max=cap,
+        )
+
+
+def _check_time_limit(level: Level, into: list[LevelViolation]) -> None:
+    if level.time_limit < MIN_TIME_LIMIT_S:
+        _add(
+            into,
+            RuleId.TIME_LIMIT_INVALID,
+            level,
+            f"time_limit {level.time_limit:g}s is below the {MIN_TIME_LIMIT_S:g}s minimum",
+            time_limit=level.time_limit,
+        )
+
+
+def _check_hazard_tile(
+    level: Level,
+    x: int,
+    y: int,
+    reachable: frozenset[int],
+    into: list[LevelViolation],
+    *,
+    label: str,
+) -> None:
+    grid = level.grid
+    if not grid.in_bounds(x, y):
+        _add(
+            into,
+            RuleId.HAZARD_ON_UNREACHABLE_TILE,
+            level,
+            f"hazard {label} at ({x},{y}) is outside the {grid.width}x{grid.height} grid",
+            x=x,
+            y=y,
+        )
+    elif _cell(grid, x, y) not in reachable:
+        _add(
+            into,
+            RuleId.HAZARD_ON_UNREACHABLE_TILE,
+            level,
+            f"hazard {label} at ({x},{y}) is not reachable from any spawn "
+            f"(not a walkable tile or cut off from the cooks)",
+            x=x,
+            y=y,
+        )
+
+
+def _check_hazards(level: Level, reachable: frozenset[int], into: list[LevelViolation]) -> None:
+    station_ids = {station.id for station in level.stations}
+    for index, hazard in enumerate(level.hazards):
+        if hazard.station is not None and hazard.station not in station_ids:
+            _add(
+                into,
+                RuleId.HAZARD_REF_UNKNOWN,
+                level,
+                f"hazard {index} ({hazard.type.value}) references unknown "
+                f"station {hazard.station!r}",
+                hazard=index,
+                station=hazard.station,
+            )
+        if hazard.x is not None and hazard.y is not None:
+            _check_hazard_tile(level, hazard.x, hazard.y, reachable, into, label=str(index))
+        for point_index, (x, y) in enumerate(hazard.path or ()):
+            _check_hazard_tile(
+                level,
+                x,
+                y,
+                reachable,
+                into,
+                label=f"{index} path point {point_index}",
+            )
+
+
+def _check_station_limits(level: Level, into: list[LevelViolation]) -> None:
+    limits = level.station_limits
+    for name, value in (("plates", limits.plates), ("pots", limits.pots)):
+        if value < 1:
+            _add(
+                into,
+                RuleId.STATION_LIMIT_ZERO,
+                level,
+                f"station_limits.{name} must be at least 1, got {value}",
+                limit=name,
+                value=value,
+            )
+
+
+def _check_travel(level: Level, into: list[LevelViolation]) -> None:
+    if level.difficulty < TRAVEL_MIN_DIFFICULTY:
+        return
+    tiles = _mean_travel_tiles(level)
+    if tiles < MIN_TRAVEL_TILES:
+        _add(
+            into,
+            RuleId.TRAVEL_TOO_SHORT,
+            level,
+            f"mean spawn-to-serve travel is {tiles:.1f} tiles, below the "
+            f"{MIN_TRAVEL_TILES:g}-tile minimum for difficulty {level.difficulty}",
+            travel_tiles=tiles,
+        )
+
+
+def _check_escalation(level: Level, prev: Level, into: list[LevelViolation]) -> None:
+    axes = escalation_axes(level)
+    before = escalation_axes(prev)
+    moved = sorted(name for name, value in axes.items() if value > before[name] + AXIS_EPS)
+    if len(moved) < MIN_ESCALATED_AXES:
+        moved_text = ", ".join(moved) if moved else "none"
+        _add(
+            into,
+            RuleId.DIFFICULTY_ORDER,
+            level,
+            f"only {len(moved)} escalation axis moves up vs {prev.id} "
+            f"({moved_text}); at least {MIN_ESCALATED_AXES} are required",
+            moved=tuple(moved),
+        )
+
+
+# --- public entry points -----------------------------------------------------------
+
+
+def validate_level(
+    level: Level, book: RecipeBook, prev: Level | None = None
+) -> ValidationReport:
+    """Run every per-level rule against one built level.
+
+    `prev` (the campaign predecessor) additionally enables the
+    `DIFFICULTY_ORDER` escalation check; without it a level stands alone.
+    """
+    errors: list[LevelViolation] = []
+    warnings: list[LevelViolation] = []
+    _check_grid_shape(level, errors)
+    reachable = reachable_cells(level)
+    _check_spawns(level, errors)
+    reachable_kinds, shelf_items = _check_stations(level, reachable, errors)
+    _check_order_pool(level, book, reachable_kinds, shelf_items, errors)
+    _check_time_limit(level, errors)
+    _check_target(level, book, errors, warnings)
+    _check_hazards(level, reachable, errors)
+    _check_station_limits(level, errors)
+    _check_travel(level, errors)
+    if prev is not None:
+        _check_escalation(level, prev, errors)
+    return ValidationReport(violations=tuple(errors), warnings=tuple(warnings))
+
+
+def validate_levels(levels: Sequence[Level], book: RecipeBook) -> ValidationReport:
+    """Every level plus the bundle-wide rules, in the campaign order given.
+
+    Pass levels `(order, id)`-ordered (what `ContentBundle.levels` already
+    is): consecutive pairs get the `DIFFICULTY_ORDER` escalation check, and
+    the campaign's declared `difficulty` must never drop. Duplicate ids are
+    reported here even though `load_content` refuses them earlier, so a
+    hand-assembled level list cannot smuggle one through.
+    """
+    violations: list[LevelViolation] = []
+    warnings: list[LevelViolation] = []
+    counts: dict[str, int] = {}
+    for level in levels:
+        counts[level.id] = counts.get(level.id, 0) + 1
+    for level_id, seen in sorted(counts.items()):
+        if seen > 1:
+            violations.append(
+                LevelViolation(
+                    rule=RuleId.DUPLICATE_ID,
+                    level_id=level_id,
+                    message=(
+                        f"level id {level_id!r} appears {seen} times in the bundle; "
+                        f"ids must be unique"
+                    ),
+                    detail={"count": seen},
+                )
+            )
+    previous: Level | None = None
+    for level in levels:
+        report = validate_level(level, book, previous)
+        violations.extend(report.violations)
+        warnings.extend(report.warnings)
+        if previous is not None and level.difficulty < previous.difficulty:
+            violations.append(
+                LevelViolation(
+                    rule=RuleId.DIFFICULTY_ORDER,
+                    level_id=level.id,
+                    message=(
+                        f"difficulty {level.difficulty} drops below {previous.id}'s "
+                        f"{previous.difficulty}; the campaign must never get easier"
+                    ),
+                    detail={"previous_id": previous.id},
+                )
+            )
+        previous = level
+    return ValidationReport(violations=tuple(violations), warnings=tuple(warnings))
+
+
+def validate_bundle(bundle: ContentBundle) -> ValidationReport:
+    """Per-level validation plus the bundle rules, in campaign order."""
+    return validate_levels(bundle.levels, bundle.recipe_book)

src/undercooked/tools/levels.py

@@ -0,0 +1,696 @@
+"""``undercooked-levels`` — the author-facing content CLI (PLAN §8, A8).
+
+Everything here is a thin shell over the *same* code the server runs at startup:
+`sim.content.load_content` (YAML -> campaign-ordered levels), `sim.validation`
+(PLAN §8's named rules) and `server.routes_content.resolve_recipes_path` (the
+recipe book's location). There is no second implementation of the content
+contract, so a level that validates here validates in `create_app` — and the
+tool never enumerates levels or recipes either: it prints whatever the loaded
+bundle contains.
+
+Subcommands (``--help`` for the full text):
+
+``validate``
+    one line per level, then `ValidationReport.render()` (or `--json`), exit 1
+    on any violation. This is the gate `.agents/validate.sh` runs.
+``list``
+    one line per level: order, id, difficulty, target, time, order pool.
+``show ID``
+    the resolved level summary; ``--grid`` adds the ASCII map plus a
+    reachability overlay (``.`` reachable, ``?`` unreachable, ``#`` blocking).
+``new ID``
+    writes a scaffolded level (built from the recipe book, never from a
+    hard-coded template kitchen) and validates that one file.
+``axes``
+    the escalation-axis table for the whole campaign (feeds phase 11).
+
+Exit codes: ``0`` success, ``1`` content violates at least one rule, ``2`` the
+tool could not do its job (content would not load, an unknown level id, a file
+`new` refuses to overwrite). argparse's own usage errors are ``2`` too.
+
+Path resolution is `Settings`-driven, never cwd-driven: `--dir` beats
+`UNDERCOOKED_LEVELS_DIR`, which beats the built-in ``levels`` — and a *relative*
+path that does not exist under the current directory is anchored to this
+checkout, so the command works from `/tmp` exactly as it does from the repo root
+(the same trick `server.app.STATIC_DIR` plays for the browser assets).
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import re
+import sys
+import textwrap
+from collections import Counter
+from collections.abc import Iterator, Mapping, Sequence
+from dataclasses import replace
+from pathlib import Path
+from typing import Final
+
+from undercooked.config import Settings, load_settings
+from undercooked.server.routes_content import resolve_recipes_path
+from undercooked.sim.content import (
+    ContentBundle,
+    ContentError,
+    load_content,
+    load_level_file,
+)
+from undercooked.sim.level import Level
+from undercooked.sim.model import ItemForm, ItemKind, StationKind
+from undercooked.sim.recipe import RecipeBook
+from undercooked.sim.validation import (
+    AXIS_NAMES,
+    LevelViolation,
+    ValidationReport,
+    escalation_axes,
+    reachable_cells,
+    validate_bundle,
+    validate_level,
+)
+
+__all__ = [
+    "EXIT_OK",
+    "EXIT_UNUSABLE",
+    "EXIT_VIOLATIONS",
+    "build_parser",
+    "main",
+    "render_starter_yaml",
+    "report_json",
+    "resolve_content_paths",
+]
+
+EXIT_OK: Final = 0
+#: Content exists but breaks a PLAN §8 rule (`validate`, `new`).
+EXIT_VIOLATIONS: Final = 1
+#: The tool could not act: load failure, unknown id, refused overwrite.
+EXIT_UNUSABLE: Final = 2
+
+#: This checkout's root (`src/undercooked/tools/levels.py` -> repo root), the
+#: anchor for a relative content path when the current directory has none.
+CHECKOUT_ROOT: Final = Path(__file__).resolve().parents[3]
+
+#: A new level id is a filename-safe slug: `NN_name` by campaign convention.
+LEVEL_ID_PATTERN: Final = re.compile(r"^[a-z0-9][a-z0-9_]*$")
+
+#: Column width of the `axes` table.
+AXIS_WIDTH: Final = 20
+
+#: Tunables of a scaffolded level: deliberately gentle, so the first thing an
+#: author sees is their own content error and not a plausibility complaint.
+STARTER_TIME_LIMIT_S: Final = 150
+STARTER_TARGET_SCORE: Final = 60
+STARTER_CONCURRENT: Final = 1
+STARTER_SPAWN_INTERVAL: Final = (12.0, 16.0)
+STARTER_DEADLINE_S: Final = 60
+STARTER_HEIGHT: Final = 7
+
+#: Grid letters the scaffold prefers per item kind (shelves stand on row 1).
+_ITEM_CHARS: Final[Mapping[str, str]] = {
+    ItemKind.TOMATO.value: "T",
+    ItemKind.ONION.value: "O",
+    ItemKind.LETTUCE.value: "L",
+    ItemKind.RICE.value: "R",
+    ItemKind.FISH.value: "F",
+    ItemKind.DOUGH.value: "H",
+    ItemKind.PATTY.value: "Y",
+    ItemKind.PLATE.value: "E",
+}
+
+#: Grid letters the scaffold prefers per station kind (fixtures stand on row 4).
+_STATION_CHARS: Final[Mapping[str, str]] = {
+    StationKind.CHOP_BOARD.value: "C",
+    StationKind.STOVE.value: "S",
+    StationKind.PREP.value: "P",
+    StationKind.SERVE.value: "X",
+    StationKind.SINK.value: "K",
+    StationKind.TRASH.value: "Z",
+}
+
+#: Fallback letters, used when a preferred one is already taken.
+_SPARE_CHARS: Final = "ABDGIJMNQUVW"
+
+
+class _CliError(Exception):
+    """A failure the CLI reports as one stderr line plus a fixed exit code."""
+
+    def __init__(self, message: str, code: int = EXIT_UNUSABLE) -> None:
+        super().__init__(message)
+        self.code = code
+
+
+# --- paths ---------------------------------------------------------------------
+
+
+def _anchor(path: Path) -> Path:
+    """Make a content path absolute without trusting the current directory.
+
+    Absolute wins; else the current directory if the path is real there; else
+    this checkout (so the documented `undercooked-levels validate` works from
+    anywhere); else the path unchanged, so the error names what was configured.
+    """
+    expanded = path.expanduser()
+    if expanded.is_absolute():
+        return expanded
+    from_cwd = Path.cwd() / expanded
+    if from_cwd.exists():
+        return from_cwd
+    anchored = CHECKOUT_ROOT / expanded
+    return anchored if anchored.exists() else expanded
+
+
+def resolve_content_paths(
+    settings: Settings,
+    levels_override: Path | None = None,
+    recipes_override: Path | None = None,
+) -> tuple[Path, Path]:
+    """Resolve ``(levels_dir, recipes_path)``: flag > environment > derived.
+
+    The derived half is `server.routes_content.resolve_recipes_path` — the same
+    resolution `create_app` uses — so the CLI and the server can never look for
+    `recipes.yaml` in two different places.
+    """
+    levels_dir = _anchor(levels_override if levels_override is not None else settings.levels_dir)
+    recipes_path = resolve_recipes_path(replace(settings, levels_dir=levels_dir))
+    if recipes_override is not None:
+        recipes_path = _anchor(recipes_override)
+    return levels_dir, recipes_path
+
+
+def _load_bundle(levels_dir: Path, recipes_path: Path) -> ContentBundle:
+    """`load_content` with the load failure turned into a reported exit 2."""
+    try:
+        return load_content(levels_dir, recipes_path)
+    except ContentError as exc:
+        raise _CliError(str(exc)) from exc
+
+
+def _bundle_for(args: argparse.Namespace) -> ContentBundle:
+    levels_dir, recipes_path = resolve_content_paths(load_settings(), args.dir, args.recipes)
+    return _load_bundle(levels_dir, recipes_path)
+
+
+# --- shared rendering -----------------------------------------------------------
+
+
+def _sorted_items(items: Sequence[LevelViolation]) -> tuple[LevelViolation, ...]:
+    """Deterministic order for JSON output (the text report sorts in `render`)."""
+    return tuple(sorted(items, key=lambda item: (item.level_id, item.rule, item.message)))
+
+
+def _field(label: str, values: Sequence[str]) -> Iterator[str]:
+    """One aligned ``label  value`` line per value (labels are 17 columns wide)."""
+    for index, value in enumerate(values):
+        yield f"{label:<18}{value}" if index == 0 else f"{'':<18}{value}"
+
+
+def _wrapped(text: str, *, placeholder: str) -> list[str]:
+    """Author prose, wrapped; an empty field says so instead of printing nothing."""
+    return textwrap.wrap(text, width=76) or [placeholder]
+
+
+def _station_summary(level: Level) -> str:
+    counts = Counter(station.kind.value for station in level.stations)
+    return ", ".join(f"{kind} x{counts[kind]}" for kind in sorted(counts)) or "none"
+
+
+def _hazard_summary(level: Level) -> str:
+    counts = Counter(hazard.type.value for hazard in level.hazards)
+    return ", ".join(f"{kind} x{counts[kind]}" for kind in sorted(counts)) or "none"
+
+
+def _recipe_line(book: RecipeBook, recipe_id: str) -> str:
+    """One order-pool entry resolved against the book (or its absence)."""
+    recipe = book.recipes.get(recipe_id)
+    if recipe is None:
+        return f"{recipe_id}: MISSING from the recipe book"
+    kinds = ", ".join(sorted(kind.value for kind in book.required_station_kinds(recipe_id)))
+    return f"{recipe_id}: {recipe.name}, {recipe.base_score} pts, needs {kinds}"
+
+
+def describe_level(level: Level, book: RecipeBook) -> list[str]:
+    """The `show` summary: the resolved level, one labelled line per fact."""
+    orders = level.orders
+    low, high = orders.spawn_interval
+    reachable = len(reachable_cells(level))
+    walkable = len(level.grid.walkable_cells())
+    limits = level.station_limits
+    facts: list[tuple[str, list[str]]] = [
+        ("order", [str(level.order)]),
+        ("difficulty", [str(level.difficulty)]),
+        ("time_limit", [f"{level.time_limit:g} s"]),
+        ("target_score", [str(level.target_score)]),
+        ("grid", [f"{level.grid.width}x{level.grid.height}"]),
+        ("spawns", [", ".join(f"({x}, {y})" for x, y in level.spawns) or "none"]),
+        ("stations", [_station_summary(level)]),
+        (
+            "orders",
+            [
+                f"pool={','.join(orders.pool)} concurrent={orders.concurrent} "
+                f"interval={low:g}-{high:g}s deadline={orders.deadline:g}s "
+                f"penalty={orders.penalty}"
+            ],
+        ),
+        ("limits", [f"plates={limits.plates} pots={limits.pots}"]),
+        ("hazards", [_hazard_summary(level)]),
+        (
+            "reachability",
+            [f"{reachable}/{walkable} walkable tiles reachable from {len(level.spawns)} spawn(s)"],
+        ),
+        ("recipes", [_recipe_line(book, recipe_id) for recipe_id in dict.fromkeys(orders.pool)]),
+        ("blurb", _wrapped(level.blurb, placeholder="(none)")),
+        ("escalation_notes", _wrapped(level.escalation_notes, placeholder="(none)")),
+    ]
+    lines = [f"{level.id} - {level.name}"]
+    for label, values in facts:
+        lines.extend(_field(label, values))
+    return lines
+
+
+def reachability_overlay(level: Level) -> list[str]:
+    """The grid re-drawn as terrain: ``.``, ``?`` (walkable, unreachable), ``#``."""
+    grid = level.grid
+    reachable = reachable_cells(level)
+    return [
+        "".join(
+            "#"
+            if not grid.is_walkable(x, y)
+            else "."
+            if y * grid.width + x in reachable
+            else "?"
+            for x in range(grid.width)
+        )
+        for y in range(grid.height)
+    ]
+
+
+def _grid_lines(level: Level) -> list[str]:
+    """The authored ASCII map plus the overlay under it."""
+    overlay = reachability_overlay(level)
+    reachable = sum(row.count(".") for row in overlay)
+    walkable = sum(row.count(c) for row in overlay for c in ".?")
+    return [
+        "",
+        "map (as authored)",
+        *(f"  {row}" for row in level.spec.grid),
+        "",
+        "reachability  . reachable   ? walkable but cut off   # blocking",
+        *(f"  {row}" for row in overlay),
+        "",
+        f"  {reachable}/{walkable} walkable tiles are reachable from the spawns",
+    ]
+
+
+# --- validate -------------------------------------------------------------------
+
+
+def report_json(report: ValidationReport, bundle: ContentBundle) -> dict[str, object]:
+    """The machine-readable report: `errors`/`warnings` lists plus per level.
+
+    Key names are part of the tool's contract with scripts
+    (`undercooked-levels validate --json | python -m json.tool`): `errors` and
+    `warnings` are the rule hits themselves — `"errors": []` means green — and
+    `counts` keeps the totals readable without counting list entries.
+    """
+    violations = _sorted_items(report.violations)
+    warnings = _sorted_items(report.warnings)
+    return {
+        "ok": report.ok,
+        "counts": {"errors": len(violations), "warnings": len(warnings)},
+        "levels": [
+            {
+                "id": level.id,
+                "order": level.order,
+                "difficulty": level.difficulty,
+                "errors": len(report.errors_for(level.id)),
+                "warnings": sum(1 for w in report.warnings if w.level_id == level.id),
+            }
+            for level in bundle.levels
+        ],
+        "errors": [_violation_json(violation) for violation in violations],
+        "warnings": [_violation_json(warning) for warning in warnings],
+    }
+
+
+def _violation_json(violation: LevelViolation) -> dict[str, object]:
+    """One rule hit as stable JSON (tuples in `detail` become arrays)."""
+    return {
+        "level": violation.level_id,
+        "rule": violation.rule,
+        "message": violation.message,
+        "detail": dict(violation.detail),
+    }
+
+
+def _level_status(level: Level, report: ValidationReport) -> str:
+    errors = len(report.errors_for(level.id))
+    warnings = sum(1 for warning in report.warnings if warning.level_id == level.id)
+    verdict = "ok" if not errors else f"{errors} error(s)"
+    if warnings:
+        verdict += f", {warnings} warning(s)"
+    geometry = f"{level.grid.width}x{level.grid.height}, {len(level.stations)} stations"
+    return f"{level.id:<28} {verdict:<24} {geometry}"
+
+
+def _cmd_validate(args: argparse.Namespace) -> int:
+    bundle = _bundle_for(args)
+    report = validate_bundle(bundle)
+    if args.json:
+        print(json.dumps(report_json(report, bundle), indent=2, sort_keys=True))
+    else:
+        for level in bundle.levels:
+            print(_level_status(level, report))
+        print(report.render())
+    return EXIT_OK if report.ok else EXIT_VIOLATIONS
+
+
+# --- list -----------------------------------------------------------------------
+
+
+LIST_HEADER: Final = f"{'ORDER':>5}  {'ID':<28}{'DIFF':>4}  {'TARGET':>6}  {'TIME':>6}  RECIPES"
+
+
+def _cmd_list(args: argparse.Namespace) -> int:
+    bundle = _bundle_for(args)
+    print(LIST_HEADER)
+    for level in bundle.levels:
+        print(
+            f"{level.order:>5}  {level.id:<28}{level.difficulty:>4}  "
+            f"{level.target_score:>6}  {level.time_limit:>6.1f}  {', '.join(level.orders.pool)}"
+        )
+    return EXIT_OK
+
+
+# --- show -----------------------------------------------------------------------
+
+
+def _cmd_show(args: argparse.Namespace) -> int:
+    bundle = _bundle_for(args)
+    level = bundle.by_id.get(args.id)
+    if level is None:
+        known = ", ".join(sorted(bundle.by_id)) or "(no levels loaded)"
+        raise _CliError(f"no level {args.id!r} in this campaign; known ids: {known}")
+    print("\n".join(describe_level(level, bundle.recipe_book)))
+    if args.grid:
+        print("\n".join(_grid_lines(level)))
+    return EXIT_OK
+
+
+# --- axes -----------------------------------------------------------------------
+
+
+def _axes_header() -> str:
+    return f"{'LEVEL':<28}{'ORDER':>6}" + "".join(f"{name:>{AXIS_WIDTH}}" for name in AXIS_NAMES)
+
+
+def _axes_row(level: Level) -> str:
+    axes = escalation_axes(level)
+    cells = "".join(f"{axes[name]:>{AXIS_WIDTH}.4f}" for name in AXIS_NAMES)
+    return f"{level.id:<28}{level.order:>6}{cells}"
+
+
+def _cmd_axes(args: argparse.Namespace) -> int:
+    bundle = _bundle_for(args)
+    print(_axes_header())
+    for level in bundle.levels:
+        print(_axes_row(level))
+    return EXIT_OK
+
+
+# --- new ------------------------------------------------------------------------
+
+
+def _char_for(preferred: str | None, used: set[str]) -> str:
+    """The preferred grid letter when it is free, else the first free spare."""
+    if preferred is not None and preferred not in used:
+        return preferred
+    return next(char for char in _SPARE_CHARS if char not in used)
+
+
+def _assign_chars(keys: Sequence[str], table: Mapping[str, str], used: set[str]) -> dict[str, str]:
+    """One unique grid character per key, deterministic in the key order given."""
+    assigned: dict[str, str] = {}
+    for key in keys:
+        char = _char_for(table.get(key), used)
+        used.add(char)
+        assigned[key] = char
+    return assigned
+
+
+def _display_name(level_id: str) -> str:
+    """`07_riverboat` -> `Riverboat`; the author edits it immediately anyway."""
+    return level_id.split("_", 1)[-1].replace("_", " ").title()
+
+
+def _starter_content(
+    book: RecipeBook,
+) -> tuple[str, list[ItemKind], list[StationKind]]:
+    """(recipe id, raw ingredients to shelf, station kinds to place) for the scaffold.
+
+    The recipe book decides the template kitchen — the first dish it defines and
+    exactly the stations that dish's steps name — because nothing in Python may
+    enumerate content (A8).
+    """
+    recipe_ids = book.ids()
+    if not recipe_ids:
+        raise _CliError("the recipe book defines no recipes, so there is nothing to scaffold")
+    recipe_id = recipe_ids[0]
+    recipe = book.get(recipe_id)
+    items = list(dict.fromkeys(
+        ingredient.kind
+        for ingredient in recipe.ingredients
+        if ingredient.form is ItemForm.RAW
+    ))
+    required = book.required_station_kinds(recipe_id) | {StationKind.SERVE}
+    kinds = sorted(required - {StationKind.SPAWN_SHELF}, key=lambda kind: kind.value)
+    return recipe_id, items, kinds
+
+
+def render_starter_yaml(*, level_id: str, order: int, difficulty: int, book: RecipeBook) -> str:
+    """The `new` template: a valid, minimal kitchen for the book's first dish.
+
+    Geometry is generated, not pasted: one spawn shelf per raw ingredient on
+    row 1, one station per required kind on row 4, and two spawns on the bottom
+    row far enough from the serve window to clear the `TRAVEL_TOO_SHORT` floor.
+    """
+    recipe_id, items, kinds = _starter_content(book)
+    used = {"#", "."}
+    shelf_chars = _assign_chars([item.value for item in items], _ITEM_CHARS, used)
+    station_chars = _assign_chars([kind.value for kind in kinds], _STATION_CHARS, used)
+    width = max(13, 2 + 2 * max(len(items), len(kinds), 1))
+    rows = [["."] * width for _ in range(STARTER_HEIGHT)]
+    for index in range(width):
+        rows[0][index] = rows[STARTER_HEIGHT - 1][index] = "#"
+    for index, item in enumerate(items):
+        rows[1][2 + 2 * index] = shelf_chars[item.value]
+    for index, kind in enumerate(kinds):
+        rows[4][2 + 2 * index] = station_chars[kind.value]
+    spawns = [(1, STARTER_HEIGHT - 2), (width - 2, STARTER_HEIGHT - 2)]
+
+    tiles: dict[str, dict[str, str]] = {"#": {"kind": "wall"}, ".": {"kind": "floor"}}
+    for item in items:
+        tiles[shelf_chars[item.value]] = {
+            "kind": "floor",
+            "station": StationKind.SPAWN_SHELF.value,
+            "item": item.value,
+        }
+    for kind in kinds:
+        tiles[station_chars[kind.value]] = {"kind": "floor", "station": kind.value}
+
+    low, high = STARTER_SPAWN_INTERVAL
+    return "\n".join([
+        f"# levels/{level_id}.yaml — scaffolded by `undercooked-levels new`.",
+        "# Schema, every field and every rule: docs/LEVELS.md.",
+        "# Next steps: edit the grid and the prose, then re-run",
+        "#   uv run undercooked-levels validate --dir <this directory>",
+        "#   uv run undercooked-levels axes --dir <this directory>",
+        "# The campaign rule (DIFFICULTY_ORDER) wants >= 2 escalation axes up",
+        "# versus the previous level, and a difficulty that never drops.",
+        "",
+        f"id: {level_id}",
+        f"name: {_display_name(level_id)}",
+        f"order: {order}",
+        f"difficulty: {difficulty}",
+        f"time_limit: {STARTER_TIME_LIMIT_S}",
+        f"target_score: {STARTER_TARGET_SCORE}",
+        "blurb: >-",
+        f"  A scaffolded kitchen around one dish ({recipe_id}). Rewrite this",
+        "  paragraph for the level a player will actually read.",
+        "grid:",
+        *(f'  - "{"".join(row)}"' for row in rows),
+        "tiles:",
+        *(f'  "{char}": {json.dumps(spec)}' for char, spec in tiles.items()),
+        f"spawns: [{', '.join(f'[{x}, {y}]' for x, y in spawns)}]",
+        "orders:",
+        f"  pool: [{recipe_id}]",
+        f"  concurrent: {STARTER_CONCURRENT}",
+        f"  spawn_interval: [{low:g}, {high:g}]",
+        f"  deadline: {STARTER_DEADLINE_S}",
+        "hazards: []",
+        "escalation_notes: >-",
+        "  TODO: name the axes this level moves up versus its predecessor.",
+        "",
+    ])
+
+
+def _new_level_id(candidate: str) -> str:
+    if not LEVEL_ID_PATTERN.match(candidate):
+        raise _CliError(
+            f"level id {candidate!r} is not a filename-safe slug "
+            "(lowercase letters, digits and underscores, starting with a letter or digit)"
+        )
+    return candidate
+
+
+def _cmd_new(args: argparse.Namespace) -> int:
+    settings = load_settings()
+    levels_dir, recipes_path = resolve_content_paths(settings, args.dir, args.recipes)
+    level_id = _new_level_id(args.id)
+    if args.order is not None and args.order < 1:
+        raise _CliError(f"--order must be at least 1, got {args.order}")
+    bundle = _load_bundle(levels_dir, recipes_path)
+    target = levels_dir / f"{level_id}.yaml"
+    if target.exists():
+        raise _CliError(f"{target} already exists; refusing to overwrite an author's level")
+    orders = [level.order for level in bundle.levels]
+    difficulties = [level.difficulty for level in bundle.levels]
+    order = args.order if args.order is not None else max(orders, default=0) + 1
+    difficulty = min(10, max(difficulties, default=0) + 1)
+    text = render_starter_yaml(
+        level_id=level_id,
+        order=order,
+        difficulty=difficulty,
+        book=bundle.recipe_book,
+    )
+    try:
+        target.write_text(text, encoding="utf-8")
+    except OSError as exc:
+        reason = exc.strerror or str(exc)
+        raise _CliError(f"cannot write {target}: {reason}") from exc
+    print(f"wrote {target}")
+    try:
+        level = load_level_file(target)
+    except ContentError as exc:
+        raise _CliError(str(exc)) from exc
+    # The new file on its own: the campaign's escalation rule needs neighbours,
+    # and a scaffolded level is not tuned against them yet by definition.
+    report = validate_level(level, bundle.recipe_book)
+    print(report.render())
+    return EXIT_OK if report.ok else EXIT_VIOLATIONS
+
+
+# --- entry point -----------------------------------------------------------------
+
+
+def _add_content_flags(parser: argparse.ArgumentParser) -> None:
+    """The `--dir` / `--recipes` pair every subcommand needs to find content."""
+    parser.add_argument(
+        "--dir",
+        type=Path,
+        default=None,
+        metavar="PATH",
+        help="levels directory (default: $UNDERCOOKED_LEVELS_DIR, else levels/)",
+    )
+    parser.add_argument(
+        "--recipes",
+        type=Path,
+        default=None,
+        metavar="PATH",
+        help="recipe book (default: $UNDERCOOKED_RECIPES_PATH, else <dir>/recipes.yaml)",
+    )
+
+
+def build_parser() -> argparse.ArgumentParser:
+    """The subcommand surface of `undercooked-levels`."""
+    parser = argparse.ArgumentParser(
+        prog="undercooked-levels",
+        description=(
+            "Inspect and validate Undercooked level content: the same loader and "
+            "the same PLAN §8 rules the server runs at startup."
+        ),
+        epilog=(
+            "exit codes: 0 ok, 1 a level breaks a validation rule, "
+            "2 nothing usable to work with (load failure, unknown id, refused overwrite)"
+        ),
+    )
+    subcommands = parser.add_subparsers(dest="command", required=True, metavar="COMMAND")
+
+    validate = subcommands.add_parser(
+        "validate",
+        help="validate every level and print the named-rule report",
+        description="Validate the campaign: one line per level, then the rule report.",
+    )
+    _add_content_flags(validate)
+    validate.add_argument(
+        "--json",
+        action="store_true",
+        help="emit a machine-readable report instead of text",
+    )
+    validate.set_defaults(handler=_cmd_validate)
+
+    listing = subcommands.add_parser(
+        "list",
+        help="one line per level, in campaign order",
+        description="List the campaign: order, id, difficulty, target, time, order pool.",
+    )
+    _add_content_flags(listing)
+    listing.set_defaults(handler=_cmd_list)
+
+    show = subcommands.add_parser(
+        "show",
+        help="resolved summary of one level",
+        description="Show one resolved level; --grid adds the ASCII map and overlay.",
+    )
+    show.add_argument("id", help="level id to inspect (see the `list` subcommand)")
+    show.add_argument(
+        "--grid",
+        action="store_true",
+        help="print the ASCII map plus the reachability overlay",
+    )
+    _add_content_flags(show)
+    show.set_defaults(handler=_cmd_show)
+
+    new = subcommands.add_parser(
+        "new",
+        help="scaffold a starter level into the levels directory",
+        description="Write a starter level (from the recipe book) and validate that file.",
+    )
+    new.add_argument("id", help="new level id, e.g. 03_riverboat")
+    new.add_argument(
+        "--order",
+        type=int,
+        default=None,
+        metavar="N",
+        help="campaign order (default: one past the highest in the directory)",
+    )
+    _add_content_flags(new)
+    new.set_defaults(handler=_cmd_new)
+
+    axes = subcommands.add_parser(
+        "axes",
+        help="escalation-axis table for the whole campaign",
+        description="Print every level's escalation axes (higher = harder), campaign order.",
+    )
+    _add_content_flags(axes)
+    axes.set_defaults(handler=_cmd_axes)
+    return parser
+
+
+def main(argv: Sequence[str] | None = None) -> int:
+    """Entry point of the `undercooked-levels` console script."""
+    try:
+        args = build_parser().parse_args(argv)
+    except SystemExit as exc:  # usage error -> 2, `--help` -> 0
+        return int(exc.code or EXIT_OK)
+    handler = args.handler
+    try:
+        return int(handler(args))
+    except _CliError as exc:
+        print(f"undercooked-levels: {exc}", file=sys.stderr)
+        return exc.code
+
+
+# `python -m undercooked.tools.levels` runs this; coverage measures the parent
+# process only (tests/unit/test_levels_cli.py drives `main` directly).
+if __name__ == "__main__":  # pragma: no cover
+    raise SystemExit(main())

tests/conftest.py

@@ -42,10 +42,13 @@ def _as_env_value(value: object) -> str:
 
 @pytest.fixture(scope="session", autouse=True)
 def isolated_levels_dir() -> Iterator[None]:
-    """Point `UNDERCOOKED_LEVELS_DIR` at an empty fixture dir for every test.
+    """Point `UNDERCOOKED_LEVELS_DIR` at this level-free fixture dir for every test.
 
-    No test may read the real `levels/` content implicitly; the shipped
+    No test may read the real `levels/` content implicitly. Since phase 02
-    content pipeline only arrives with phase 02.
+    task 04 `create_app()` also needs a recipe book, and the dir carries a
+    committed minimal `recipes.yaml` (zero levels — a legal campaign) so
+    fixture-isolated apps still boot; tests that need real levels build one
+    under `tmp_path` or point explicitly at `levels/`.
     """
     FIXTURE_LEVELS_DIR.mkdir(parents=True, exist_ok=True)
     previous = os.environ.get("UNDERCOOKED_LEVELS_DIR")

tests/fixtures/levels/_bad/99_broken_room.yaml

@@ -0,0 +1,46 @@
+# A deliberately broken level for `undercooked-levels validate` (phase 02 task 05).
+#
+# It *loads* (the schema is satisfied) and then breaks three PLAN §8 rules, so
+# the CLI's exit-1 path, its named-rule report and the `show --grid` "?" overlay
+# all have a stable input to be tested against:
+#
+#   * the right-hand box (x6..x9, y1..y2) is sealed on all four sides, so the
+#     serve window inside it is cut off from the spawn -> SERVE_REACHABLE, and
+#     those same cells show up as `?` in `show --grid`;
+#   * the recipe's steps name a stove this kitchen does not have
+#     -> RECIPE_STATION_MISSING;
+#   * target_score 999 is far past 60 % of the theoretical maximum
+#     -> TARGET_IMPLAUSIBLE.
+#
+# The `_bad` directory name opts out of content discovery (`_is_draft`), so this
+# campaign is only loaded when something points `--dir` straight at it — the
+# server and the rest of the test suite never see it.
+id: 99_broken_room
+name: Broken Room
+order: 1
+difficulty: 5
+time_limit: 60
+target_score: 999
+blurb: A kitchen with a sealed-off servery, the fixture for the validator itself.
+grid:
+  - "###########"
+  - "#T..C#..V.#"
+  - "#....#....#"
+  - "#.#########"
+  - "#S........#"
+  - "###########"
+tiles:
+  "#": {kind: wall}
+  ".": {kind: floor}
+  "T": {kind: floor, station: spawn_shelf, item: tomato}
+  "C": {kind: floor, station: chop_board}
+  "V": {kind: floor, station: serve}
+  "S": {kind: floor, station: spawn_shelf, item: onion}
+spawns: [[1, 4]]
+orders:
+  pool: [soup]
+  concurrent: 1
+  spawn_interval: [12, 16]
+  deadline: 60
+hazards: []
+escalation_notes: Fixture only, never shipped and never part of a real campaign.

tests/fixtures/levels/_bad/recipes.yaml

@@ -0,0 +1,23 @@
+# The recipe book for the `_bad` fixture campaign (phase 02 task 05).
+#
+# `undercooked-levels validate --dir tests/fixtures/levels/_bad` needs a book of
+# its own so the fixture stays self-contained: the tests assert on the rule hits
+# of `99_broken_room.yaml`, and those numbers depend on this book, not on the
+# shipped `levels/recipes.yaml`. One dish, `soup`, whose steps name a
+# `chop_board`, a `stove` and a `prep` station — the fixture kitchen omits the
+# stove on purpose.
+ingredients: [tomato, onion]
+
+recipes:
+  soup:
+    id: soup
+    name: Tomato Soup
+    base_score: 40
+    ingredients:
+      - {kind: tomato, form: raw}
+      - {kind: onion, form: raw}
+    steps:
+      - {kind: pick, station: spawn_shelf}
+      - {kind: chop, station: chop_board, seconds: 0.9}
+      - {kind: cook, station: stove, seconds: 5.0}
+      - {kind: plate, station: prep, seconds: 0.4}

tests/fixtures/levels/recipes.yaml

@@ -0,0 +1,18 @@
+# tests/fixtures/levels/recipes.yaml — the minimal book for the test-isolated
+# levels directory (see tests/conftest.py `isolated_levels_dir`).
+#
+# A campaign with zero levels is legal (docs/LEVELS.md), but every
+# `create_app()` still needs a recipe book to load. Tests that want actual
+# levels build their own under `tmp_path`; the shipped campaign lives in
+# `levels/` and is covered by tests/unit/test_shipped_content.py.
+ingredients: [tomato]
+recipes:
+  tomato_soup:
+    id: tomato_soup
+    name: Test Soup
+    base_score: 10
+    ingredients:
+      - {kind: tomato, form: raw}
+    steps:
+      - {kind: pick, station: spawn_shelf}
+      - {kind: serve, station: serve}

tests/integration/test_content_api.py

@@ -0,0 +1,234 @@
+"""Content API contract over in-process HTTP (task 04).
+
+`/api/levels`, `/api/levels/{id}` and `/api/config` are the only channels the
+client uses for content and settings. These tests build the real app against
+the shipped ``levels/`` campaign and assert the exact wire shape: shipped
+ids in campaign order, responses that validate against the strict response
+models (which are ``extra="forbid"``, so a leaked field is a failure, not a
+silence), 404 for unknown ids, and the four-key client-config subset with no
+secret-, operator- or origin-named key. A content bug must crash startup
+(see ``tests/unit/test_content_startup.py``), never a request.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Iterator
+from pathlib import Path
+from typing import Any
+
+import pytest
+from fastapi.testclient import TestClient
+
+from undercooked.config import Settings
+from undercooked.server.app import create_app
+from undercooked.server.routes_content import (
+    PROTOCOL_VERSION,
+    ClientConfig,
+    LevelDetail,
+    LevelSummary,
+)
+
+REPO_ROOT = Path(__file__).resolve().parents[2]
+SHIPPED_LEVELS = REPO_ROOT / "levels"
+SHIPPED_IDS = ["01_first_kitchen", "02_double_shift"]
+
+#: Keys that must never appear anywhere in a content response. Layout and
+#: internals stay server-side (PLAN §8); operator settings stay in config.py.
+FORBIDDEN_KEYS = {
+    "grid",
+    "tiles",
+    "legend",
+    "spawns",
+    "spawn",
+    "stations",
+    "station",
+    "station_limits",
+    "orders",
+    "pool",
+    "steps",
+    "secret",
+    "password",
+    "origins",
+}
+
+
+@pytest.fixture(scope="module")
+def campaign_client() -> Iterator[TestClient]:
+    with TestClient(create_app(Settings(levels_dir=SHIPPED_LEVELS))) as test_client:
+        yield test_client
+
+
+def walk_keys(node: Any) -> set[str]:
+    """Every JSON object key at any depth of a parsed response body."""
+    if isinstance(node, dict):
+        return set(node) | {k for child in node.values() for k in walk_keys(child)}
+    if isinstance(node, list):
+        return {k for child in node for k in walk_keys(child)}
+    return set()
+
+
+# --- GET /api/levels -----------------------------------------------------------
+
+
+def test_lists_the_shipped_levels_in_campaign_order(campaign_client: TestClient) -> None:
+    response = campaign_client.get("/api/levels")
+    assert response.status_code == 200
+    body = response.json()
+    assert [level["id"] for level in body] == SHIPPED_IDS
+    assert [level["order"] for level in body] == [1, 2]
+
+
+def test_level_cards_match_the_response_model_exactly(
+    campaign_client: TestClient,
+) -> None:
+    body = campaign_client.get("/api/levels").json()
+    for card in body:
+        # extra="forbid" on LevelSummary: an unexpected key fails this, and
+        # the explicit key-set check catches a *missing* field too.
+        LevelSummary.model_validate(card)
+        assert set(card) == set(LevelSummary.model_fields)
+
+
+@pytest.mark.parametrize("level_id", SHIPPED_IDS)
+def test_level_card_exposes_the_expected_fields(
+    campaign_client: TestClient, level_id: str
+) -> None:
+    card = next(
+        level for level in campaign_client.get("/api/levels").json()
+        if level["id"] == level_id
+    )
+    assert isinstance(card["name"], str) and card["name"]
+    assert isinstance(card["blurb"], str) and len(card["blurb"]) >= 20
+    assert isinstance(card["difficulty"], int) and 1 <= card["difficulty"] <= 10
+    assert isinstance(card["target_score"], int)
+    assert isinstance(card["time_limit"], (int, float))
+    # Recipes exist — the ids the client may later send back in `join.ready`.
+    assert card["recipes"]
+    known_recipe_names = {"Tomato Soup", "Garden Salad", "Fried Rice", "Chef's Burger"}
+    assert set(card["recipes"]) <= known_recipe_names
+
+
+# --- GET /api/levels/{id} -------------------------------------------------------
+
+
+@pytest.mark.parametrize("level_id", SHIPPED_IDS)
+def test_level_detail_matches_the_response_model(
+    campaign_client: TestClient, level_id: str
+) -> None:
+    response = campaign_client.get(f"/api/levels/{level_id}")
+    assert response.status_code == 200
+    detail = response.json()
+    LevelDetail.model_validate(detail)
+    assert set(detail) == set(LevelSummary.model_fields) | {"hazards"}
+    assert detail["id"] == level_id
+
+
+def test_level_detail_hazards_are_type_only(campaign_client: TestClient) -> None:
+    detail = campaign_client.get("/api/levels/02_double_shift").json()
+    # Neither shipped level carries a hazard; when one does, each entry is
+    # exactly {"type": ...} — kind only, never a position (US-07).
+    assert detail["hazards"] == []
+
+
+def test_unknown_level_is_404(campaign_client: TestClient) -> None:
+    response = campaign_client.get("/api/levels/99_ghost_kitchen")
+    assert response.status_code == 404
+
+
+def test_path_style_level_ids_cannot_reach_the_filesystem(
+    campaign_client: TestClient,
+) -> None:
+    # Raw ".." would be normalised away by the HTTP client before it leaves;
+    # the encoded forms are what a hostile browser can actually send.
+    for hostile in ["..%2F..%2Fetc%2Fpasswd", "%2e%2e/levels", "%2e%2e"]:
+        response = campaign_client.get(f"/api/levels/{hostile}")
+        assert response.status_code == 404, hostile
+
+
+# --- no internals leak ----------------------------------------------------------
+
+
+@pytest.mark.parametrize(
+    "path",
+    [
+        "/api/levels",
+        "/api/levels/01_first_kitchen",
+        "/api/levels/02_double_shift",
+        "/api/config",
+    ],
+)
+def test_no_internal_or_secret_keys_any_in_response(
+    campaign_client: TestClient, path: str
+) -> None:
+    response = campaign_client.get(path)
+    assert response.status_code == 200
+    leaked = walk_keys(response.json()) & FORBIDDEN_KEYS
+    assert not leaked, f"{path} leaked keys: {sorted(leaked)}"
+
+
+def test_no_grid_rows_leak_into_any_content_response(
+    campaign_client: TestClient,
+) -> None:
+    for path in [
+        "/api/levels",
+        "/api/levels/01_first_kitchen",
+        "/api/levels/02_double_shift",
+        "/api/config",
+    ]:
+        body = campaign_client.get(path).text
+        assert "####" not in body, f"{path} appears to carry a raw grid row"
+
+
+# --- GET /api/config --------------------------------------------------------------
+
+
+def test_config_exposes_exactly_the_client_safe_fields(
+    campaign_client: TestClient,
+) -> None:
+    response = campaign_client.get("/api/config")
+    assert response.status_code == 200
+    body = response.json()
+    ClientConfig.model_validate(body)
+    assert set(body) == {"max_players", "tick_hz", "snapshot_hz", "protocol_version"}
+    assert body == {
+        "max_players": 4,
+        "tick_hz": 20,
+        "snapshot_hz": 20,
+        "protocol_version": PROTOCOL_VERSION,
+    }
+
+
+def test_config_has_no_secret_operator_or_origin_named_key(
+    campaign_client: TestClient,
+) -> None:
+    body = campaign_client.get("/api/config").json()
+    for key in body:
+        lowered = key.lower()
+        assert not any(
+            marker in lowered
+            for marker in ("secret", "password", "origins", "token", "key")
+        ), key
+
+
+def test_config_reflects_the_settings_built_into_this_app() -> None:
+    settings = Settings(levels_dir=SHIPPED_LEVELS, max_players=2, tick_hz=20, snapshot_hz=10)
+    with TestClient(create_app(settings)) as test_client:
+        body = test_client.get("/api/config").json()
+    assert body == {
+        "max_players": 2,
+        "tick_hz": 20,
+        "snapshot_hz": 10,
+        "protocol_version": PROTOCOL_VERSION,
+    }
+
+
+# --- the test-isolated campaign ----------------------------------------------------
+
+
+def test_the_env_isolated_app_starts_with_zero_levels(client: TestClient) -> None:
+    # The session's isolated levels dir holds only a recipe book (A8 allows
+    # a zero-level campaign); the app must still boot and answer cleanly.
+    response = client.get("/api/levels")
+    assert response.status_code == 200
+    assert response.json() == []
+    assert client.get("/api/levels/01_first_kitchen").status_code == 404

tests/sim/test_grid.py

@@ -0,0 +1,207 @@
+"""Unit tests for `undercooked.sim.grid` (phase 02 task 01).
+
+Geometry, walkability, BFS around walls, and every `GridError` the ASCII
+parser must refuse: ragged rows, unknown legend chars, empty/oversized maps.
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from undercooked.sim.grid import MAX_GRID, Grid, GridError
+from undercooked.sim.model import TileKind, Vec2
+
+pytestmark = pytest.mark.unit
+
+LEGEND = {
+    ".": TileKind.FLOOR,
+    "#": TileKind.WALL,
+    "C": TileKind.COUNTER,
+    "o": TileKind.VOID,
+}
+
+
+def make_grid(rows: list[str]) -> Grid:
+    return Grid.from_rows(rows, LEGEND)
+
+
+# --- Grid.from_rows --------------------------------------------------------------
+
+
+def test_from_rows_lays_tiles_out_row_major() -> None:
+    grid = make_grid(["..#", "#.o", "C#."])
+    assert (grid.width, grid.height) == (3, 3)
+    assert grid.tiles == (
+        TileKind.FLOOR,
+        TileKind.FLOOR,
+        TileKind.WALL,
+        TileKind.WALL,
+        TileKind.FLOOR,
+        TileKind.VOID,
+        TileKind.COUNTER,
+        TileKind.WALL,
+        TileKind.FLOOR,
+    )
+    assert dict(grid.station_at) == {}
+
+
+def test_from_rows_rejects_no_rows() -> None:
+    with pytest.raises(GridError, match="no rows"):
+        make_grid([])
+
+
+def test_from_rows_rejects_ragged_rows() -> None:
+    with pytest.raises(GridError, match="ragged grid: row 1"):
+        make_grid(["...", ".."])
+
+
+def test_from_rows_rejects_unknown_legend_char() -> None:
+    with pytest.raises(GridError, match=r"unknown legend char '@' at \(2,0\)"):
+        make_grid(["..@"])
+
+
+def test_from_rows_rejects_empty_row() -> None:
+    with pytest.raises(GridError, match=r"1\.\.64"):
+        make_grid([""])
+
+
+def test_from_rows_rejects_oversized_map() -> None:
+    with pytest.raises(GridError, match=r"1\.\.64"):
+        make_grid(["." * (MAX_GRID + 1)])
+
+
+@pytest.mark.parametrize(
+    ("width", "height", "tiles"),
+    [
+        (0, 1, ()),
+        (1, 0, ()),
+        (MAX_GRID + 1, 1, (TileKind.FLOOR,) * (MAX_GRID + 1)),
+        (2, 2, (TileKind.FLOOR,) * 3),  # length mismatch
+    ],
+)
+def test_direct_construction_validates_dimensions(
+    width: int, height: int, tiles: tuple[TileKind, ...]
+) -> None:
+    with pytest.raises(GridError):
+        Grid(width=width, height=height, tiles=tiles)
+
+
+# --- geometry ---------------------------------------------------------------------
+
+
+@pytest.fixture
+def grid() -> Grid:
+    return make_grid(["..#", "#.o", "C#."])
+
+
+@pytest.mark.parametrize(
+    ("x", "y", "expected"),
+    [(0, 0, True), (2, 2, True), (-1, 0, False), (0, -1, False), (3, 0, False), (0, 3, False)],
+)
+def test_in_bounds(grid: Grid, x: int, y: int, expected: bool) -> None:
+    assert grid.in_bounds(x, y) is expected
+
+
+def test_tile_at_in_and_out_of_bounds(grid: Grid) -> None:
+    assert grid.tile_at(0, 0) is TileKind.FLOOR
+    assert grid.tile_at(2, 0) is TileKind.WALL
+    assert grid.tile_at(1, 1) is TileKind.FLOOR
+    assert grid.tile_at(-1, 0) is None
+    assert grid.tile_at(3, 2) is None
+
+
+def test_is_walkable(grid: Grid) -> None:
+    assert grid.is_walkable(0, 0) is True
+    assert grid.is_walkable(1, 2) is False  # wall
+    assert grid.is_walkable(2, 1) is False  # void
+    assert grid.is_walkable(0, 2) is False  # counter
+    assert grid.is_walkable(9, 9) is False  # off-map
+
+
+def test_walkable_cells_is_row_major(grid: Grid) -> None:
+    assert grid.walkable_cells() == (0, 1, 4, 8)
+
+
+def test_blocking_mask_covers_every_cell(grid: Grid) -> None:
+    assert grid.blocking_mask() == (
+        False,
+        False,
+        True,
+        True,
+        False,
+        True,
+        True,
+        True,
+        False,
+    )
+
+
+# --- neighbours ---------------------------------------------------------------------
+
+
+def test_neighbours_four_way_order_is_north_west_east_south() -> None:
+    grid = make_grid(["...", "...", "..."])
+    assert grid.neighbours(4) == (1, 3, 5, 7)
+
+
+def test_neighbours_eight_way_is_complete_around_a_center() -> None:
+    grid = make_grid(["...", "...", "..."])
+    assert grid.neighbours(4, four_way=False) == (1, 0, 2, 3, 5, 6, 8, 7)
+
+
+def test_neighbours_are_clipped_to_the_map() -> None:
+    grid = make_grid(["..", ".."])
+    assert grid.neighbours(0) == (1, 2)
+    assert set(grid.neighbours(0, four_way=False)) == {1, 2, 3}
+
+
+def test_neighbours_reject_out_of_range_cells() -> None:
+    grid = make_grid([".."])
+    with pytest.raises(GridError, match="out of range"):
+        grid.neighbours(2)
+
+
+# --- bfs_reachable ---------------------------------------------------------------------
+
+
+def test_bfs_finds_the_whole_open_region() -> None:
+    grid = make_grid(["...", "..."])
+    assert grid.bfs_reachable(0) == frozenset(range(6))
+
+
+def test_bfs_walks_around_a_wall_column() -> None:
+    grid = make_grid([".##.", "....", ".##."])
+    # every floor cell is connected through the middle row
+    assert grid.bfs_reachable(0) == frozenset({0, 3, 4, 5, 6, 7, 8, 11})
+
+
+def test_bfs_cannot_cross_a_full_wall() -> None:
+    grid = make_grid(["...###", "...###"])
+    assert grid.bfs_reachable(0) == frozenset({0, 1, 2, 6, 7, 8})
+
+
+def test_bfs_honours_extra_blocked_cells() -> None:
+    grid = make_grid(["..."])
+    assert grid.bfs_reachable(0) == frozenset({0, 1, 2})
+    assert grid.bfs_reachable(0, extra_blocked=frozenset({1})) == frozenset({0})
+    assert grid.bfs_reachable(0, extra_blocked=frozenset({0})) == frozenset()
+
+
+def test_bfs_from_a_non_walkable_start_reaches_nothing() -> None:
+    grid = make_grid([".#"])
+    assert grid.bfs_reachable(1) == frozenset()
+
+
+def test_bfs_rejects_an_out_of_range_start() -> None:
+    grid = make_grid([".."])
+    with pytest.raises(GridError, match="out of range"):
+        grid.bfs_reachable(2)
+
+
+# --- manhattan --------------------------------------------------------------------------
+
+
+def test_manhattan_distance() -> None:
+    assert Grid.manhattan(Vec2(1.0, 1.0), Vec2(4.0, 5.0)) == 7.0
+    assert Grid.manhattan(Vec2(-2.5, 0.25), Vec2(0.5, 0.25)) == 3.0
+    assert Grid.manhattan(Vec2(2.0, 2.0), Vec2(2.0, 2.0)) == 0.0

tests/sim/test_model.py

@@ -0,0 +1,431 @@
+"""Unit tests for the pure simulation data model (phase 02 task 01, PLAN §5).
+
+Covers the enum vocabulary, `Vec2` geometry, the frozen entities and their
+value-semantics helpers, and the `state_hash` contract the replay tests
+(A21) will rely on: scored changes change the hash, cosmetic ones never do.
+"""
+
+from __future__ import annotations
+
+from dataclasses import replace
+from typing import Any
+
+import pytest
+
+from undercooked.sim.grid import Grid
+from undercooked.sim.model import (
+    OPPOSITE,
+    WALKABLE,
+    Event,
+    EventKind,
+    Facing,
+    Hazard,
+    HazardKind,
+    Item,
+    ItemForm,
+    ItemKind,
+    KitchenState,
+    Order,
+    Player,
+    PlayerState,
+    RoundResult,
+    Station,
+    StationKind,
+    StationState,
+    TileKind,
+    Vec2,
+    _canonical,
+)
+
+pytestmark = pytest.mark.unit
+
+
+@pytest.fixture
+def state() -> KitchenState:
+    grid = Grid(width=3, height=2, tiles=(TileKind.FLOOR,) * 6)
+    return KitchenState(level_id="01_test", seed=42, grid=grid, target_score=40, time_left=90.0)
+
+
+# --- Facing -------------------------------------------------------------------
+
+
+@pytest.mark.parametrize(
+    ("facing", "delta"),
+    [
+        (Facing.UP, Vec2(0.0, -1.0)),
+        (Facing.RIGHT, Vec2(1.0, 0.0)),
+        (Facing.DOWN, Vec2(0.0, 1.0)),
+        (Facing.LEFT, Vec2(-1.0, 0.0)),
+    ],
+)
+def test_facing_delta_is_a_unit_step(facing: Facing, delta: Vec2) -> None:
+    assert facing.delta() == delta
+
+
+def test_facing_wire_values() -> None:
+    assert [f.value for f in Facing] == ["up", "right", "down", "left"]
+
+
+def test_opposite_pairs_and_is_involutive() -> None:
+    assert len(OPPOSITE) == 4
+    for facing in Facing:
+        assert OPPOSITE[OPPOSITE[facing]] is facing
+
+
+# --- Vec2 ---------------------------------------------------------------------
+
+
+def test_vec2_defaults_to_zero() -> None:
+    assert Vec2() == Vec2(0.0, 0.0)
+
+
+def test_vec2_add_and_sub() -> None:
+    a, b = Vec2(2.0, 3.0), Vec2(0.5, -1.0)
+    assert a + b == Vec2(2.5, 2.0)
+    assert a - b == Vec2(1.5, 4.0)
+
+
+def test_vec2_scale() -> None:
+    assert Vec2(1.5, -2.0).scale(2.0) == Vec2(3.0, -4.0)
+
+
+@pytest.mark.parametrize(
+    ("vec", "expected"),
+    [(Vec2(3.0, 4.0), 5.0), (Vec2(0.0, 0.0), 0.0), (Vec2(-1.0, 0.0), 1.0)],
+)
+def test_vec2_length(vec: Vec2, expected: float) -> None:
+    assert vec.length == pytest.approx(expected)
+
+
+def test_vec2_normalized() -> None:
+    unit = Vec2(3.0, 4.0).normalized()
+    assert unit.x == pytest.approx(0.6)
+    assert unit.y == pytest.approx(0.8)
+    assert unit.length == pytest.approx(1.0)
+
+
+def test_vec2_zero_normalizes_to_zero_without_division_error() -> None:
+    assert Vec2(0.0, 0.0).normalized() == Vec2(0.0, 0.0)
+
+
+def test_vec2_quantized() -> None:
+    assert Vec2(1.23456, -2.98765).quantized() == Vec2(1.23, -2.99)
+    assert Vec2(1.23456, -2.98765).quantized(digits=1) == Vec2(1.2, -3.0)
+
+
+@pytest.mark.parametrize(
+    ("vec", "key"),
+    [
+        (Vec2(0.0, 0.0), (0, 0)),
+        (Vec2(1.5, 2.25), (96, 144)),
+        (Vec2(0.02, 0.01), (1, 1)),
+        (Vec2(-0.03, -0.007), (-2, 0)),
+    ],
+)
+def test_vec2_key_is_64th_tile_resolution(vec: Vec2, key: tuple[int, int]) -> None:
+    assert key == vec.KEY
+
+
+# --- TileKind / WALKABLE -------------------------------------------------------
+
+
+def test_only_floor_is_walkable() -> None:
+    assert frozenset({TileKind.FLOOR}) == WALKABLE
+    assert {t.value for t in TileKind} == {"floor", "counter", "wall", "void"}
+
+
+# --- Item ----------------------------------------------------------------------
+
+
+def test_item_defaults() -> None:
+    item = Item(kind=ItemKind.TOMATO, form=ItemForm.RAW)
+    assert (item.quality, item.ttl, item.dish_id) == (1.0, 0.0, None)
+
+
+def test_item_is_trash_only_when_burnt() -> None:
+    for form in ItemForm:
+        item = Item(kind=ItemKind.PATTY, form=form)
+        assert item.is_trash() is (form is ItemForm.BURNT)
+
+
+def test_item_with_form_returns_a_new_instance() -> None:
+    raw = Item(kind=ItemKind.ONION, form=ItemForm.RAW, quality=0.5)
+    chopped = raw.with_form(ItemForm.CHOPPED)
+    assert chopped is not raw
+    assert chopped.form is ItemForm.CHOPPED
+    assert chopped.kind is ItemKind.ONION and chopped.quality == 0.5
+    assert raw.form is ItemForm.RAW  # value semantics: the original never changes
+
+
+# --- Station / Player / Order / Hazard ------------------------------------------
+
+
+def test_station_defaults() -> None:
+    station = Station(id="stove_1", kind=StationKind.STOVE, x=3.0, y=2.0)
+    assert (station.contents, station.state) == ((), StationState.IDLE)
+    assert (station.progress, station.duration, station.cooldown) == (0.0, 0.0, 0.0)
+    assert station.worker_id is None and station.dirty is False
+
+
+def test_station_kinds_and_states_match_the_plan_vocabulary() -> None:
+    assert {k.value for k in StationKind} == {
+        "spawn_shelf",
+        "chop_board",
+        "stove",
+        "prep",
+        "serve",
+        "sink",
+        "trash",
+    }
+    assert {s.value for s in StationState} == {
+        "idle",
+        "working",
+        "ready",
+        "busy",
+        "dirty",
+        "blocked",
+    }
+
+
+def test_player_defaults() -> None:
+    player = Player(id="tok")
+    assert player.pos == Vec2(0.0, 0.0) and player.vel == Vec2(0.0, 0.0)
+    assert player.facing is Facing.DOWN and player.carrying is None
+    assert player.state is PlayerState.IDLE
+    assert (player.stun, player.score, player.dishes_served) == (0.0, 0, 0)
+    assert player.connected is True and player.seat == 0
+
+
+def test_player_state_vocabulary() -> None:
+    assert {s.value for s in PlayerState} == {"idle", "moving", "using", "stunned", "disconnected"}
+
+
+def test_order_defaults() -> None:
+    order = Order(id=17, recipe_id="soup")
+    assert (order.time_left, order.deadline, order.value) == (0.0, 0.0, 0)
+
+
+def test_hazard_defaults() -> None:
+    hazard = Hazard(id="hz1", kind=HazardKind.FIRE)
+    assert hazard.pos == Vec2(0.0, 0.0)
+    assert hazard.life == 0.0 and hazard.path == () and hazard.period_s == 0.0
+    assert hazard.station_id is None and hazard.chance_per_s == 0.0
+
+
+def test_hazard_and_result_vocabularies() -> None:
+    assert {k.value for k in HazardKind} == {"fire", "spill", "moving_obstacle", "rush"}
+    assert {r.value for r in RoundResult} == {"won", "lost", "aborted"}
+
+
+def test_item_kind_and_form_vocabularies() -> None:
+    assert {k.value for k in ItemKind} == {
+        "tomato",
+        "onion",
+        "lettuce",
+        "dough",
+        "patty",
+        "rice",
+        "fish",
+        "plate",
+    }
+    assert {f.value for f in ItemForm} == {
+        "raw",
+        "chopped",
+        "cooking",
+        "cooked",
+        "burnt",
+        "plated",
+        "dirty",
+    }
+
+
+# --- Event / EventKind -----------------------------------------------------------
+
+
+@pytest.mark.parametrize(
+    ("kind", "value"),
+    [
+        (EventKind.CHOPPED, "chopped"),
+        (EventKind.SERVED, "served"),
+        (EventKind.BURNT, "burnt"),
+        (EventKind.ORDER_MISSED, "order_missed"),
+        (EventKind.FIRE_STARTED, "fire_started"),
+        (EventKind.DENIED, "denied"),
+        (EventKind.SCORE, "score"),
+        (EventKind.STUNNED, "stunned"),
+        (EventKind.WASHED, "washed"),
+        (EventKind.TRASHED, "trashed"),
+        (EventKind.TICK_SLOW, "tick_slow"),
+    ],
+)
+def test_event_kind_constants(kind: EventKind, value: str) -> None:
+    assert kind == value
+
+
+def test_event_defaults() -> None:
+    event = Event(kind=EventKind.SERVED)
+    assert event.kind == "served"
+    assert (event.x, event.y) == (0.0, 0.0)
+    assert dict(event.payload) == {}
+
+
+def test_event_carries_a_payload() -> None:
+    event = Event(kind=EventKind.SCORE, x=9.0, y=1.0, payload={"points": 40})
+    assert event.payload["points"] == 40
+
+
+# --- KitchenState.state_hash -------------------------------------------------------
+
+
+def test_state_hash_is_stable_hex(state: KitchenState) -> None:
+    first = state.state_hash()
+    assert first == state.state_hash()
+    assert len(first) == 64
+    assert all(c in "0123456789abcdef" for c in first)
+
+
+@pytest.mark.parametrize(
+    "mutation",
+    [
+        lambda s: replace(s, score=s.score + 1),
+        lambda s: replace(s, target_score=s.target_score + 10),
+        lambda s: replace(s, tick=s.tick + 1),
+        lambda s: replace(s, time_left=s.time_left + 0.01),
+        lambda s: replace(s, streak=s.streak + 1),
+        lambda s: replace(s, result=RoundResult.WON),
+        lambda s: replace(s, seed=s.seed + 1),
+        lambda s: replace(s, level_id="02_other"),
+        lambda s: replace(s, rng_state=s.rng_state ^ 1),
+        lambda s: replace(s, counters={"served": 1}),
+        lambda s: replace(s, players=(Player(id="p1"),)),
+        lambda s: replace(s, stations=(Station(id="s1", kind=StationKind.STOVE),)),
+        lambda s: replace(s, orders=(Order(id=1, recipe_id="soup"),)),
+        lambda s: replace(s, hazards=(Hazard(id="h1", kind=HazardKind.FIRE),)),
+        lambda s: replace(
+            s, grid=Grid(width=3, height=2, tiles=(TileKind.FLOOR,) * 5 + (TileKind.WALL,))
+        ),
+    ],
+    ids=[
+        "score",
+        "target_score",
+        "tick",
+        "time_left",
+        "streak",
+        "result",
+        "seed",
+        "level_id",
+        "rng_state",
+        "counters",
+        "players",
+        "stations",
+        "orders",
+        "hazards",
+        "grid",
+    ],
+)
+def test_state_hash_changes_for_scored_fields(
+    state: KitchenState, mutation: Any
+) -> None:
+    assert mutation(state).state_hash() != state.state_hash()
+
+
+def test_state_hash_ignores_counter_insertion_order() -> None:
+    grid = Grid(width=1, height=1, tiles=(TileKind.FLOOR,))
+    left = KitchenState("01_a", 1, grid, counters={"served": 2, "missed": 1})
+    right = KitchenState("01_a", 1, grid, counters={"missed": 1, "served": 2})
+    assert left.state_hash() == right.state_hash()
+
+
+def test_state_hash_ignores_station_at_insertion_order() -> None:
+    tiles = (TileKind.FLOOR,) * 6
+    grid_a = Grid(width=3, height=2, tiles=tiles, station_at={0: "shelf", 5: "serve"})
+    grid_b = Grid(width=3, height=2, tiles=tiles, station_at={5: "serve", 0: "shelf"})
+    assert (
+        KitchenState("01_a", 1, grid_a).state_hash()
+        == KitchenState("01_a", 1, grid_b).state_hash()
+    )
+
+
+def test_state_hash_ignores_float_repr_noise(state: KitchenState) -> None:
+    # cosmetic float noise (< 1e-6) is below the canonical quantization...
+    assert replace(state, time_left=90.0 + 1e-9).state_hash() == state.state_hash()
+    assert replace(state, time_left=-0.0).state_hash() == replace(
+        state, time_left=0.0
+    ).state_hash()
+    # ...but a real difference above it always changes the hash
+    assert replace(state, time_left=90.000001).state_hash() != state.state_hash()
+
+
+# --- KitchenState.with_* helpers -----------------------------------------------------
+
+
+def test_with_tick_and_time_left_return_new_instances(state: KitchenState) -> None:
+    advanced = state.with_tick(7).with_time_left(83.5)
+    assert advanced.tick == 7 and advanced.time_left == 83.5
+    assert advanced is not state
+    assert state.tick == 0 and state.time_left == 90.0
+
+
+def test_with_score_optionally_sets_streak(state: KitchenState) -> None:
+    scored = state.with_score(40)
+    assert scored.score == 40 and scored.streak == state.streak
+    with_streak = state.with_score(40, streak=3)
+    assert with_streak.score == 40 and with_streak.streak == 3
+    assert state.score == 0
+
+
+@pytest.mark.parametrize("result", [RoundResult.WON, RoundResult.LOST, None])
+def test_with_result(state: KitchenState, result: RoundResult | None) -> None:
+    updated = state.with_result(result)
+    assert updated.result is result
+    assert state.result is None
+
+
+def test_with_entity_collections(state: KitchenState) -> None:
+    stations = (Station(id="s1", kind=StationKind.CHOP_BOARD),)
+    players = (Player(id="p1", seat=0),)
+    orders = (Order(id=1, recipe_id="soup"),)
+    hazards = (Hazard(id="h1", kind=HazardKind.SPILL),)
+    updated = (
+        state.with_stations(stations)
+        .with_players(players)
+        .with_orders(orders)
+        .with_hazards(hazards)
+    )
+    assert updated.stations == stations
+    assert updated.players == players
+    assert updated.orders == orders
+    assert updated.hazards == hazards
+    assert state.stations == () and state.players == ()
+    assert state.orders == () and state.hazards == ()
+
+
+def test_with_counters_merges_without_mutating(state: KitchenState) -> None:
+    merged = state.with_counters(served=1).with_counters(served=2, missed=1)
+    assert dict(merged.counters) == {"served": 2, "missed": 1}
+    assert dict(state.counters) == {}
+
+
+# --- _canonical ------------------------------------------------------------------------
+
+
+def test_canonical_scalars_are_typed_tokens() -> None:
+    assert _canonical(None) == "n:"
+    assert _canonical(True) == "b:1"
+    assert _canonical(1) == "i:1"  # bool is checked before int
+    assert _canonical("served") == "s:served"
+    assert _canonical(Facing.UP) == "s:up"  # StrEnums canonicalise by value
+    assert _canonical(2.5) == "f:2.500000"
+
+
+def test_canonical_sorts_mapping_keys() -> None:
+    assert _canonical({"b": 1, "a": 2}) == _canonical({"a": 2, "b": 1})
+
+
+def test_canonical_renders_dataclasses_field_name_sorted() -> None:
+    assert _canonical(Vec2(1.0, 2.0)) == "d:Vec2(x=f:1.000000,y=f:2.000000)"
+
+
+def test_canonical_rejects_unknown_types_so_fields_cannot_hide() -> None:
+    with pytest.raises(TypeError, match="cannot canonicalise"):
+        _canonical(object())

tests/sim/test_purity.py

@@ -0,0 +1,81 @@
+"""Unit tests for the A5 purity guard (phase 02 task 01).
+
+Proves the shipped `sim/` package imports nothing forbidden, and — just as
+important — that the guard actually catches a violation when one exists.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import pytest
+
+from undercooked.sim.purity import (
+    FORBIDDEN_ROOTS,
+    PurityError,
+    assert_pure_imports,
+    find_violations,
+)
+
+pytestmark = pytest.mark.unit
+
+
+def test_forbidden_roots_are_the_plan_a5_list() -> None:
+    assert frozenset({"asyncio", "fastapi", "socket", "time", "uvicorn"}) == FORBIDDEN_ROOTS
+
+
+def test_shipped_sim_package_is_pure() -> None:
+    assert find_violations() == []
+    assert_pure_imports()  # must not raise
+
+
+def test_clean_module_with_stdlib_sibling_and_relative_imports(tmp_path: Path) -> None:
+    (tmp_path / "ok.py").write_text(
+        "import math\n"
+        "from hashlib import sha256\n"
+        "from undercooked.sim.model import Vec2\n"
+        "from . import model\n",
+        encoding="utf-8",
+    )
+    assert find_violations(tmp_path) == []
+
+
+def test_comments_and_strings_are_not_imports(tmp_path: Path) -> None:
+    (tmp_path / "prose.py").write_text(
+        '# we must never "import asyncio" here\nDOC = "import time is forbidden"\n',
+        encoding="utf-8",
+    )
+    assert find_violations(tmp_path) == []
+
+
+@pytest.mark.parametrize(
+    ("statement", "root"),
+    [
+        ("import asyncio", "asyncio"),
+        ("import time", "time"),
+        ("from time import monotonic", "time"),
+        ("import socket", "socket"),
+        ("from socket import socket", "socket"),
+        ("from fastapi import FastAPI", "fastapi"),
+        ("import uvicorn", "uvicorn"),
+        ("import uvicorn.protocols", "uvicorn"),
+    ],
+)
+def test_find_violations_detects_each_forbidden_import(
+    tmp_path: Path, statement: str, root: str
+) -> None:
+    (tmp_path / "bad.py").write_text(f"{statement}\n", encoding="utf-8")
+    assert find_violations(tmp_path) == [f"bad.py: imports {root}"]
+
+
+def test_find_violations_scans_subpackages(tmp_path: Path) -> None:
+    sub = tmp_path / "deep"
+    sub.mkdir()
+    (sub / "leak.py").write_text("import asyncio\n", encoding="utf-8")
+    assert find_violations(tmp_path) == ["deep/leak.py: imports asyncio"]
+
+
+def test_assert_pure_imports_raises_with_the_violation_list(tmp_path: Path) -> None:
+    (tmp_path / "bad.py").write_text("import time\nimport socket\n", encoding="utf-8")
+    with pytest.raises(PurityError, match=r"bad\.py: imports socket"):
+        assert_pure_imports(tmp_path)

tests/sim/test_validation.py

@@ -0,0 +1,490 @@
+"""Unit tests for `undercooked.sim.validation` (phase 02 task 03).
+
+One failing case per rule id (asserted against `RuleId.ALL`) plus the passing
+counterpart, a planted-unreachable-station exactness test, the bundle rules,
+and the escalation/target math. Every fixture level is built by `make_level`
+so a rule's failing case is one readable override away.
+"""
+
+from __future__ import annotations
+
+import copy
+from collections.abc import Callable
+from dataclasses import replace
+from typing import Any
+
+import pytest
+
+from undercooked.sim.content import ContentBundle
+from undercooked.sim.grid import Grid
+from undercooked.sim.level import Level, LevelDef
+from undercooked.sim.model import TileKind
+from undercooked.sim.recipe import RecipeBook
+from undercooked.sim.validation import (
+    TARGET_WARN_FRACTION,
+    RuleId,
+    escalation_axes,
+    theoretical_max_score,
+    validate_bundle,
+    validate_level,
+    validate_levels,
+)
+
+pytestmark = pytest.mark.unit
+
+# --- fixtures ---------------------------------------------------------------------
+
+RECIPES: dict[str, dict[str, Any]] = {
+    "soup": {
+        "id": "soup",
+        "name": "Tomato Soup",
+        "base_score": 40,
+        "ingredients": [{"kind": "tomato", "form": "raw"}, {"kind": "onion", "form": "raw"}],
+        "steps": [
+            {"kind": "chop", "station": "chop_board", "seconds": 0.9},
+            {"kind": "cook", "station": "stove", "seconds": 5.0},
+            {"kind": "plate", "station": "prep", "seconds": 0.5},
+            {"kind": "serve", "station": "serve"},
+        ],
+    },
+    "salad": {
+        "id": "salad",
+        "name": "Garden Salad",
+        "base_score": 30,
+        "ingredients": [{"kind": "lettuce", "form": "chopped"}],
+        "steps": [
+            {"kind": "plate", "station": "prep", "seconds": 0.4},
+            {"kind": "serve", "station": "serve"},
+        ],
+    },
+}
+
+CLEAN_GRID = [
+    "###########",
+    "#S.O....C.#",
+    "#.........#",
+    "#..X...P..#",
+    "#.........#",
+    "#......V..#",
+    "###########",
+]
+
+CLEAN_TILES: dict[str, dict[str, str]] = {
+    ".": {"kind": "floor"},
+    "#": {"kind": "wall"},
+    "S": {"kind": "floor", "station": "spawn_shelf", "item": "tomato"},
+    "O": {"kind": "floor", "station": "spawn_shelf", "item": "onion"},
+    "C": {"kind": "floor", "station": "chop_board"},
+    "X": {"kind": "floor", "station": "stove"},
+    "P": {"kind": "floor", "station": "prep"},
+    "V": {"kind": "floor", "station": "serve"},
+}
+
+#: The clean kitchen grown by two rows: a sealed 3-cell pocket at (7..9, 7).
+POCKET_GRID = [
+    "###########",
+    "#S.O....C.#",
+    "#.........#",
+    "#..X...P..#",
+    "#.........#",
+    "#......V..#",
+    "#....######",
+    "#.....#..K#",
+    "###########",
+]
+
+#: Same pocket, empty floor inside — only the second spawn lives down there.
+POCKET_FLOOR_GRID = [
+    "###########",
+    "#S.O....C.#",
+    "#.........#",
+    "#..X...P..#",
+    "#.........#",
+    "#......V..#",
+    "#....######",
+    "#.....#...#",
+    "###########",
+]
+
+#: The serve window itself sealed inside the pocket (nothing else serves).
+SERVE_POCKET_GRID = [
+    "###########",
+    "#S.O....C.#",
+    "#.........#",
+    "#..X...P..#",
+    "#.........#",
+    "#....######",
+    "#.....#...#",
+    "#.....#..V#",
+    "###########",
+]
+
+NO_STOVE_GRID = [
+    "###########",
+    "#S.O....C.#",
+    "#.........#",
+    "#......P..#",
+    "#.........#",
+    "#......V..#",
+    "###########",
+]
+
+NO_ONION_GRID = [
+    "###########",
+    "#S......C.#",
+    "#.........#",
+    "#..X...P..#",
+    "#.........#",
+    "#......V..#",
+    "###########",
+]
+
+TINY_GRID = ["#####", "#SV.#", "#####"]
+
+TINY_TILES: dict[str, dict[str, str]] = {
+    ".": {"kind": "floor"},
+    "#": {"kind": "wall"},
+    "S": {"kind": "floor", "station": "spawn_shelf", "item": "tomato"},
+    "V": {"kind": "floor", "station": "serve"},
+}
+
+ORDERS: dict[str, Any] = {
+    "pool": ["soup"],
+    "concurrent": 1,
+    "spawn_interval": [16, 20],
+    "deadline": 60,
+}
+
+
+def level_data(**overrides: Any) -> dict[str, Any]:
+    data: dict[str, Any] = {
+        "id": "01_clean",
+        "name": "Clean Kitchen",
+        "order": 1,
+        "time_limit": 150,
+        "target_score": 45,
+        "difficulty": 1,
+        "grid": copy.deepcopy(CLEAN_GRID),
+        "tiles": copy.deepcopy(CLEAN_TILES),
+        "spawns": [[1, 1]],
+        "orders": copy.deepcopy(ORDERS),
+    }
+    data.update(overrides)
+    return data
+
+
+def make_level(**overrides: Any) -> Level:
+    """Build the clean fixture level with the given `LevelDef` overrides."""
+    return Level.build(LevelDef.model_validate(level_data(**overrides)))
+
+
+def make_book(*recipe_ids: str) -> RecipeBook:
+    ids = recipe_ids or ("soup",)
+    return RecipeBook.model_validate(
+        {"recipes": {recipe_id: copy.deepcopy(RECIPES[recipe_id]) for recipe_id in ids}}
+    )
+
+
+# --- the clean level is silent ------------------------------------------------------
+
+
+def test_clean_level_reports_no_errors_or_warnings() -> None:
+    report = validate_level(make_level(), make_book())
+    assert report.ok
+    assert report.violations == ()
+    assert report.warnings == ()
+    assert report.render() == "content OK: no violations"
+
+
+# --- failing cases, one per rule id ---------------------------------------------------
+
+Case = Callable[[], tuple[Level, RecipeBook, Level | None]]
+
+
+def _pocket_level() -> Level:
+    tiles = copy.deepcopy(CLEAN_TILES)
+    tiles["K"] = {"kind": "floor", "station": "sink"}
+    return make_level(grid=POCKET_GRID, tiles=tiles)
+
+
+def _grid_shape_height() -> tuple[Level, RecipeBook, Level | None]:
+    level = make_level()
+    shrunk = Grid(
+        width=level.grid.width,
+        height=level.grid.height - 1,
+        tiles=level.grid.tiles[: level.grid.width * (level.grid.height - 1)],
+    )
+    return replace(level, grid=shrunk), make_book(), None
+
+
+def _spawn_on_floor() -> tuple[Level, RecipeBook, Level | None]:
+    return make_level(spawns=[[0, 0]]), make_book(), None
+
+
+def _spawn_count() -> tuple[Level, RecipeBook, Level | None]:
+    return replace(make_level(), spawns=()), make_book(), None
+
+
+def _station_reachable() -> tuple[Level, RecipeBook, Level | None]:
+    return _pocket_level(), make_book(), None
+
+
+def _serve_reachable() -> tuple[Level, RecipeBook, Level | None]:
+    return make_level(grid=SERVE_POCKET_GRID), make_book(), None
+
+
+def _spawns_mutual() -> tuple[Level, RecipeBook, Level | None]:
+    level = make_level(grid=POCKET_FLOOR_GRID, spawns=[[1, 1], [9, 7], [5, 4]])
+    return level, make_book(), None
+
+
+def _recipe_station_missing() -> tuple[Level, RecipeBook, Level | None]:
+    tiles = {c: spec for c, spec in copy.deepcopy(CLEAN_TILES).items() if c != "X"}
+    return make_level(grid=NO_STOVE_GRID, tiles=tiles), make_book(), None
+
+
+def _recipe_ingredient_missing() -> tuple[Level, RecipeBook, Level | None]:
+    tiles = {c: spec for c, spec in copy.deepcopy(CLEAN_TILES).items() if c != "O"}
+    return make_level(grid=NO_ONION_GRID, tiles=tiles), make_book(), None
+
+
+def _pool_unknown() -> tuple[Level, RecipeBook, Level | None]:
+    level = make_level(orders=dict(ORDERS, pool=["soup", "ghost"]))
+    return level, make_book(), None
+
+
+def _target_too_high() -> tuple[Level, RecipeBook, Level | None]:
+    book = make_book()
+    cap = theoretical_max_score(make_level(), book)
+    assert cap is not None
+    return make_level(target_score=int(cap * 2)), book, None
+
+
+def _time_limit() -> tuple[Level, RecipeBook, Level | None]:
+    return make_level(time_limit=10), make_book(), None
+
+
+def _hazard_ref() -> tuple[Level, RecipeBook, Level | None]:
+    level = make_level(hazards=[{"type": "fire", "station": "stove_9"}])
+    return level, make_book(), None
+
+
+def _hazard_tile() -> tuple[Level, RecipeBook, Level | None]:
+    level = make_level(hazards=[{"type": "spill", "x": 0, "y": 0}])
+    return level, make_book(), None
+
+
+def _limit_zero() -> tuple[Level, RecipeBook, Level | None]:
+    return make_level(station_limits={"plates": 0, "pots": 1}), make_book(), None
+
+
+def _difficulty_order() -> tuple[Level, RecipeBook, Level | None]:
+    return make_level(), make_book(), make_level()
+
+
+def _travel_short() -> tuple[Level, RecipeBook, Level | None]:
+    level = make_level(difficulty=3, grid=TINY_GRID, tiles=TINY_TILES)
+    return level, make_book(), None
+
+
+FAILING: dict[str, Case] = {
+    RuleId.GRID_SHAPE: _grid_shape_height,
+    RuleId.SPAWN_ON_FLOOR: _spawn_on_floor,
+    RuleId.SPAWN_COUNT: _spawn_count,
+    RuleId.STATION_REACHABLE: _station_reachable,
+    RuleId.SERVE_REACHABLE: _serve_reachable,
+    RuleId.SPAWNS_MUTUALLY_REACHABLE: _spawns_mutual,
+    RuleId.RECIPE_STATION_MISSING: _recipe_station_missing,
+    RuleId.RECIPE_INGREDIENT_UNAVAILABLE: _recipe_ingredient_missing,
+    RuleId.ORDER_POOL_UNKNOWN_RECIPE: _pool_unknown,
+    RuleId.TARGET_IMPLAUSIBLE: _target_too_high,
+    RuleId.TIME_LIMIT_INVALID: _time_limit,
+    RuleId.HAZARD_REF_UNKNOWN: _hazard_ref,
+    RuleId.HAZARD_ON_UNREACHABLE_TILE: _hazard_tile,
+    RuleId.STATION_LIMIT_ZERO: _limit_zero,
+    RuleId.DIFFICULTY_ORDER: _difficulty_order,
+    RuleId.TRAVEL_TOO_SHORT: _travel_short,
+}
+
+
+def test_failing_cases_cover_every_rule_id() -> None:
+    """Every rule id has a failing case here; DUPLICATE_ID is a bundle rule."""
+    assert set(FAILING) | {RuleId.DUPLICATE_ID} == set(RuleId.ALL)
+    assert len(FAILING) + 1 == len(RuleId.ALL)
+
+
+@pytest.mark.parametrize("rule", sorted(FAILING))
+def test_rule_fires_with_a_human_message(rule: str) -> None:
+    level, book, prev = FAILING[rule]()
+    report = validate_level(level, book, prev)
+    assert not report.ok
+    fired = [violation for violation in report.violations if violation.rule == rule]
+    assert fired, f"{rule} did not fire:\n{report.render()}"
+
+
+def test_planted_unreachable_station_reports_exactly_station_reachable() -> None:
+    report = validate_level(_pocket_level(), make_book())
+    assert len(report.violations) == 1, report.render()
+    violation = report.violations[0]
+    assert violation.rule == RuleId.STATION_REACHABLE
+    assert "(9,7)" in violation.message
+    assert dict(violation.detail) == {"x": 9, "y": 7, "station_id": "sink_0"}
+    assert report.errors_for("01_clean") == (violation,)
+    assert report.errors_for("nope") == ()
+
+
+# --- passing counterparts beyond the clean level ---------------------------------------
+
+
+def test_grid_shape_also_flags_a_width_disagreement() -> None:
+    level = make_level()
+    wider = Grid(
+        width=level.grid.width + 1,
+        height=level.grid.height,
+        tiles=level.grid.tiles + (TileKind.WALL,) * level.grid.height,
+    )
+    report = validate_level(replace(level, grid=wider), make_book())
+    fired = [v.rule for v in report.violations]
+    assert RuleId.GRID_SHAPE in fired
+
+
+def test_spawn_on_the_serve_window_travels_zero_tiles() -> None:
+    assert escalation_axes(make_level(spawns=[[7, 5]]))["travel_distance"] == 0.0
+
+
+def test_non_raw_ingredients_are_not_shelf_checked() -> None:
+    level = make_level(orders=dict(ORDERS, pool=["salad"]))
+    report = validate_level(level, make_book("salad"))
+    assert report.ok, report.render()
+
+
+def test_tight_target_warns_without_failing() -> None:
+    cap = theoretical_max_score(make_level(), make_book())
+    assert cap is not None
+    tight = make_level(target_score=int(TARGET_WARN_FRACTION * cap) + 10)
+    report = validate_level(tight, make_book())
+    assert report.ok  # warnings never fail
+    assert [warning.rule for warning in report.warnings] == [RuleId.TARGET_IMPLAUSIBLE]
+
+
+def test_two_axes_moved_satisfies_difficulty_order() -> None:
+    prev = make_level()
+    nxt = make_level(
+        order=2,
+        hazards=[{"type": "spill", "x": 5, "y": 2}],
+        orders=dict(ORDERS, concurrent=2),
+    )
+    report = validate_level(nxt, make_book(), prev)
+    assert report.ok, report.render()
+
+
+def test_high_difficulty_level_with_real_travel_is_clean() -> None:
+    report = validate_level(make_level(difficulty=3), make_book())
+    assert report.ok, report.render()
+
+
+def test_hazard_path_point_off_the_floor_fires() -> None:
+    level = make_level(
+        hazards=[{"type": "moving_obstacle", "path": [[1, 1], [0, 0]], "period_s": 8.0}]
+    )
+    fired = [
+        violation.message
+        for violation in validate_level(level, make_book()).violations
+        if violation.rule == RuleId.HAZARD_ON_UNREACHABLE_TILE
+    ]
+    assert len(fired) == 1
+    assert "path point 1" in fired[0]
+
+
+def test_hazard_outside_the_grid_fires() -> None:
+    level = make_level(hazards=[{"type": "spill", "x": 99, "y": 99}])
+    fired = [
+        violation.message
+        for violation in validate_level(level, make_book()).violations
+        if violation.rule == RuleId.HAZARD_ON_UNREACHABLE_TILE
+    ]
+    assert len(fired) == 1
+    assert "outside the 11x7 grid" in fired[0]  # clean fixture is 11x7
+
+
+# --- bundle-level rules -------------------------------------------------------------
+
+
+def test_duplicate_id_is_a_bundle_violation() -> None:
+    first = make_level()
+    second = make_level(order=2)  # forgets to rename: same id again
+    report = validate_levels([first, second], make_book())
+    dupes = [violation for violation in report.violations if violation.rule == RuleId.DUPLICATE_ID]
+    assert len(dupes) == 1
+    assert dupes[0].level_id == "01_clean"
+    assert dupes[0].detail["count"] == 2
+
+
+def test_validate_bundle_accepts_a_clean_two_level_campaign() -> None:
+    first = make_level()
+    second = make_level(
+        id="02_later",
+        order=2,
+        difficulty=2,
+        hazards=[{"type": "spill", "x": 5, "y": 2}],
+        orders=dict(ORDERS, concurrent=2),
+    )
+    bundle = ContentBundle(recipe_book=make_book(), levels=(first, second))
+    report = validate_bundle(bundle)
+    assert report.ok and not report.warnings, report.render()
+
+
+def test_validate_bundle_refuses_a_difficulty_drop() -> None:
+    first = make_level(difficulty=2)
+    second = make_level(id="02_later", order=2, difficulty=1)
+    bundle = ContentBundle(recipe_book=make_book(), levels=(first, second))
+    report = validate_bundle(bundle)
+    drops = [
+        violation
+        for violation in report.violations
+        if violation.rule == RuleId.DIFFICULTY_ORDER and "drops" in violation.message
+    ]
+    assert len(drops) == 1
+    assert drops[0].level_id == "02_later"
+
+
+# --- the bound itself ---------------------------------------------------------------
+
+
+def test_theoretical_max_score_needs_known_recipes_and_valid_time() -> None:
+    book = make_book()
+    ghost = make_level(orders=dict(ORDERS, pool=["ghost"]))
+    assert theoretical_max_score(ghost, book) is None
+    assert theoretical_max_score(make_level(time_limit=10), book) is None
+
+
+def test_theoretical_max_score_counts_orders_the_time_allows() -> None:
+    book = make_book()
+    level = make_level()
+    cap = theoretical_max_score(level, book)
+    assert cap is not None
+    # 150 s / (7.0 s soup + 2*10 tiles / 3.5 tiles/s travel) -> 11 orders x 40 pts.
+    assert cap == 11 * 40
+
+
+def test_escalation_axes_are_named_and_higher_is_harder() -> None:
+    level = make_level()
+    axes = escalation_axes(level)
+    assert set(axes) == {
+        "recipe_complexity",
+        "order_pressure",
+        "time_budget",
+        "station_scarcity",
+        "layout_hostility",
+        "hazard_density",
+        "travel_distance",
+    }
+    assert axes["hazard_density"] == 0.0
+    assert axes["travel_distance"] == 10.0
+    harder = make_level(
+        hazards=[{"type": "spill", "x": 5, "y": 2}],
+        orders=dict(ORDERS, concurrent=2),
+    )
+    harder_axes = escalation_axes(harder)
+    assert harder_axes["hazard_density"] > axes["hazard_density"]
+    assert harder_axes["order_pressure"] > axes["order_pressure"]

tests/sim/test_validation_render.py

@@ -0,0 +1,81 @@
+"""`ValidationReport.render()` is deterministic, rule-naming, and bounded.
+
+The CLI (task 05) prints this text verbatim and the content-gate criterion
+greps it for rule names, so the render must be stable (same report, same
+bytes — input order cannot leak in), mention every rule id it holds, stay
+free of ANSI escapes, and not explode into a wall of text for a broken level.
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from undercooked.sim.validation import LevelViolation, RuleId, ValidationReport
+
+pytestmark = pytest.mark.unit
+
+
+def _violation(rule: str, level_id: str, message: str) -> LevelViolation:
+    return LevelViolation(rule=rule, level_id=level_id, message=message)
+
+
+def _ten_violation_report() -> ValidationReport:
+    errors = tuple(
+        _violation(rule, f"level_{index % 2}", f"synthetic {rule} message {index}")
+        for index, rule in enumerate(RuleId.ALL[:10])
+    )
+    warnings = (_violation(RuleId.TARGET_IMPLAUSIBLE, "level_0", "tight but fair"),)
+    return ValidationReport(violations=errors, warnings=warnings)
+
+
+def test_render_is_deterministic_across_runs_and_input_order() -> None:
+    report = _ten_violation_report()
+    assert report.render() == report.render()
+    shuffled = ValidationReport(
+        violations=tuple(reversed(report.violations)),
+        warnings=report.warnings,
+    )
+    assert shuffled.render() == report.render()
+
+
+def test_render_names_every_rule_level_and_severity_without_ansi() -> None:
+    rendered = _ten_violation_report().render()
+    for violation in _ten_violation_report().violations:
+        assert violation.rule in rendered
+        assert violation.message in rendered
+    assert "level_0" in rendered and "level_1" in rendered
+    assert "ERROR" in rendered and "WARN" in rendered
+    assert "\x1b" not in rendered
+
+
+def test_render_stays_under_40_lines_for_a_10_violation_level() -> None:
+    assert len(_ten_violation_report().render().splitlines()) <= 40
+
+
+def test_render_orders_errors_by_rule_within_a_level() -> None:
+    report = ValidationReport(
+        violations=(
+            _violation(RuleId.TRAVEL_TOO_SHORT, "l1", "b travels too little"),
+            _violation(RuleId.GRID_SHAPE, "l1", "a shape disagrees"),
+        )
+    )
+    error_lines = [line for line in report.render().splitlines() if "ERROR" in line]
+    assert len(error_lines) == 2
+    assert RuleId.GRID_SHAPE in error_lines[0]
+    assert RuleId.TRAVEL_TOO_SHORT in error_lines[1]
+
+
+def test_clean_report_renders_one_ok_line() -> None:
+    report = ValidationReport()
+    assert report.ok
+    assert report.render() == "content OK: no violations"
+
+
+def test_warnings_only_report_renders_ok_and_warn_lines() -> None:
+    report = ValidationReport(
+        warnings=(_violation(RuleId.TARGET_IMPLAUSIBLE, "l1", "tight but fair"),)
+    )
+    rendered = report.render()
+    assert report.ok  # warnings never fail
+    assert "WARN" in rendered and "ERROR" not in rendered
+    assert "0 error(s), 1 warning(s)" in rendered

tests/unit/test_content_loader.py

@@ -0,0 +1,281 @@
+"""Unit tests for `undercooked.sim.content` (phase 02 task 02).
+
+The loader is the "no hard-coded content lists" rule (A8) in executable form:
+a directory of YAML in, a campaign-ordered `ContentBundle` out. These tests
+pin discovery (recursion, `_`-drafts, the recipe book's own name), the
+`(order, id)` sort, and the promise that *every* file-level problem — bad
+YAML, schema violations, unknown order-pool recipes, duplicate ids — is
+reported in one `ContentError`, not one per run.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Any
+
+import pytest
+import yaml
+
+from undercooked.sim.content import (
+    LEVEL_FILE_SUFFIXES,
+    ContentBundle,
+    ContentError,
+    load_content,
+    load_level_file,
+)
+from undercooked.sim.grid import GridError
+from undercooked.sim.level import Level, LevelDef
+from undercooked.sim.model import ItemKind
+from undercooked.sim.recipe import RecipeBook
+
+pytestmark = pytest.mark.unit
+
+BOOK_YAML = """
+ingredients: [tomato, onion]
+recipes:
+  soup:
+    id: soup
+    name: Tomato Soup
+    base_score: 40
+    ingredients: [{kind: tomato, form: raw}, {kind: onion, form: raw}]
+    steps:
+      - {kind: chop, station: chop_board, seconds: 0.9}
+      - {kind: cook, station: stove, seconds: 5}
+"""
+
+
+def level_body(level_id: str, order: int, **overrides: Any) -> dict[str, Any]:
+    body: dict[str, Any] = {
+        "id": level_id,
+        "name": level_id.replace("_", " ").title(),
+        "order": order,
+        "time_limit": 150,
+        "target_score": 45,
+        "grid": ["#####", "#S..#", "#.C.#", "#..V#", "#####"],
+        "tiles": {
+            ".": {"kind": "floor"},
+            "#": {"kind": "wall"},
+            "S": {"kind": "floor", "station": "spawn_shelf", "item": "tomato"},
+            "C": {"kind": "floor", "station": "chop_board"},
+            "V": {"kind": "floor", "station": "serve"},
+        },
+        "spawns": [[1, 1]],
+        "orders": {
+            "pool": ["soup"],
+            "concurrent": 1,
+            "spawn_interval": [16, 20],
+            "deadline": 60,
+        },
+    }
+    body.update(overrides)
+    return body
+
+
+def write_yaml(directory: Path, name: str, body: dict[str, Any]) -> Path:
+    path = directory / name
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(yaml.safe_dump(body, sort_keys=False), encoding="utf-8")
+    return path
+
+
+def make_content_dir(tmp_path: Path, *levels: tuple[str, int]) -> tuple[Path, Path]:
+    """A content dir with a recipe book and the named (filename, order) levels."""
+    levels_dir = tmp_path / "levels"
+    levels_dir.mkdir()
+    recipes_path = levels_dir / "recipes.yaml"
+    recipes_path.write_text(BOOK_YAML, encoding="utf-8")
+    for name, order in levels:
+        write_yaml(levels_dir, name, level_body(Path(name).stem, order))
+    return levels_dir, recipes_path
+
+
+# ---------------------------------------------------------------------------- happy path
+
+
+def test_loads_the_book_and_every_level_file(tmp_path: Path) -> None:
+    levels_dir, recipes_path = make_content_dir(tmp_path, ("01_a.yaml", 1), ("02_b.yaml", 2))
+    bundle = load_content(levels_dir, recipes_path)
+    assert bundle.recipe_book.ids() == ("soup",)
+    assert [level.id for level in bundle.levels] == ["01_a", "02_b"]
+    assert set(bundle.by_id) == {"01_a", "02_b"}
+    assert bundle.by_id["02_b"].order == 2
+
+
+def test_levels_order_by_order_then_id_not_by_filename(tmp_path: Path) -> None:
+    levels_dir, recipes_path = make_content_dir(
+        tmp_path, ("zzz_first.yaml", 1), ("01_second.yaml", 2), ("aaa_tie_b.yaml", 1)
+    )
+    write_yaml(levels_dir, "aab_tie.yaml", level_body("aab_tie", 1))
+    bundle = load_content(levels_dir, recipes_path)
+    assert [level.id for level in bundle.levels] == [
+        "aaa_tie_b",
+        "aab_tie",
+        "zzz_first",
+        "01_second",
+    ]
+
+
+def test_an_empty_levels_dir_is_a_legal_empty_campaign(tmp_path: Path) -> None:
+    levels_dir, recipes_path = make_content_dir(tmp_path)
+    bundle = load_content(levels_dir, recipes_path)
+    assert bundle.levels == ()
+    assert dict(bundle.by_id) == {}
+
+
+def test_paths_may_be_plain_strings(tmp_path: Path) -> None:
+    levels_dir, recipes_path = make_content_dir(tmp_path, ("01_a.yaml", 1))
+    bundle = load_content(str(levels_dir), str(recipes_path))
+    assert [level.id for level in bundle.levels] == ["01_a"]
+
+
+# --------------------------------------------------------------------------- discovery
+
+
+def test_discovery_recurses_into_subdirectories(tmp_path: Path) -> None:
+    levels_dir, recipes_path = make_content_dir(tmp_path, ("01_a.yaml", 1))
+    write_yaml(levels_dir / "campaign" / "deep", "02_b.yaml", level_body("02_b", 2))
+    bundle = load_content(levels_dir, recipes_path)
+    assert [level.id for level in bundle.levels] == ["01_a", "02_b"]
+
+
+def test_underscore_prefixed_files_are_ignored_even_when_broken(tmp_path: Path) -> None:
+    levels_dir, recipes_path = make_content_dir(tmp_path, ("01_a.yaml", 1))
+    (levels_dir / "_draft.yaml").write_text("grid: [unclosed\n", encoding="utf-8")
+    (levels_dir / "_wip").mkdir()
+    write_yaml(levels_dir / "_wip", "scratch.yaml", {"id": "nothing"})
+    bundle = load_content(levels_dir, recipes_path)
+    assert [level.id for level in bundle.levels] == ["01_a"]
+
+
+def test_recipes_yaml_in_the_levels_dir_is_not_a_level(tmp_path: Path) -> None:
+    levels_dir, recipes_path = make_content_dir(tmp_path)
+    # recipes.yaml itself would not validate as a level; loading must not even try.
+    bundle = load_content(levels_dir, recipes_path)
+    assert bundle.levels == ()
+
+
+def test_non_yaml_files_are_ignored(tmp_path: Path) -> None:
+    levels_dir, recipes_path = make_content_dir(tmp_path, ("01_a.yaml", 1))
+    (levels_dir / "notes.md").write_text("# ideas\n- more soup\n", encoding="utf-8")
+    (levels_dir / "level.json").write_text("{not yaml at all", encoding="utf-8")
+    bundle = load_content(levels_dir, recipes_path)
+    assert [level.id for level in bundle.levels] == ["01_a"]
+
+
+def test_suffix_set_covers_yaml_and_yml(tmp_path: Path) -> None:
+    assert set(LEVEL_FILE_SUFFIXES) == {".yaml", ".yml"}
+    levels_dir, recipes_path = make_content_dir(tmp_path)
+    (levels_dir / "02_b.yml").write_text(
+        yaml.safe_dump(level_body("02_b", 2), sort_keys=False), encoding="utf-8"
+    )
+    bundle = load_content(levels_dir, recipes_path)
+    assert [level.id for level in bundle.levels] == ["02_b"]
+
+
+# ---------------------------------------------------------------- error aggregation
+
+
+def test_unknown_order_pool_recipe_is_a_load_error(tmp_path: Path) -> None:
+    levels_dir, recipes_path = make_content_dir(tmp_path)
+    write_yaml(levels_dir, "01_a.yaml", level_body("01_a", 1, orders={
+        "pool": ["soup", "ghost"],
+        "concurrent": 1,
+        "spawn_interval": [16, 20],
+        "deadline": 60,
+    }))
+    with pytest.raises(ContentError, match="unknown recipes: ghost"):
+        load_content(levels_dir, recipes_path)
+
+
+def test_duplicate_level_ids_are_refused(tmp_path: Path) -> None:
+    levels_dir, recipes_path = make_content_dir(tmp_path, ("01_a.yaml", 1))
+    write_yaml(levels_dir, "90_copy.yaml", level_body("01_a", 5))
+    with pytest.raises(ContentError, match="duplicate level ids: 01_a"):
+        load_content(levels_dir, recipes_path)
+
+
+def test_missing_levels_directory_is_reported(tmp_path: Path) -> None:
+    _, recipes_path = make_content_dir(tmp_path)
+    with pytest.raises(ContentError, match="does not exist"):
+        load_content(tmp_path / "nowhere", recipes_path)
+
+
+def test_missing_recipe_book_is_reported(tmp_path: Path) -> None:
+    levels_dir, _ = make_content_dir(tmp_path)
+    with pytest.raises(ContentError, match="cannot read content file"):
+        load_content(levels_dir, tmp_path / "no" / "recipes.yaml")
+
+
+def test_all_file_errors_survive_to_one_content_error(tmp_path: Path) -> None:
+    levels_dir, recipes_path = make_content_dir(tmp_path)
+    (levels_dir / "bad_one.yaml").write_text("grid: [unclosed\n", encoding="utf-8")
+    write_yaml(levels_dir, "bad_two.yaml", {"id": "x"})  # mostly-missing schema
+    write_yaml(levels_dir, "bad_three.yaml", level_body("ghost", 1, orders={
+        "pool": ["phantom"],
+        "concurrent": 1,
+        "spawn_interval": [16, 20],
+        "deadline": 60,
+    }))
+    with pytest.raises(ContentError) as excinfo:
+        load_content(levels_dir, recipes_path)
+    message = str(excinfo.value)
+    assert "bad_one.yaml" in message
+    assert "invalid YAML" in message
+    assert "bad_two.yaml" in message
+    assert "Field required" in message
+    assert "bad_three.yaml" in message
+    assert "phantom" in message
+    # every broken file opens its own issue block
+    assert f"{levels_dir / 'bad_one.yaml'}: " in message
+    assert f"{levels_dir / 'bad_two.yaml'}: " in message
+    assert f"{levels_dir / 'bad_three.yaml'}: " in message
+
+
+# ------------------------------------------------------------------ load_level_file
+
+
+def test_load_level_file_returns_a_built_level(tmp_path: Path) -> None:
+    path = write_yaml(tmp_path, "01_a.yaml", level_body("01_a", 1))
+    level = load_level_file(path)
+    assert isinstance(level, Level)
+    assert level.station_items == {"spawn_shelf_0": ItemKind.TOMATO}
+    assert level.grid.width == 5
+
+
+def test_load_level_file_names_the_file_on_schema_errors(tmp_path: Path) -> None:
+    path = write_yaml(tmp_path, "broken.yaml", level_body("x", 1, grid=["###", "#..#"]))
+    with pytest.raises(ContentError, match=r"broken.yaml.*ragged grid"):
+        load_level_file(path)
+
+
+def test_load_level_file_wraps_geometry_failures(tmp_path: Path, monkeypatch: Any) -> None:
+    def explode(_spec: Any) -> Level:
+        raise GridError("planted geometry failure")
+
+    monkeypatch.setattr(Level, "build", staticmethod(explode))
+    path = write_yaml(tmp_path, "01_a.yaml", level_body("01_a", 1))
+    with pytest.raises(ContentError, match=r"01_a\.yaml: planted geometry failure"):
+        load_level_file(path)
+
+
+# --------------------------------------------------------------------- ContentBundle
+
+
+def empty_book() -> RecipeBook:
+    return RecipeBook.model_validate({"recipes": {}})
+
+
+def built_level(level_id: str, order: int) -> Level:
+    return Level.build(LevelDef.model_validate(level_body(level_id, order)))
+
+
+def test_bundle_by_id_is_derived_and_can_be_read_back() -> None:
+    bundle = ContentBundle(recipe_book=empty_book(), levels=(built_level("a", 2),))
+    assert list(bundle.by_id) == ["a"]
+    assert bundle.by_id["a"].order == 2
+
+
+def test_a_hand_built_bundle_cannot_smuggle_duplicate_ids() -> None:
+    levels = (built_level("a", 1), built_level("a", 2))
+    with pytest.raises(ContentError, match="duplicate level ids: a"):
+        ContentBundle(recipe_book=empty_book(), levels=levels)

tests/unit/test_content_not_hardcoded.py

@@ -0,0 +1,234 @@
+"""A8, enforced: the campaign is YAML on disk and never a list in Python.
+
+Phase 02 task 05's regression lock. The loader (`sim/content.py`), the content
+API (`server/routes_content.py`) and the author CLI (`tools/levels.py`) all
+discover levels by *reading a directory*, so a new `.yaml` file is the only
+thing an author writes to ship a level. This test proves that twice over:
+
+* **addition without code** — copy `levels/` to `tmp_path`, drop in
+  `90_zerocool.yaml` with `order: 99`, and the loader, `/api/levels` and
+  `undercooked-levels list`/`show`/`validate` all carry it, while the id
+  appears in no Python file at all (and deleting the file removes it again);
+* **absence of literals** — no file under `src/undercooked/` (and no browser
+  asset) contains a shipped level id or recipe id as a literal, so a hard-coded
+  list cannot be quietly re-added later.
+
+The session fixture points `UNDERCOOKED_LEVELS_DIR` at the empty
+`tests/fixtures/levels/` directory, so every path here is set explicitly.
+"""
+
+from __future__ import annotations
+
+import shutil
+from collections.abc import Mapping
+from pathlib import Path
+
+import pytest
+from fastapi.testclient import TestClient
+
+from undercooked.config import load_settings
+from undercooked.server.app import create_app
+from undercooked.server.routes_content import RECIPES_FILENAME
+from undercooked.sim.content import load_content
+from undercooked.sim.recipe import load_recipe_book
+from undercooked.tools.levels import EXIT_OK, EXIT_UNUSABLE, main
+
+pytestmark = pytest.mark.unit
+
+REPO_ROOT = Path(__file__).resolve().parents[2]
+SHIPPED_LEVELS = REPO_ROOT / "levels"
+SHIPPED_RECIPES = SHIPPED_LEVELS / RECIPES_FILENAME
+SOURCE_ROOT = REPO_ROOT / "src" / "undercooked"
+STATIC_ROOT = REPO_ROOT / "app" / "static"
+
+
+def shipped_level_ids() -> set[str]:
+    """Every id the shipped `levels/` directory carries (the book is not a level)."""
+    return {
+        path.stem
+        for path in SHIPPED_LEVELS.glob("*.yaml")
+        if path.name not in {RECIPES_FILENAME, "recipes.yml"}
+    }
+
+#: The level this test adds. `order: 99` puts it last in the campaign whatever
+#: the shipped levels do, and three dishes with three orders live at once move
+#: the escalation axes the campaign rule (`DIFFICULTY_ORDER`) asks about.
+NEW_LEVEL_ID = "90_zerocool"
+NEW_LEVEL_YAML = """\
+id: 90_zerocool
+name: Zero Cool
+order: 99
+difficulty: 3
+time_limit: 120
+target_score: 110
+blurb: >-
+  A manifest probe, not a campaign level: three dishes, three orders live at
+  once, and the same pinched floor as the second kitchen.
+grid:
+  - "#############"
+  - "#T....C....O#"
+  - "#.....S.....#"
+  - "####=...=####"
+  - "#....P.X....#"
+  - "#...........#"
+  - "####=...=####"
+  - "#R.K.....L..#"
+  - "#############"
+tiles:
+  "#": {kind: wall}
+  ".": {kind: floor}
+  "=": {kind: counter}
+  "T": {kind: floor, station: spawn_shelf, item: tomato}
+  "O": {kind: floor, station: spawn_shelf, item: onion}
+  "L": {kind: floor, station: spawn_shelf, item: lettuce}
+  "R": {kind: floor, station: spawn_shelf, item: rice}
+  "C": {kind: floor, station: chop_board}
+  "S": {kind: floor, station: stove}
+  "P": {kind: floor, station: prep}
+  "X": {kind: floor, station: serve}
+  "K": {kind: floor, station: sink}
+spawns: [[2, 2], [10, 5]]
+orders:
+  pool: [soup, salad, fried_rice]
+  concurrent: 3
+  spawn_interval: [10, 14]
+  deadline: 45
+hazards: []
+escalation_notes: >-
+  Moves recipe_complexity, order_pressure and time_budget up versus the last
+  shipped level, which is what the campaign rule asks for.
+"""
+
+
+def copy_shipped_levels(destination: Path) -> Path:
+    """`levels/` copied verbatim — making a level here edits no Python."""
+    shutil.copytree(SHIPPED_LEVELS, destination)
+    return destination
+
+
+def text_files(root: Path, suffixes: tuple[str, ...]) -> Mapping[str, str]:
+    """Every file under `root` with one of `suffixes`, keyed by path."""
+    return {
+        str(path): path.read_text(encoding="utf-8")
+        for path in sorted(root.rglob("*"))
+        if path.is_file()
+        and path.suffix in suffixes
+        and "__pycache__" not in path.parts
+    }
+
+
+def offenders(literals: set[str], texts: Mapping[str, str]) -> dict[str, list[str]]:
+    """Which literals appear in which file — empty means the scan passed."""
+    return {
+        path: sorted(literal for literal in literals if literal in text)
+        for path, text in texts.items()
+        if any(literal in text for literal in literals)
+    }
+
+
+@pytest.fixture()
+def extended_campaign(tmp_path: Path) -> Path:
+    """A copy of the shipped campaign plus one level nothing ever registered."""
+    campaign = copy_shipped_levels(tmp_path / "campaign")
+    (campaign / f"{NEW_LEVEL_ID}.yaml").write_text(NEW_LEVEL_YAML, encoding="utf-8")
+    return campaign
+
+
+# --- a file is the only thing an author writes ----------------------------------
+
+
+def test_a_new_yaml_file_reaches_the_loader_the_api_and_the_cli_without_code(
+    extended_campaign: Path,
+    monkeypatch: pytest.MonkeyPatch,
+    capsys: pytest.CaptureFixture[str],
+) -> None:
+    shipped_ids = [level.id for level in load_content(SHIPPED_LEVELS, SHIPPED_RECIPES).levels]
+
+    # 1. the pure loader
+    bundle = load_content(extended_campaign, extended_campaign / RECIPES_FILENAME)
+    assert [level.id for level in bundle.levels] == [*shipped_ids, NEW_LEVEL_ID]
+    assert list(bundle.by_id) == [*shipped_ids, NEW_LEVEL_ID]
+    assert bundle.by_id[NEW_LEVEL_ID].order == 99
+
+    # 2. the HTTP content API, addressed only through the environment
+    monkeypatch.setenv("UNDERCOOKED_LEVELS_DIR", str(extended_campaign))
+    monkeypatch.delenv("UNDERCOOKED_RECIPES_PATH", raising=False)
+    with TestClient(create_app(load_settings())) as client:
+        listing = client.get("/api/levels")
+        assert listing.status_code == 200
+        assert [level["id"] for level in listing.json()] == [*shipped_ids, NEW_LEVEL_ID]
+        detail = client.get(f"/api/levels/{NEW_LEVEL_ID}")
+        assert detail.status_code == 200
+        assert detail.json()["order"] == 99
+        assert detail.json()["difficulty"] == 3
+
+    # 3. the author CLI, on the same environment and no `--dir`
+    assert main(["list"]) == EXIT_OK
+    assert capsys.readouterr().out.splitlines()[-1].split()[1] == NEW_LEVEL_ID
+
+    assert main(["show", NEW_LEVEL_ID, "--grid"]) == EXIT_OK
+    shown = capsys.readouterr().out
+    assert shown.startswith(f"{NEW_LEVEL_ID} - Zero Cool")
+    assert "concurrent=3" in shown
+
+    # 4. and the gate accepts it: a row in a table that breaks PLAN §8 is not a level
+    assert main(["validate"]) == EXIT_OK
+    assert NEW_LEVEL_ID in capsys.readouterr().out
+
+    # The proof of the whole point: the new level is everywhere except the code.
+    assert offenders({NEW_LEVEL_ID}, text_files(SOURCE_ROOT, (".py", ".pyi"))) == {}
+
+
+def test_the_copied_campaign_on_its_own_is_still_the_shipped_two(tmp_path: Path) -> None:
+    """The copy is faithful, so the test above measures exactly one added file."""
+    pristine = copy_shipped_levels(tmp_path / "pristine")
+    copied = load_content(pristine, pristine / RECIPES_FILENAME)
+    shipped = load_content(SHIPPED_LEVELS, SHIPPED_RECIPES)
+    assert [level.id for level in copied.levels] == [level.id for level in shipped.levels]
+
+
+def test_a_deletion_is_as_invisible_as_an_addition(
+    extended_campaign: Path, capsys: pytest.CaptureFixture[str]
+) -> None:
+    """Nothing caches the campaign in Python either: remove the file, it is gone."""
+    (extended_campaign / f"{NEW_LEVEL_ID}.yaml").unlink()
+    assert main(["list", "--dir", str(extended_campaign)]) == EXIT_OK
+    assert NEW_LEVEL_ID not in capsys.readouterr().out
+
+    assert main(["show", NEW_LEVEL_ID, "--dir", str(extended_campaign)]) == EXIT_UNUSABLE
+    assert NEW_LEVEL_ID not in load_content(
+        extended_campaign, extended_campaign / RECIPES_FILENAME
+    ).by_id
+
+
+# --- no literals in the source --------------------------------------------------
+
+
+def test_no_python_source_names_a_shipped_level_id() -> None:
+    """`levels/*.yaml` owns level ids: nothing in `src/undercooked/` repeats one."""
+    shipped = load_content(SHIPPED_LEVELS, SHIPPED_RECIPES)
+    literals = shipped_level_ids() | {level.id for level in shipped.levels}
+    assert literals  # a scan of an empty set would pass and prove nothing
+    assert offenders(literals, text_files(SOURCE_ROOT, (".py", ".pyi"))) == {}
+
+
+def test_no_python_source_names_a_recipe_from_the_book() -> None:
+    """A8 covers recipes too: `recipes.yaml` is the only list of dishes."""
+    literals = set(load_recipe_book(SHIPPED_RECIPES).ids())
+    assert literals
+    assert offenders(literals, text_files(SOURCE_ROOT, (".py", ".pyi"))) == {}
+
+
+def test_no_browser_asset_names_a_shipped_level_id() -> None:
+    """The client asks the API for the campaign; it does not carry one."""
+    literals = shipped_level_ids()
+    assets = text_files(STATIC_ROOT, (".js", ".css", ".html", ".svg"))
+    assert assets
+    assert offenders(literals, assets) == {}
+
+
+def test_the_loader_discovers_levels_by_walking_the_directory() -> None:
+    """The discovery mechanism behind every test above: a directory walk."""
+    loader = (SOURCE_ROOT / "sim" / "content.py").read_text(encoding="utf-8")
+    assert "rglob" in loader
+    assert "iterdir" not in loader  # one walk, at any depth, not a flat listing

tests/unit/test_content_startup.py

@@ -0,0 +1,186 @@
+"""Startup content gate (task 04): broken content refuses to start the app.
+
+``create_app`` loads and validates the campaign once through
+``load_validated_content``. These tests drive that function and the factory
+directly: the shipped content loads; a level whose serve window is sealed
+into a wall pocket raises ``ContentError`` and prints the named-rule report
+on stderr; a missing recipe book does the same; and a *warning* band target
+score logs a warning without refusing startup.
+"""
+
+from __future__ import annotations
+
+import logging
+import shutil
+import textwrap
+from pathlib import Path
+
+import pytest
+
+from undercooked.config import Settings
+from undercooked.server.app import create_app
+from undercooked.server.routes_content import (
+    RECIPES_FILENAME,
+    load_validated_content,
+    resolve_recipes_path,
+)
+from undercooked.sim.recipe import ContentError
+
+REPO_ROOT = Path(__file__).resolve().parents[2]
+SHIPPED_LEVELS = REPO_ROOT / "levels"
+
+
+def make_settings(levels_dir: Path) -> Settings:
+    return Settings(levels_dir=levels_dir, recipes_path=levels_dir / RECIPES_FILENAME)
+
+
+# --- path resolution ----------------------------------------------------------
+
+
+def test_recipes_path_defaults_inside_the_levels_dir() -> None:
+    settings = Settings(levels_dir=Path("/content/campaign"))
+    assert settings.recipes_path is None  # the shipped default is "derive it"
+    assert resolve_recipes_path(settings) == Path("/content/campaign/recipes.yaml")
+
+
+def test_recipes_path_override_wins() -> None:
+    settings = Settings(
+        levels_dir=Path("/content/campaign"), recipes_path=Path("/other/book.yaml")
+    )
+    assert resolve_recipes_path(settings) == Path("/other/book.yaml")
+
+
+# --- the happy path ------------------------------------------------------------
+
+
+def test_startup_loads_the_shipped_bundle() -> None:
+    application = create_app(Settings(levels_dir=SHIPPED_LEVELS))
+    loaded = application.state.content
+    assert [level.id for level in loaded.levels] == [
+        "01_first_kitchen",
+        "02_double_shift",
+    ]
+    assert "burger" in loaded.recipe_book.recipes
+
+
+# --- broken content refuses startup ---------------------------------------------
+
+
+def copy_shipped_content(tmp_path: Path) -> Path:
+    """A writable copy of the shipped campaign the tests may vandalise."""
+    target = tmp_path / "levels"
+    shutil.copytree(SHIPPED_LEVELS, target)
+    return target
+
+
+def test_missing_recipe_book_refuses_startup_with_a_stderr_report(
+    tmp_path: Path, capsys: pytest.CaptureFixture[str]
+) -> None:
+    capsys.readouterr()  # discard anything earlier tests printed
+    with pytest.raises(ContentError, match=r"recipes\.yaml"):
+        load_validated_content(make_settings(tmp_path))
+    err = capsys.readouterr().err
+    assert "content failed to load" in err
+    assert "recipes.yaml" in err
+
+
+def test_unparsable_level_file_refuses_startup(
+    tmp_path: Path, capsys: pytest.CaptureFixture[str]
+) -> None:
+    levels = copy_shipped_content(tmp_path)
+    (levels / "03_broken.yaml").write_text("\tthis: [is not", encoding="utf-8")
+    capsys.readouterr()
+    with pytest.raises(ContentError, match=r"03_broken\.yaml"):
+        create_app(make_settings(levels))
+    assert "03_broken.yaml" in capsys.readouterr().err
+
+
+def test_sealed_serve_window_refuses_startup_with_named_rule_report(
+    tmp_path: Path, capsys: pytest.CaptureFixture[str]
+) -> None:
+    levels = copy_shipped_content(tmp_path)
+    path = levels / "02_double_shift.yaml"
+    text = path.read_text(encoding="utf-8")
+    vandalism = text.replace('- "####=...=####"', '- "####=..#=####"', 1)
+    vandalism = vandalism.replace('- "#....P.X....#"', '- "#....P#X#...#"')
+    vandalism = vandalism.replace('- "#...........#"', '- "#......#....#"')
+    assert vandalism != text, "the grid rows this test vandalises changed"
+    path.write_text(vandalism, encoding="utf-8")
+
+    capsys.readouterr()
+    with pytest.raises(ContentError, match="SERVE_REACHABLE"):
+        create_app(make_settings(levels))
+    err = capsys.readouterr().err
+    assert "SERVE_REACHABLE" in err
+    assert "02_double_shift" in err
+
+
+# --- warnings do not refuse startup ----------------------------------------------
+
+
+WARN_BAND_CONTENT = """\
+id: probe_level
+name: Probe
+order: 1
+difficulty: 1
+time_limit: 30
+target_score: 40
+blurb: A synthetic level that lands just inside the target-score warning band.
+grid:
+  - "#######"
+  - "#T....#"
+  - "#.....#"
+  - "#....X#"
+  - "#######"
+tiles:
+  "#": {kind: wall}
+  ".": {kind: floor}
+  "T": {kind: floor, station: spawn_shelf, item: tomato}
+  "X": {kind: floor, station: serve}
+spawns: [[1, 1]]
+orders:
+  pool: [probe]
+  concurrent: 1
+  spawn_interval: [10, 14]
+  deadline: 40
+"""
+
+WARN_BAND_RECIPES = """\
+recipes:
+  probe:
+    id: probe
+    name: Probe Dish
+    base_score: 10
+    ingredients:
+      - {kind: tomato, form: raw}
+    steps:
+      - {kind: pick, station: spawn_shelf}
+      - {kind: serve, station: serve}
+"""
+
+
+def write_warn_band_content(tmp_path: Path) -> Path:
+    (tmp_path / RECIPES_FILENAME).write_text(
+        textwrap.dedent(WARN_BAND_RECIPES), encoding="utf-8"
+    )
+    (tmp_path / "probe_level.yaml").write_text(
+        textwrap.dedent(WARN_BAND_CONTENT), encoding="utf-8"
+    )
+    return tmp_path
+
+
+def test_warn_band_target_starts_but_logs_the_warning(
+    tmp_path: Path, caplog: pytest.LogCaptureFixture
+) -> None:
+    levels = write_warn_band_content(tmp_path)
+    with caplog.at_level(logging.WARNING, logger="undercooked.server.routes_content"):
+        bundle = load_validated_content(make_settings(levels))
+    assert [level.id for level in bundle.levels] == ["probe_level"]
+    messages = [record.getMessage() for record in caplog.records]
+    assert any("TARGET_IMPLAUSIBLE" in message for message in messages), messages
+
+
+def test_warn_band_level_still_builds_an_app(tmp_path: Path) -> None:
+    levels = write_warn_band_content(tmp_path)
+    application = create_app(make_settings(levels))
+    assert application.state.content.by_id["probe_level"].target_score == 40

tests/unit/test_docs_index.py

@@ -204,9 +204,14 @@ def test_no_doc_file_exists_that_nothing_promises() -> None:
     )
 
 
-def test_the_docs_on_disk_are_the_three_phase_01_documents() -> None:
+def test_the_docs_on_disk_are_the_shipped_documents() -> None:
     """Pins today's tree so that adding a document is a deliberate, named act."""
-    assert written_docs() == {"docs/DESIGN.md", "docs/DEPLOYMENT.md", "docs/TESTING.md"}
+    assert written_docs() == {
+        "docs/DESIGN.md",
+        "docs/DEPLOYMENT.md",
+        "docs/LEVELS.md",
+        "docs/TESTING.md",
+    }
 
 
 def stub_without_owner(relative: str, text: str) -> str | None:

tests/unit/test_harness.py

@@ -50,7 +50,12 @@ def test_fixture_levels_dir_exists(isolated_levels_dir: None) -> None:
     levels_dir = Path(os.environ["UNDERCOOKED_LEVELS_DIR"])
     assert levels_dir == REPO_ROOT / "tests" / "fixtures" / "levels"
     assert levels_dir.is_dir()
-    assert not list(levels_dir.glob("*.yaml"))
+    # Hermeticity is about *levels*: a default test must never see the shipped
+    # campaign. The committed `recipes.yaml` is not one (the loader never treats
+    # it as a level) — startup has required a recipe book since phase 02 task 04,
+    # and this minimal one is what lets fixture-isolated apps boot at all.
+    levels = [p.name for p in levels_dir.glob("*.yaml") if p.name != "recipes.yaml"]
+    assert not levels
 
 
 def test_e2e_tests_follow_the_story_naming_convention() -> None:

tests/unit/test_level_schema.py

@@ -0,0 +1,440 @@
+"""Unit tests for `undercooked.sim.level` (phase 02 task 02).
+
+The level schema is what makes a YAML file *well-formed* — rectangular grid,
+complete legend both directions, spawns on the map, exactly one serve, ordered
+spawn intervals, placed hazards — and `build_grid`/`Level.build` are the
+deterministic geometry those rules guarantee. Each validator branch gets a
+failing document; the happy path asserts the documented defaults and station
+id conventions from `docs/LEVELS.md`.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+import pytest
+from pydantic import ValidationError
+
+from undercooked.sim.grid import MAX_GRID, Grid
+from undercooked.sim.level import (
+    HazardSpec,
+    Level,
+    LevelDef,
+    OrdersSpec,
+    StationLimits,
+    TileSpec,
+    build_grid,
+    build_station_items,
+    build_stations,
+)
+from undercooked.sim.model import HazardKind, ItemKind, StationKind, TileKind
+
+pytestmark = pytest.mark.unit
+
+TILES: dict[str, Any] = {
+    ".": {"kind": "floor"},
+    "#": {"kind": "wall"},
+    "S": {"kind": "floor", "station": "spawn_shelf", "item": "tomato"},
+    "C": {"kind": "floor", "station": "chop_board"},
+    "V": {"kind": "floor", "station": "serve"},
+}
+
+GRID = ["#####", "#S..#", "#.C.#", "#..V#", "#####"]
+
+
+def level_dict(**overrides: Any) -> dict[str, Any]:
+    """A minimal valid level body; overrides replace top-level fields."""
+    body: dict[str, Any] = {
+        "id": "01_first_kitchen",
+        "name": "First Kitchen",
+        "order": 1,
+        "time_limit": 150,
+        "target_score": 45,
+        "grid": list(GRID),
+        "tiles": {key: dict(value) for key, value in TILES.items()},
+        "spawns": [[1, 1]],
+        "orders": {
+            "pool": ["soup"],
+            "concurrent": 1,
+            "spawn_interval": [16, 20],
+            "deadline": 60,
+        },
+    }
+    body.update(overrides)
+    return body
+
+
+def make_level(**overrides: Any) -> LevelDef:
+    return LevelDef.model_validate(level_dict(**overrides))
+
+
+# ------------------------------------------------------------------------- TileSpec
+
+
+def test_plain_tiles_default_to_no_station_no_item() -> None:
+    spec = TileSpec.model_validate({"kind": "floor"})
+    assert spec.kind is TileKind.FLOOR
+    assert spec.station is None
+    assert spec.item is None
+    assert spec.id_prefix is None
+
+
+def test_spawn_shelf_must_name_the_item_it_spawns() -> None:
+    with pytest.raises(ValidationError, match="must name the item"):
+        TileSpec.model_validate({"kind": "floor", "station": "spawn_shelf"})
+
+
+def test_only_a_spawn_shelf_may_carry_an_item() -> None:
+    with pytest.raises(ValidationError, match="only a spawn_shelf"):
+        TileSpec.model_validate({"kind": "floor", "station": "chop_board", "item": "tomato"})
+
+
+def test_item_without_any_station_is_meaningless() -> None:
+    with pytest.raises(ValidationError, match="only a spawn_shelf"):
+        TileSpec.model_validate({"kind": "floor", "item": "tomato"})
+
+
+def test_id_prefix_belongs_to_a_station_tile() -> None:
+    with pytest.raises(ValidationError, match="id_prefix belongs to a station"):
+        TileSpec.model_validate({"kind": "wall", "id_prefix": "north"})
+
+
+def test_id_prefix_must_not_be_empty() -> None:
+    with pytest.raises(ValidationError):
+        TileSpec.model_validate({"kind": "floor", "station": "stove", "id_prefix": ""})
+
+
+def test_tilespec_rejects_unknown_keys_and_values() -> None:
+    with pytest.raises(ValidationError):
+        TileSpec.model_validate({"kind": "floor", "glow": True})
+    with pytest.raises(ValidationError):
+        TileSpec.model_validate({"kind": "swamp"})
+
+
+# ------------------------------------------------------------------------ OrdersSpec
+
+
+def test_orders_defaults() -> None:
+    orders = OrdersSpec.model_validate(
+        {"pool": ["soup"], "concurrent": 1, "spawn_interval": [7, 11], "deadline": 42}
+    )
+    assert orders.penalty == 10
+    assert orders.spawn_interval == (7.0, 11.0)
+
+
+def test_spawn_interval_min_must_not_exceed_max() -> None:
+    with pytest.raises(ValidationError, match="greater than its end"):
+        OrdersSpec.model_validate(
+            {"pool": ["soup"], "concurrent": 1, "spawn_interval": [12, 8], "deadline": 42}
+        )
+
+
+def test_spawn_interval_must_start_above_zero() -> None:
+    with pytest.raises(ValidationError, match="above 0"):
+        OrdersSpec.model_validate(
+            {"pool": ["soup"], "concurrent": 1, "spawn_interval": [0, 8], "deadline": 42}
+        )
+
+
+def test_orders_pool_needs_recipes_and_sane_numbers() -> None:
+    with pytest.raises(ValidationError, match="pool"):
+        OrdersSpec.model_validate(
+            {"pool": [], "concurrent": 1, "spawn_interval": [7, 11], "deadline": 42}
+        )
+    with pytest.raises(ValidationError, match="concurrent"):
+        OrdersSpec.model_validate(
+            {"pool": ["soup"], "concurrent": 0, "spawn_interval": [7, 11], "deadline": 42}
+        )
+    with pytest.raises(ValidationError, match="deadline"):
+        OrdersSpec.model_validate(
+            {"pool": ["soup"], "concurrent": 1, "spawn_interval": [7, 11], "deadline": 0}
+        )
+    with pytest.raises(ValidationError, match="penalty"):
+        OrdersSpec.model_validate(
+            {"pool": ["soup"], "concurrent": 1, "spawn_interval": [7, 11], "deadline": 42,
+             "penalty": -1}
+        )
+
+
+# ------------------------------------------------------------------------ HazardSpec
+
+
+def hazard_dict(**overrides: Any) -> dict[str, Any]:
+    body: dict[str, Any] = {"type": "spill", "x": 3, "y": 4}
+    body.update(overrides)
+    return body
+
+
+def test_hazard_defaults() -> None:
+    hazard = HazardSpec.model_validate(hazard_dict())
+    assert hazard.type is HazardKind.SPILL
+    assert hazard.chance_per_s is None
+    assert hazard.respawn_s is None
+    assert hazard.path is None
+    assert hazard.intensity == 1.0
+
+
+def test_fire_may_attach_to_a_station() -> None:
+    hazard = HazardSpec.model_validate({"type": "fire", "station": "stove_2"})
+    assert hazard.station == "stove_2"
+
+
+def test_fire_needs_a_station_or_coordinates() -> None:
+    with pytest.raises(ValidationError, match="fire hazard needs a station or x/y"):
+        HazardSpec.model_validate({"type": "fire"})
+
+
+def test_spill_needs_a_location_too() -> None:
+    with pytest.raises(ValidationError, match="spill hazard needs a station or x/y"):
+        HazardSpec.model_validate({"type": "spill"})
+
+
+def test_station_and_coordinates_are_exclusive() -> None:
+    with pytest.raises(ValidationError, match="not both"):
+        HazardSpec.model_validate({"type": "fire", "station": "stove_0", "x": 2, "y": 3})
+
+
+def test_coordinates_come_as_a_pair() -> None:
+    with pytest.raises(ValidationError, match="given together"):
+        HazardSpec.model_validate({"type": "spill", "x": 2})
+    with pytest.raises(ValidationError, match="given together"):
+        HazardSpec.model_validate({"type": "spill", "y": 3})
+
+
+def test_moving_obstacle_needs_a_route_and_a_period() -> None:
+    with pytest.raises(ValidationError, match="path of at least 2 points"):
+        HazardSpec.model_validate({"type": "moving_obstacle", "period_s": 9})
+    with pytest.raises(ValidationError, match="path of at least 2 points"):
+        HazardSpec.model_validate(
+            {"type": "moving_obstacle", "path": [[4, 3]], "period_s": 9}
+        )
+    with pytest.raises(ValidationError, match="needs period_s"):
+        HazardSpec.model_validate({"type": "moving_obstacle", "path": [[4, 3], [10, 3]]})
+
+
+def test_moving_obstacle_happy_path_needs_no_placement() -> None:
+    hazard = HazardSpec.model_validate(
+        {"type": "moving_obstacle", "path": [[4, 3], [10, 3]], "period_s": 9}
+    )
+    assert hazard.path == [(4, 3), (10, 3)]
+    assert hazard.period_s == 9.0
+
+
+def test_rush_is_global_but_needs_positive_intensity() -> None:
+    assert HazardSpec.model_validate({"type": "rush"}).intensity == 1.0
+    with pytest.raises(ValidationError, match="positive intensity"):
+        HazardSpec.model_validate({"type": "rush", "intensity": 0})
+
+
+def test_hazard_numeric_bounds() -> None:
+    with pytest.raises(ValidationError, match="chance_per_s"):
+        HazardSpec.model_validate(hazard_dict(chance_per_s=-0.01))
+    with pytest.raises(ValidationError, match="respawn_s"):
+        HazardSpec.model_validate(hazard_dict(respawn_s=0))
+
+
+# -------------------------------------------------------------------------- LevelDef
+
+
+def test_level_defaults_match_the_documented_values() -> None:
+    level = make_level()
+    assert level.hazards == []
+    assert level.station_limits.plates == 99
+    assert level.station_limits.pots == 99
+    assert level.difficulty == 1
+    assert level.blurb == ""
+    assert level.escalation_notes == ""
+
+
+def test_ragged_grid_is_refused() -> None:
+    with pytest.raises(ValidationError, match="ragged grid: row 4 is 4 chars, want 5"):
+        make_level(grid=["#####", "#S..#", "#.C.#", "#..V#", "####"])
+
+
+def test_empty_rows_are_refused() -> None:
+    with pytest.raises(ValidationError, match="rows must not be empty"):
+        make_level(grid=["", ""], tiles={".": {"kind": "floor"}, "V": {"kind": "floor",
+                "station": "serve"}}, spawns=[])
+
+
+def test_grid_larger_than_the_engine_supports_is_refused() -> None:
+    wide = "#" * (MAX_GRID + 1)
+    with pytest.raises(ValidationError, match=f"at most {MAX_GRID}x{MAX_GRID}"):
+        make_level(grid=[wide, wide])
+
+
+def test_grid_character_without_a_legend_entry_is_refused() -> None:
+    with pytest.raises(ValidationError, match="grid character 'Z' has no tiles"):
+        make_level(grid=["#####", "#Z..#", "#.C.#", "#..V#", "#####"])
+
+
+def test_unused_legend_character_is_refused() -> None:
+    with pytest.raises(ValidationError, match="legend character 'X' is never used"):
+        make_level(tiles={**TILES, "X": {"kind": "wall"}})
+
+
+def test_level_needs_at_least_one_spawn() -> None:
+    with pytest.raises(ValidationError, match="at least one spawn"):
+        make_level(spawns=[])
+
+
+def test_spawns_must_be_inside_the_grid() -> None:
+    with pytest.raises(ValidationError, match=r"spawn \(9,9\) is outside the 5x5 grid"):
+        make_level(spawns=[[9, 9]])
+    with pytest.raises(ValidationError, match=r"spawn \(-1,2\) is outside"):
+        make_level(spawns=[[-1, 2]])
+
+
+def test_level_needs_exactly_one_serve_legend() -> None:
+    no_serve = {key: dict(value) for key, value in TILES.items() if key != "V"}
+    with pytest.raises(ValidationError, match="found 0"):
+        make_level(grid=["#####", "#S..#", "#.C.#", "#...#", "#####"], tiles=no_serve)
+    with pytest.raises(ValidationError, match="found 2"):
+        make_level(
+            grid=["#####", "#S..#", "#.C.#", "#V.U#", "#####"],
+            tiles={**TILES, "U": {"kind": "floor", "station": "serve"}},
+        )
+
+
+def test_level_numbers_are_validated() -> None:
+    with pytest.raises(ValidationError, match="difficulty"):
+        make_level(difficulty=11)
+    with pytest.raises(ValidationError, match="difficulty"):
+        make_level(difficulty=0)
+    with pytest.raises(ValidationError, match="order"):
+        make_level(order=0)
+    with pytest.raises(ValidationError, match="time_limit"):
+        make_level(time_limit=0)
+    with pytest.raises(ValidationError, match="target_score"):
+        make_level(target_score=0)
+
+
+def test_level_rejects_unknown_keys() -> None:
+    with pytest.raises(ValidationError):
+        make_level(weather="rain")
+
+
+def test_level_requires_a_grid_and_a_legend() -> None:
+    with pytest.raises(ValidationError, match="grid"):
+        make_level(grid=[])
+    with pytest.raises(ValidationError, match="tiles"):
+        make_level(tiles={})
+
+
+def test_hazards_and_station_limits_round_trip_on_the_level() -> None:
+    level = make_level(
+        hazards=[
+            {"type": "fire", "station": "stove_2", "chance_per_s": 0.02},
+            {"type": "moving_obstacle", "path": [[4, 3], [10, 3]], "period_s": 9},
+        ],
+        station_limits={"plates": 3, "pots": 2},
+        difficulty=7,
+        blurb="A kitchen.",
+        escalation_notes="tighter",
+    )
+    assert [hazard.type for hazard in level.hazards] == [
+        HazardKind.FIRE,
+        HazardKind.MOVING_OBSTACLE,
+    ]
+    assert level.station_limits == StationLimits(plates=3, pots=2)
+    assert level.difficulty == 7
+
+
+# ------------------------------------------------------------------- building geometry
+
+
+def test_build_grid_maps_the_legend_and_marks_station_cells() -> None:
+    level = make_level()
+    grid = build_grid(level)
+    assert isinstance(grid, Grid)
+    assert (grid.width, grid.height) == (5, 5)
+    assert grid.tile_at(0, 0) is TileKind.WALL
+    assert grid.tile_at(1, 1) is TileKind.FLOOR
+    assert dict(grid.station_at) == {6: "spawn_shelf_0", 12: "chop_board_0", 18: "serve_0"}
+
+
+def test_station_ids_count_per_kind_in_reading_order() -> None:
+    stations = build_stations(make_level())
+    assert [(station.id, station.kind) for station in stations] == [
+        ("spawn_shelf_0", StationKind.SPAWN_SHELF),
+        ("chop_board_0", StationKind.CHOP_BOARD),
+        ("serve_0", StationKind.SERVE),
+    ]
+
+
+def test_two_legend_chars_of_one_station_kind_share_one_counter() -> None:
+    level = make_level(
+        grid=["######", "#T...#", "#..t.#", "#...V#", "######"],
+        tiles={
+            ".": {"kind": "floor"},
+            "#": {"kind": "wall"},
+            "T": {"kind": "floor", "station": "stove"},
+            "t": {"kind": "floor", "station": "stove", "id_prefix": "burner"},
+            "V": {"kind": "floor", "station": "serve"},
+        },
+    )
+    stations = build_stations(level)
+    assert [station.id for station in stations] == ["stove_0", "burner_1", "serve_0"]
+
+
+def test_station_positions_are_float_tile_coordinates() -> None:
+    chop = next(station for station in build_stations(make_level()) if station.kind
+                is StationKind.CHOP_BOARD)
+    assert (chop.x, chop.y) == (2.0, 2.0)
+
+
+def test_a_station_on_a_non_floor_tile_still_builds_and_is_task_03_problem() -> None:
+    level = make_level(
+        grid=["#####", "#S..#", "#.C.#", "#.=V#", "#####"],
+        tiles={**TILES, "=": {"kind": "counter", "station": "stove"}},
+    )
+    grid = build_grid(level)
+    assert grid.tile_at(2, 3) is TileKind.COUNTER
+    assert grid.station_at[17] == "stove_0"
+
+
+def test_station_items_cover_only_spawn_shelves() -> None:
+    level = make_level(
+        grid=["######", "#Ss..#", "#.C..#", "#...V#", "######"],
+        tiles={
+            ".": {"kind": "floor"},
+            "#": {"kind": "wall"},
+            "S": {"kind": "floor", "station": "spawn_shelf", "item": "tomato"},
+            "s": {"kind": "floor", "station": "spawn_shelf", "item": "onion"},
+            "C": {"kind": "floor", "station": "chop_board"},
+            "V": {"kind": "floor", "station": "serve"},
+        },
+        spawns=[[1, 1], [2, 1]],
+    )
+    assert build_station_items(level) == {
+        "spawn_shelf_0": ItemKind.TOMATO,
+        "spawn_shelf_1": ItemKind.ONION,
+    }
+
+
+def test_level_build_pairs_the_spec_with_grid_stations_and_spawns() -> None:
+    spec = make_level()
+    level = Level.build(spec)
+    assert level.spec is spec
+    assert level.grid == build_grid(spec)
+    assert level.stations == build_stations(spec)
+    assert level.spawns == ((1, 1),)
+    assert level.station_items == {"spawn_shelf_0": ItemKind.TOMATO}
+
+
+def test_level_delegates_the_authoring_fields_to_its_spec() -> None:
+    level = Level.build(
+        make_level(difficulty=3, blurb="Cook!", escalation_notes="harder",
+                   station_limits={"plates": 4})
+    )
+    assert level.id == "01_first_kitchen"
+    assert level.name == "First Kitchen"
+    assert level.order == 1
+    assert level.difficulty == 3
+    assert level.time_limit == 150
+    assert level.target_score == 45
+    assert level.orders.pool == ["soup"]
+    assert level.hazards == []
+    assert level.station_limits.plates == 4
+    assert level.blurb == "Cook!"
+    assert level.escalation_notes == "harder"

tests/unit/test_levels_cli.py

@@ -0,0 +1,839 @@
+"""Unit tests for the ``undercooked-levels`` author CLI (phase 02 task 05).
+
+The CLI is the author's half of the content contract, so these tests pin the
+three things a human depends on:
+
+* **the same code path as the server** — `validate` is `load_content` +
+  `validate_bundle`, so a green run here is a green `create_app()`;
+* **the output contract** — one line per level, the named-rule report, a
+  `--json` document a script can parse, the `show --grid` reachability overlay
+  that makes a sealed pocket visible, and the `axes` table phase 11 will feed;
+* **the exit codes** — 0 usable, 1 a rule is broken, 2 nothing usable to work
+  with (load failure, unknown id, refused overwrite, usage error).
+
+Content is addressed explicitly with ``--dir`` (the session-wide
+`UNDERCOOKED_LEVELS_DIR` fixture points at an empty campaign), plus one test
+that deletes the variable and ``chdir``s to prove the documented commands still
+resolve ``levels/`` from the checkout rather than from luck about cwd.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import shutil
+import subprocess
+import sys
+import tomllib
+from collections.abc import Iterator, Sequence
+from dataclasses import replace
+from pathlib import Path
+
+import pytest
+import yaml
+
+from undercooked.config import Settings, load_settings
+from undercooked.server.routes_content import RECIPES_FILENAME
+from undercooked.sim.content import ContentBundle, load_content, load_level_file
+from undercooked.sim.recipe import ContentError, RecipeBook, load_recipe_book
+from undercooked.sim.validation import (
+    AXIS_NAMES,
+    LevelViolation,
+    ValidationReport,
+    escalation_axes,
+)
+from undercooked.tools import levels
+from undercooked.tools.levels import (
+    CHECKOUT_ROOT,
+    EXIT_OK,
+    EXIT_UNUSABLE,
+    EXIT_VIOLATIONS,
+    LIST_HEADER,
+    STARTER_HEIGHT,
+    _assign_chars,
+    _char_for,
+    _CliError,
+    _level_status,
+    describe_level,
+    main,
+    reachability_overlay,
+    render_starter_yaml,
+    resolve_content_paths,
+)
+
+pytestmark = pytest.mark.unit
+
+REPO_ROOT = Path(__file__).resolve().parents[2]
+SHIPPED_LEVELS = REPO_ROOT / "levels"
+SHIPPED_RECIPES = SHIPPED_LEVELS / RECIPES_FILENAME
+BAD_LEVELS = REPO_ROOT / "tests" / "fixtures" / "levels" / "_bad"
+PYPROJECT = REPO_ROOT / "pyproject.toml"
+
+
+def run(argv: Sequence[str]) -> int:
+    """The console script's entry point, in-process (same target as the wrapper)."""
+    return main(list(argv))
+
+
+@pytest.fixture()
+def shipped_bundle() -> ContentBundle:
+    return load_content(SHIPPED_LEVELS, SHIPPED_RECIPES)
+
+
+@pytest.fixture()
+def campaign_copy(tmp_path: Path) -> Iterator[Path]:
+    """A writable copy of the shipped campaign that `new` may extend."""
+    target = tmp_path / "campaign"
+    shutil.copytree(SHIPPED_LEVELS, target)
+    yield target
+
+
+def write_campaign(tmp_path: Path, level_yaml: str, book_yaml: str) -> Path:
+    """A throwaway campaign directory built from two YAML documents."""
+    directory = tmp_path / "mini"
+    directory.mkdir(parents=True, exist_ok=True)
+    (directory / RECIPES_FILENAME).write_text(book_yaml, encoding="utf-8")
+    (directory / "01_mini.yaml").write_text(level_yaml, encoding="utf-8")
+    return directory
+
+
+# --- path resolution -----------------------------------------------------------
+
+
+def test_absolute_flags_win_over_everything() -> None:
+    levels_dir, recipes = resolve_content_paths(
+        Settings(levels_dir=SHIPPED_LEVELS, recipes_path=SHIPPED_RECIPES),
+        Path("/content/levels"),
+        Path("/content/book.yaml"),
+    )
+    assert levels_dir == Path("/content/levels")
+    assert recipes == Path("/content/book.yaml")
+
+
+def test_the_recipe_book_defaults_to_the_one_inside_the_levels_dir() -> None:
+    levels_dir, recipes = resolve_content_paths(Settings(levels_dir=SHIPPED_LEVELS))
+    assert levels_dir == SHIPPED_LEVELS
+    assert recipes == SHIPPED_RECIPES
+
+
+def test_the_recipe_book_flag_beats_the_environment(tmp_path: Path, monkeypatch) -> None:
+    env_book = tmp_path / "env-book.yaml"
+    monkeypatch.setenv("UNDERCOOKED_RECIPES_PATH", str(env_book))
+    _, from_env = resolve_content_paths(load_settings())
+    assert from_env == env_book
+    _, from_flag = resolve_content_paths(load_settings(), None, SHIPPED_RECIPES)
+    assert from_flag == SHIPPED_RECIPES
+
+
+def test_a_relative_path_that_exists_here_is_used_here(tmp_path: Path, monkeypatch) -> None:
+    base = tmp_path.resolve()
+    (base / "levels").mkdir()
+    monkeypatch.chdir(base)
+    levels_dir, recipes = resolve_content_paths(Settings(levels_dir=Path("levels")))
+    assert levels_dir == base / "levels"
+    assert recipes == base / "levels" / RECIPES_FILENAME
+
+
+def test_a_relative_path_with_nothing_local_anchors_to_the_checkout(
+    tmp_path: Path, monkeypatch
+) -> None:
+    """The documented `undercooked-levels validate` works from any directory."""
+    monkeypatch.chdir(tmp_path)
+    levels_dir, recipes = resolve_content_paths(Settings(levels_dir=Path("levels")))
+    assert levels_dir == CHECKOUT_ROOT / "levels"
+    assert recipes == CHECKOUT_ROOT / "levels" / RECIPES_FILENAME
+    assert levels_dir.is_dir()
+
+
+def test_a_path_that_exists_nowhere_is_reported_as_it_was_typed(
+    tmp_path: Path, monkeypatch
+) -> None:
+    monkeypatch.chdir(tmp_path)
+    levels_dir, recipes = resolve_content_paths(
+        Settings(levels_dir=Path("nope/levels"), recipes_path=Path("nope/book.yaml"))
+    )
+    assert levels_dir == Path("nope/levels")
+    assert recipes == Path("nope/book.yaml")
+
+
+def test_a_home_relative_path_is_expanded() -> None:
+    levels_dir, _ = resolve_content_paths(Settings(levels_dir=Path("~/levels")))
+    assert levels_dir == Path.home() / "levels"
+
+
+# --- validate -------------------------------------------------------------------
+
+
+def test_validate_is_green_on_the_shipped_campaign_and_prints_one_line_per_level(
+    capsys: pytest.CaptureFixture[str], shipped_bundle: ContentBundle
+) -> None:
+    assert run(["validate", "--dir", str(SHIPPED_LEVELS)]) == EXIT_OK
+    lines = capsys.readouterr().out.splitlines()
+    assert lines[-1] == "content OK: no violations"
+    assert len(lines) == len(shipped_bundle.levels) + 1
+    for level, line in zip(shipped_bundle.levels, lines[:-1], strict=True):
+        assert line.startswith(level.id)
+        assert "ok" in line
+        assert f"{level.grid.width}x{level.grid.height}" in line
+
+
+def test_validate_exits_1_and_names_every_rule_it_hit(
+    capsys: pytest.CaptureFixture[str]
+) -> None:
+    assert run(["validate", "--dir", str(BAD_LEVELS)]) == EXIT_VIOLATIONS
+    out = capsys.readouterr().out
+    assert "99_broken_room" in out
+    assert "4 error(s)" in out
+    for rule in ("SERVE_REACHABLE", "RECIPE_STATION_MISSING", "TARGET_IMPLAUSIBLE"):
+        assert rule in out
+
+
+def test_validate_json_is_machine_readable_when_green(
+    capsys: pytest.CaptureFixture[str], shipped_bundle: ContentBundle
+) -> None:
+    assert run(["validate", "--dir", str(SHIPPED_LEVELS), "--json"]) == EXIT_OK
+    payload = json.loads(capsys.readouterr().out)
+    assert payload["ok"] is True
+    assert payload["errors"] == []
+    assert payload["warnings"] == []
+    assert payload["counts"] == {"errors": 0, "warnings": 0}
+    assert [level["id"] for level in payload["levels"]] == [
+        level.id for level in shipped_bundle.levels
+    ]
+
+
+def test_validate_json_carries_the_rule_hits_and_their_detail(
+    capsys: pytest.CaptureFixture[str]
+) -> None:
+    assert run(["validate", "--dir", str(BAD_LEVELS), "--json"]) == EXIT_VIOLATIONS
+    payload = json.loads(capsys.readouterr().out)
+    errors = payload["errors"]
+    assert payload["ok"] is False
+    assert payload["counts"] == {"errors": len(errors), "warnings": len(payload["warnings"])}
+    assert {error["rule"] for error in errors} >= {
+        "SERVE_REACHABLE",
+        "RECIPE_STATION_MISSING",
+        "TARGET_IMPLAUSIBLE",
+    }
+    assert {error["level"] for error in errors} == {"99_broken_room"}
+    assert {"recipe": "soup", "station": "stove"} in [error["detail"] for error in errors]
+    keys = [(error["level"], error["rule"], error["message"]) for error in errors]
+    assert keys == sorted(keys)
+
+
+def test_validate_json_is_what_python_m_json_tool_accepts(
+    capsys: pytest.CaptureFixture[str], tmp_path: Path
+) -> None:
+    """The documented pipe: `undercooked-levels validate --json | python -m json.tool`."""
+    assert run(["validate", "--dir", str(BAD_LEVELS), "--json"]) == EXIT_VIOLATIONS
+    raw = capsys.readouterr().out
+    formatted = subprocess.run(
+        [sys.executable, "-m", "json.tool"],
+        input=raw,
+        capture_output=True,
+        text=True,
+        timeout=60,
+        check=False,
+    )
+    assert formatted.returncode == 0, formatted.stderr
+    assert "SERVE_REACHABLE" in formatted.stdout
+    assert json.loads(formatted.stdout)["ok"] is False
+    assert tmp_path  # nothing is written; the report only goes to stdout
+
+
+def test_validate_names_a_missing_directory_and_exits_2(
+    capsys: pytest.CaptureFixture[str], tmp_path: Path
+) -> None:
+    missing = tmp_path / "not-here"
+    assert run(["validate", "--dir", str(missing)]) == EXIT_UNUSABLE
+    err = capsys.readouterr().err
+    assert err.startswith("undercooked-levels:")
+    assert "not-here" in err
+
+
+def test_validate_reports_a_broken_recipe_book_as_exit_2(
+    capsys: pytest.CaptureFixture[str], tmp_path: Path
+) -> None:
+    directory = tmp_path / "levels"
+    directory.mkdir()
+    (directory / RECIPES_FILENAME).write_text("\tthis: [is not", encoding="utf-8")
+    assert run(["validate", "--dir", str(directory)]) == EXIT_UNUSABLE
+    assert "recipes.yaml" in capsys.readouterr().err
+
+
+def test_validate_follows_the_levels_dir_from_the_environment(
+    capsys: pytest.CaptureFixture[str], monkeypatch, campaign_copy: Path
+) -> None:
+    monkeypatch.setenv("UNDERCOOKED_LEVELS_DIR", str(campaign_copy))
+    assert run(["validate"]) == EXIT_OK
+    assert "02_double_shift" in capsys.readouterr().out
+
+
+@pytest.mark.parametrize(
+    "argv",
+    [
+        ["validate"],
+        ["list"],
+        ["axes"],
+        ["show", "01_first_kitchen", "--grid"],
+    ],
+)
+def test_every_subcommand_resolves_levels_from_the_checkout_not_from_cwd(
+    argv: list[str], capsys: pytest.CaptureFixture[str], tmp_path: Path, monkeypatch
+) -> None:
+    """No `--dir`, empty cwd, unset env: the built-in `levels` still finds campaign 02."""
+    monkeypatch.delenv("UNDERCOOKED_LEVELS_DIR", raising=False)
+    monkeypatch.delenv("UNDERCOOKED_RECIPES_PATH", raising=False)
+    monkeypatch.chdir(tmp_path)
+    assert run(argv) == EXIT_OK
+    assert "01_first_kitchen" in capsys.readouterr().out
+
+
+def test_level_status_counts_warnings_of_a_warning_only_level() -> None:
+    bundle = load_content(SHIPPED_LEVELS, SHIPPED_RECIPES)
+    level = bundle.levels[0]
+    warning = LevelViolation(rule="TARGET_IMPLAUSIBLE", level_id=level.id, message="tight")
+    clean = _level_status(level, ValidationReport())
+    noisy = _level_status(level, ValidationReport(violations=(), warnings=(warning,)))
+    error = LevelViolation(rule="SERVE_REACHABLE", level_id=level.id, message="sealed")
+    broken = _level_status(level, ValidationReport(violations=(error,), warnings=(warning,)))
+    assert clean == f"{level.id:<28} {'ok':<24} {level.grid.width}x{level.grid.height}" + (
+        f", {len(level.stations)} stations"
+    )
+    assert "0 error(s)" not in clean
+    assert ", 1 warning(s)" in noisy
+    assert "1 error(s), 1 warning(s)" in broken
+
+
+# --- list -----------------------------------------------------------------------
+
+
+def test_list_prints_one_line_per_level_in_campaign_order(
+    capsys: pytest.CaptureFixture[str], shipped_bundle: ContentBundle
+) -> None:
+    assert run(["list", "--dir", str(SHIPPED_LEVELS)]) == EXIT_OK
+    lines = capsys.readouterr().out.splitlines()
+    assert lines[0] == LIST_HEADER
+    assert len(lines) == 1 + len(shipped_bundle.levels)
+    for level, line in zip(shipped_bundle.levels, lines[1:], strict=True):
+        assert line.startswith(f"{level.order:>5}  {level.id:<28}")
+        assert str(level.target_score) in line
+        assert f"{level.time_limit:>6.1f}" in line
+        for recipe_id in level.orders.pool:
+            assert recipe_id in line
+
+
+def test_list_of_an_empty_campaign_is_only_the_header(
+    capsys: pytest.CaptureFixture[str], tmp_path: Path
+) -> None:
+    empty = tmp_path / "empty"
+    empty.mkdir()
+    shutil.copy(SHIPPED_RECIPES, empty / RECIPES_FILENAME)
+    assert run(["list", "--dir", str(empty)]) == EXIT_OK
+    assert capsys.readouterr().out.splitlines() == [LIST_HEADER]
+
+
+# --- show -----------------------------------------------------------------------
+
+
+def test_show_prints_the_resolved_level(
+    capsys: pytest.CaptureFixture[str], shipped_bundle: ContentBundle
+) -> None:
+    level = shipped_bundle.levels[1]
+    assert run(["show", level.id, "--dir", str(SHIPPED_LEVELS)]) == EXIT_OK
+    out = capsys.readouterr().out
+    assert out.startswith(f"{level.id} - {level.name}")
+    for label in (
+        "order",
+        "difficulty",
+        "time_limit",
+        "target_score",
+        "grid",
+        "spawns",
+        "stations",
+        "orders",
+        "limits",
+        "hazards",
+        "reachability",
+        "recipes",
+        "blurb",
+        "escalation_notes",
+    ):
+        assert f"{label:<18}" in out
+    assert "spawn_shelf x" in out
+    assert "map (as authored)" not in out  # --grid is opt-in
+
+
+def test_show_grid_adds_the_authored_map_and_the_reachability_overlay(
+    capsys: pytest.CaptureFixture[str], shipped_bundle: ContentBundle
+) -> None:
+    level = shipped_bundle.levels[0]
+    assert run(["show", level.id, "--grid", "--dir", str(SHIPPED_LEVELS)]) == EXIT_OK
+    out = capsys.readouterr().out
+    assert "map (as authored)" in out
+    for row in level.spec.grid:
+        assert f"  {row}" in out
+    assert ". reachable" in out
+    assert "?" not in out.split("reachability")[1]  # the shipped kitchen has no dead pockets
+    assert f"{level.grid.width}x{level.grid.height}" in out
+
+
+def test_show_grid_marks_a_sealed_pocket_as_unreachable(
+    capsys: pytest.CaptureFixture[str]
+) -> None:
+    assert run(["show", "99_broken_room", "--grid", "--dir", str(BAD_LEVELS)]) == EXIT_OK
+    out = capsys.readouterr().out
+    overlay = out.split("? walkable but cut off")[1]
+    assert "??" in overlay
+    assert "????#" in overlay  # the walled-in servery row
+    assert "walkable tiles are reachable from the spawns" in overlay
+
+
+def test_show_rejects_an_unknown_level_id(
+    capsys: pytest.CaptureFixture[str], shipped_bundle: ContentBundle
+) -> None:
+    assert run(["show", "07_no_such_kitchen", "--dir", str(SHIPPED_LEVELS)]) == EXIT_UNUSABLE
+    err = capsys.readouterr().err
+    assert "no level '07_no_such_kitchen'" in err
+    assert "01_first_kitchen" in err  # it says what does exist
+
+
+# --- axes -----------------------------------------------------------------------
+
+
+def test_axes_table_lists_every_axis_for_every_level_in_campaign_order(
+    capsys: pytest.CaptureFixture[str], shipped_bundle: ContentBundle
+) -> None:
+    assert run(["axes", "--dir", str(SHIPPED_LEVELS)]) == EXIT_OK
+    lines = capsys.readouterr().out.splitlines()
+    header = lines[0]
+    assert header.startswith("LEVEL")
+    for name in AXIS_NAMES:
+        assert name in header
+    assert set(AXIS_NAMES) == set(escalation_axes(shipped_bundle.levels[0]))
+    assert len(lines) == 1 + len(shipped_bundle.levels)
+    for level, line in zip(shipped_bundle.levels, lines[1:], strict=True):
+        assert line.startswith(f"{level.id:<28}{level.order:>6}")
+        axes = escalation_axes(level)
+        for name in AXIS_NAMES:
+            assert f"{axes[name]:>20.4f}" in line
+
+
+def test_axes_of_an_empty_campaign_is_only_the_header(
+    capsys: pytest.CaptureFixture[str], tmp_path: Path
+) -> None:
+    empty = tmp_path / "empty"
+    empty.mkdir()
+    shutil.copy(SHIPPED_RECIPES, empty / RECIPES_FILENAME)
+    assert run(["axes", "--dir", str(empty)]) == EXIT_OK
+    assert len(capsys.readouterr().out.splitlines()) == 1
+
+
+# --- new ------------------------------------------------------------------------
+
+
+def test_new_writes_a_starter_level_and_validates_it(
+    capsys: pytest.CaptureFixture[str], campaign_copy: Path, shipped_bundle: ContentBundle
+) -> None:
+    assert run(["new", "03_zero_cool", "--dir", str(campaign_copy)]) == EXIT_OK
+    out = capsys.readouterr().out
+    assert f"wrote {campaign_copy / '03_zero_cool.yaml'}" in out
+    assert "content OK: no violations" in out
+
+    written = campaign_copy / "03_zero_cool.yaml"
+    body = yaml.safe_load(written.read_text(encoding="utf-8"))
+    top_difficulty = max(level.difficulty for level in shipped_bundle.levels)
+    top_order = max(level.order for level in shipped_bundle.levels)
+    assert body["id"] == "03_zero_cool"
+    assert body["order"] == top_order + 1
+    assert body["difficulty"] == top_difficulty + 1
+
+    # It is a real level: it loads, and the campaign now carries three of them.
+    level = load_level_file(written)
+    assert level.name == "Zero Cool"
+    assert load_content(campaign_copy, SHIPPED_RECIPES).levels[-1].id == "03_zero_cool"
+    # ...and the scaffolded kitchen is self-consistent: every grid letter used is
+    # declared, and nothing declared is left unused.
+    assert set(body["tiles"]) == {cell for row in body["grid"] for cell in row}
+
+
+def test_new_uses_the_requested_order(
+    capsys: pytest.CaptureFixture[str], campaign_copy: Path
+) -> None:
+    assert run(["new", "07_late_shift", "--order", "7", "--dir", str(campaign_copy)]) == EXIT_OK
+    body = yaml.safe_load((campaign_copy / "07_late_shift.yaml").read_text(encoding="utf-8"))
+    assert body["order"] == 7
+    capsys.readouterr()
+
+
+def test_new_refuses_to_overwrite_an_existing_level(
+    capsys: pytest.CaptureFixture[str], campaign_copy: Path
+) -> None:
+    target = campaign_copy / "01_first_kitchen.yaml"
+    before = target.read_text(encoding="utf-8")
+    assert run(["new", "01_first_kitchen", "--dir", str(campaign_copy)]) == EXIT_UNUSABLE
+    err = capsys.readouterr().err
+    assert "already exists" in err
+    assert "refusing to overwrite" in err
+    assert target.read_text(encoding="utf-8") == before
+
+
+@pytest.mark.parametrize("candidate", ["ZeroCool", "03 zero cool", "_private", "03-second"])
+def test_new_rejects_an_id_that_is_not_a_filename_safe_slug(
+    candidate: str, capsys: pytest.CaptureFixture[str], campaign_copy: Path
+) -> None:
+    assert run(["new", candidate, "--dir", str(campaign_copy)]) == EXIT_UNUSABLE
+    assert "slug" in capsys.readouterr().err
+    assert not (campaign_copy / f"{candidate}.yaml").exists()
+
+
+def test_new_rejects_an_order_below_one(
+    capsys: pytest.CaptureFixture[str], campaign_copy: Path
+) -> None:
+    assert run(["new", "03_probe", "--order", "0", "--dir", str(campaign_copy)]) == EXIT_UNUSABLE
+    assert "--order must be at least 1" in capsys.readouterr().err
+
+
+def test_new_needs_a_readable_campaign(
+    capsys: pytest.CaptureFixture[str], tmp_path: Path
+) -> None:
+    assert run(["new", "03_probe", "--dir", str(tmp_path / "nothing")]) == EXIT_UNUSABLE
+    assert "undercooked-levels:" in capsys.readouterr().err
+
+
+@pytest.mark.parametrize(
+    "error",
+    [OSError("disk full"), OSError(28, "No space left on device")],
+    ids=["no-strerror", "with-strerror"],
+)
+def test_new_reports_a_write_failure(
+    error: OSError,
+    capsys: pytest.CaptureFixture[str],
+    campaign_copy: Path,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    def refuse(self: Path, *args: object, **kwargs: object) -> int:
+        raise error
+
+    monkeypatch.setattr(Path, "write_text", refuse)
+    assert run(["new", "03_probe", "--dir", str(campaign_copy)]) == EXIT_UNUSABLE
+    err = capsys.readouterr().err
+    assert "cannot write" in err
+    assert (error.strerror or str(error)) in err
+
+
+def test_new_reports_a_file_the_loader_rejects(
+    capsys: pytest.CaptureFixture[str], campaign_copy: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    def refuse(path: Path | str) -> object:
+        raise ContentError("the loader hated the scaffold")
+
+    monkeypatch.setattr(levels, "load_level_file", refuse)
+    assert run(["new", "03_probe", "--dir", str(campaign_copy)]) == EXIT_UNUSABLE
+    assert "the loader hated the scaffold" in capsys.readouterr().err
+
+
+# --- the scaffold itself --------------------------------------------------------
+
+
+def test_the_scaffold_is_built_from_the_recipe_book_not_a_template_kitchen() -> None:
+    book = RecipeBook.model_validate(
+        {
+            "recipes": {
+                "broth": {
+                    "id": "broth",
+                    "name": "Plain Broth",
+                    "base_score": 20,
+                    "ingredients": [
+                        {"kind": "rice", "form": "raw"},
+                        {"kind": "fish", "form": "raw"},
+                    ],
+                    "steps": [
+                        {"kind": "pick", "station": "spawn_shelf"},
+                        {"kind": "wash", "station": "sink"},
+                        {"kind": "cook", "station": "stove"},
+                        {"kind": "serve", "station": "serve"},
+                    ],
+                }
+            }
+        }
+    )
+    text = render_starter_yaml(level_id="05_broth_house", order=5, difficulty=4, book=book)
+    body = yaml.safe_load(text)
+
+    assert body["id"] == "05_broth_house"
+    assert body["name"] == "Broth House"
+    assert body["order"] == 5
+    assert body["difficulty"] == 4
+    assert body["orders"]["pool"] == ["broth"]
+
+    stations = {spec["station"] for spec in body["tiles"].values() if "station" in spec}
+    items = {spec["item"] for spec in body["tiles"].values() if "item" in spec}
+    assert stations == {"spawn_shelf", "sink", "stove", "serve"}  # no chop board: no chop step
+    assert items == {"rice", "fish"}
+    assert set(body["tiles"]) == {cell for row in body["grid"] for cell in row}
+    assert len({len(row) for row in body["grid"]}) == 1
+
+
+def test_the_scaffold_of_the_shipped_book_shelves_that_books_first_dish() -> None:
+    book = load_content(SHIPPED_LEVELS, SHIPPED_RECIPES).recipe_book
+    first = book.ids()[0]
+    body = yaml.safe_load(
+        render_starter_yaml(level_id="05_probe", order=5, difficulty=1, book=book)
+    )
+    expected_items = {
+        ingredient.kind.value
+        for ingredient in book.get(first).ingredients
+        if str(ingredient.form) == "raw"
+    }
+    assert {spec["item"] for spec in body["tiles"].values() if "item" in spec} == expected_items
+    assert body["orders"]["pool"] == [first]
+
+
+def test_the_scaffold_needs_something_to_cook() -> None:
+    with pytest.raises(_CliError):
+        render_starter_yaml(level_id="05_probe", order=1, difficulty=1, book=RecipeBook(recipes={}))
+
+
+def test_the_scaffold_keeps_the_spawn_far_enough_from_the_window() -> None:
+    book = load_content(SHIPPED_LEVELS, SHIPPED_RECIPES).recipe_book
+    body = yaml.safe_load(
+        render_starter_yaml(level_id="05_probe", order=1, difficulty=5, book=book)
+    )
+    width = len(body["grid"][0])
+    assert body["grid"][0] == "#" * width == body["grid"][-1]
+    assert len(body["grid"]) == STARTER_HEIGHT
+    fixtures = body["grid"][4]
+    serve_x = fixtures.index(
+        next(char for char, spec in body["tiles"].items() if spec.get("station") == "serve")
+    )
+    assert fixtures[serve_x] != "."  # the window really stands on the fixture row
+    for x, y in body["spawns"]:
+        assert abs(x - serve_x) + abs(y - 4) >= 3  # TRAVEL_TOO_SHORT's floor
+
+
+def test_grid_letters_stay_unique_when_a_preferred_letter_is_taken() -> None:
+    free = {"#", "."}
+    assert _char_for("T", free) == "T"  # an untouched preference wins
+    assert _char_for("T", free | {"T"}) not in free | {"T"}  # taken: a spare instead
+    assert _char_for(None, free) not in free
+
+    used = set(free)
+    assigned = _assign_chars(
+        ["tomato", "onion", "lettuce", "rice"],
+        {"tomato": "T", "onion": "T", "lettuce": "T", "rice": "R"},
+        used,
+    )
+    assert list(assigned) == ["tomato", "onion", "lettuce", "rice"]
+    assert assigned["tomato"] == "T"  # first in line keeps its preferred letter
+    assert assigned["onion"] != "T"  # the collisions fall back to spares
+    assert assigned["rice"] == "R"
+    assert len(set(assigned.values())) == 4
+    assert used == {"#", ".", *assigned.values()}  # the caller's set is the ledger
+
+
+# --- describe_level / overlay directly ------------------------------------------
+
+
+MINI_BOOK = """\
+recipes:
+  snack:
+    id: snack
+    name: Plain Snack
+    base_score: 20
+    ingredients:
+      - {kind: tomato, form: raw}
+    steps:
+      - {kind: pick, station: spawn_shelf}
+      - {kind: serve, station: serve}
+"""
+
+MINI_LEVEL = """\
+id: 01_mini
+name: Mini
+order: 1
+time_limit: 90
+target_score: 40
+grid:
+  - "#######"
+  - "#T....#"
+  - "#.#.#.#"
+  - "#....X#"
+  - "#######"
+tiles:
+  "#": {kind: wall}
+  ".": {kind: floor}
+  "T": {kind: floor, station: spawn_shelf, item: tomato}
+  "X": {kind: floor, station: serve}
+spawns: [[1, 1]]
+orders:
+  pool: [snack]
+  concurrent: 1
+  spawn_interval: [10, 14]
+  deadline: 40
+"""
+
+#: The same kitchen with a cupboard walled in at (3, 3) and (4, 3): walkable
+#: floor no spawn can reach, which is exactly what the overlay exists to show.
+POCKET_LEVEL = MINI_LEVEL.replace(
+    """\
+id: 01_mini
+name: Mini""",
+    """\
+id: 02_pocket
+name: Pocket""",
+).replace(
+    """\
+  - "#######"
+  - "#T....#"
+  - "#.#.#.#"
+  - "#....X#"
+  - "#######""",
+    """\
+  - "########"
+  - "#T....X#"
+  - "#.####.#"
+  - "#.#BB#.#"
+  - "#.####.#"
+  - "########""",
+).replace(
+    '  "X": {kind: floor, station: serve}',
+    '  "X": {kind: floor, station: serve}\n  "B": {kind: floor}',
+)
+
+
+def test_describe_marks_missing_prose_as_none(tmp_path: Path) -> None:
+    directory = write_campaign(tmp_path, MINI_LEVEL, MINI_BOOK)
+    level = load_level_file(directory / "01_mini.yaml")
+    out = "\n".join(describe_level(level, load_recipe_book(directory / RECIPES_FILENAME)))
+    assert "snack: Plain Snack" in out  # the prose is missing, the recipe is not
+    assert out.count("(none)") == 2  # blurb and escalation_notes say so out loud
+
+
+def test_describe_reports_a_pool_recipe_the_book_does_not_define(tmp_path: Path) -> None:
+    directory = write_campaign(tmp_path, MINI_LEVEL, MINI_BOOK)
+    level = load_level_file(directory / "01_mini.yaml")
+    out = "\n".join(describe_level(level, RecipeBook(recipes={})))
+    assert "snack: MISSING from the recipe book" in out
+
+
+def test_describe_counts_hazards(tmp_path: Path) -> None:
+    level_yaml = MINI_LEVEL.replace("hazards: []", "") + (
+        "hazards:\n  - {type: spill, x: 4, y: 1}\n  - {type: fire, x: 2, y: 1}\n"
+    )
+    directory = tmp_path / "haz"
+    directory.mkdir()
+    (directory / RECIPES_FILENAME).write_text(MINI_BOOK, encoding="utf-8")
+    (directory / "01_mini.yaml").write_text(level_yaml, encoding="utf-8")
+    level = load_level_file(directory / "01_mini.yaml")
+    lines = describe_level(level, load_recipe_book(directory / RECIPES_FILENAME))
+    hazards = next(line for line in lines if line.startswith("hazards"))
+    assert "fire x1" in hazards
+    assert "spill x1" in hazards
+
+
+def test_describe_says_none_when_a_level_has_no_stations(tmp_path: Path) -> None:
+    """A hand-built level with no stations (content forbids one: no serve window).
+
+    `LevelDef` requires exactly one serve tile, so the station-less `Level` can
+    only come from a hand-built dataclass — which is what a phase-03 caller
+    holding a partially-built level does.
+    """
+    directory = write_campaign(tmp_path, MINI_LEVEL, MINI_BOOK)
+    level = load_level_file(directory / "01_mini.yaml")
+    assert level.stations  # the authored level has some; the stripped one has none
+    lines = describe_level(replace(level, stations=()), RecipeBook(recipes={}))
+    empty = "stations".ljust(18) + "none"
+    assert next(line for line in lines if line.startswith("stations")) == empty
+
+
+def test_reachability_overlay_only_uses_the_three_terrain_characters(tmp_path: Path) -> None:
+    directory = write_campaign(tmp_path, POCKET_LEVEL, MINI_BOOK)
+    level = load_level_file(directory / "01_mini.yaml")
+    overlay = reachability_overlay(level)
+    assert {cell for row in overlay for cell in row} == {".", "?", "#"}
+    # the walled-in pocket reads as `??`; everything else the cooks can walk to.
+    assert overlay[3] == "#.#??#.#"
+    assert sum(row.count("?") for row in overlay) == 2
+
+
+# --- entry point ----------------------------------------------------------------
+
+
+def test_no_subcommand_is_a_usage_error(capsys: pytest.CaptureFixture[str]) -> None:
+    assert run([]) == EXIT_UNUSABLE
+    err = capsys.readouterr().err
+    assert "usage: undercooked-levels" in err
+    assert "arguments are required: COMMAND" in err
+
+
+def test_an_unknown_subcommand_is_a_usage_error(capsys: pytest.CaptureFixture[str]) -> None:
+    assert run(["scullery"]) == EXIT_UNUSABLE
+    err = capsys.readouterr().err
+    assert "usage: undercooked-levels" in err
+    assert "invalid choice: 'scullery'" in err
+
+
+def test_an_unknown_flag_is_a_usage_error(capsys: pytest.CaptureFixture[str]) -> None:
+    assert run(["validate", "--nope"]) == EXIT_UNUSABLE
+    assert "unrecognized arguments" in capsys.readouterr().err
+
+
+@pytest.mark.parametrize("argv", [["--help"], ["validate", "--help"], ["new", "-h"]])
+def test_help_exits_cleanly(argv: list[str], capsys: pytest.CaptureFixture[str]) -> None:
+    assert run(argv) == EXIT_OK
+    out = capsys.readouterr().out
+    assert "usage: undercooked-levels" in out
+
+
+def test_help_documents_the_subcommands_and_exit_codes(
+    capsys: pytest.CaptureFixture[str],
+) -> None:
+    assert run(["--help"]) == EXIT_OK
+    out = capsys.readouterr().out
+    for command in ("validate", "list", "show", "new", "axes"):
+        assert command in out
+    assert "0 ok, 1 a level breaks a validation rule" in out
+
+
+def test_the_parser_builds_with_a_handler_for_every_subcommand() -> None:
+    parser = levels.build_parser()
+    for command, expected in (
+        (["validate"], "validate"),
+        (["list"], "list"),
+        (["show", "x"], "show"),
+        (["new", "x"], "new"),
+        (["axes"], "axes"),
+    ):
+        args = parser.parse_args(command)
+        assert args.command == expected
+        assert callable(args.handler)
+        assert args.dir is None or isinstance(args.dir, Path)
+
+
+def test_console_script_is_declared_and_runs_as_a_module() -> None:
+    """The real entry point: `[project.scripts]` -> this module's `main`."""
+    declared = tomllib.loads(PYPROJECT.read_text(encoding="utf-8"))["project"]["scripts"]
+    assert declared["undercooked-levels"] == "undercooked.tools.levels:main"
+
+    result = subprocess.run(  # fixed argv, the repo's own interpreter
+        [
+            sys.executable,
+            "-m",
+            "undercooked.tools.levels",
+            "list",
+        ],
+        capture_output=True,
+        text=True,
+        timeout=120,
+        check=False,
+        cwd=REPO_ROOT,
+        env={**os.environ, "UNDERCOOKED_LEVELS_DIR": str(SHIPPED_LEVELS)},
+    )
+    assert result.returncode == EXIT_OK, result.stderr
+    assert "01_first_kitchen" in result.stdout
+    assert "02_double_shift" in result.stdout
+
+
+def test_exit_codes_are_the_documented_ones() -> None:
+    assert (EXIT_OK, EXIT_VIOLATIONS, EXIT_UNUSABLE) == (0, 1, 2)

tests/unit/test_levels_doc.py

@@ -0,0 +1,104 @@
+"""docs/LEVELS.md must document exactly the fields the models have.
+
+The doc is the content author's contract; the Pydantic models are its
+executable form. This test parses every ``### <Model>`` section's *Field*
+table (the first table whose header row starts with ``Field``, so the
+HazardSpec placement table is not mistaken for a field list) and requires the
+back-quoted field names to equal ``model_fields`` — no undocumented field,
+no documented ghost. A planted-drift case proves the extraction can fail.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import pytest
+from pydantic import BaseModel
+
+from undercooked.sim.level import (
+    HazardSpec,
+    LevelDef,
+    OrdersSpec,
+    StationLimits,
+    TileSpec,
+)
+from undercooked.sim.recipe import Ingredient, Recipe, RecipeBook, Step
+
+pytestmark = pytest.mark.unit
+
+DOC_PATH = Path(__file__).resolve().parents[2] / "docs" / "LEVELS.md"
+
+MODELS: tuple[type[BaseModel], ...] = (
+    Ingredient,
+    Step,
+    Recipe,
+    RecipeBook,
+    TileSpec,
+    OrdersSpec,
+    HazardSpec,
+    StationLimits,
+    LevelDef,
+)
+
+
+def _section(lines: list[str], model_name: str) -> list[str]:
+    """Lines of the `### model_name` section, up to the next heading."""
+    heading = f"### {model_name}"
+    collected: list[str] = []
+    in_section = False
+    for line in lines:
+        if line.startswith("#"):
+            if in_section:
+                break
+            in_section = line.strip() == heading
+            continue
+        if in_section:
+            collected.append(line)
+    return collected
+
+
+def _table_fields(section: list[str]) -> set[str]:
+    """Back-quoted first-column entries of the section's `Field` table."""
+    start = None
+    for index, line in enumerate(section):
+        stripped = line.strip()
+        if stripped.startswith("|") and stripped.strip("|").split("|")[0].strip() == "Field":
+            start = index
+            break
+    if start is None:
+        return set()
+    fields: set[str] = set()
+    for line in section[start + 2 :]:  # skip header + |---| separator
+        stripped = line.strip()
+        if not stripped.startswith("|"):
+            break
+        first = stripped.strip("|").split("|")[0].strip()
+        if len(first) > 2 and first.startswith("`") and first.endswith("`"):
+            fields.add(first[1:-1])
+    return fields
+
+
+@pytest.fixture(scope="module")
+def doc_lines() -> list[str]:
+    assert DOC_PATH.is_file(), f"{DOC_PATH} is missing"
+    return DOC_PATH.read_text(encoding="utf-8").splitlines()
+
+
+@pytest.mark.parametrize("model", MODELS, ids=lambda model: model.__name__)
+def test_every_model_has_an_exact_field_table(doc_lines: list[str], model: type[BaseModel]) -> None:
+    documented = _table_fields(_section(doc_lines, model.__name__))
+    assert documented, f"docs/LEVELS.md has no Field table for {model.__name__}"
+    assert documented == set(model.model_fields), (
+        f"docs/LEVELS.md vs {model.__name__}: "
+        f"undocumented={sorted(set(model.model_fields) - documented)}, "
+        f"not-in-model={sorted(documented - set(model.model_fields))}"
+    )
+
+
+def test_the_comparison_catches_planted_drift(doc_lines: list[str]) -> None:
+    planted = [
+        line.replace("| `penalty` |", "| `reward` |") for line in doc_lines
+    ]
+    documented = _table_fields(_section(planted, "OrdersSpec"))
+    assert "reward" in documented
+    assert documented != set(OrdersSpec.model_fields)

tests/unit/test_main.py

@@ -37,6 +37,7 @@ from undercooked.config import (
     get_settings,
     reset_settings_cache,
 )
+from undercooked.sim.recipe import ContentError
 
 if TYPE_CHECKING:
     from collections.abc import Iterator
@@ -214,6 +215,39 @@ def test_main_reports_an_invalid_environment(
     assert "UNDERCOOKED_TICK_HZ=99" in capsys.readouterr().err
 
 
+def test_main_exits_two_when_the_content_gate_refuses_startup(
+    monkeypatch: pytest.MonkeyPatch, capsys: pytest.CaptureFixture[str]
+) -> None:
+    """Unusable levels are a startup error, not a traceback.
+
+    Content is read when uvicorn imports the app string, so `create_app`'s
+    `ContentError` surfaces out of `uvicorn.run` — and the CLI answers with exit 2,
+    the same code an unusable `UNDERCOOKED_PORT` gets. The named-rule report itself
+    is written by `create_app` before it raises, so this only adds the closing line.
+    """
+
+    def refusing_run(_target: str, **_kwargs: Any) -> None:
+        raise ContentError("undercooked: level content failed validation:\nBOGUS_RULE: nope")
+
+    monkeypatch.setattr(uvicorn, "run", refusing_run)
+    assert main([]) == EXIT_CONFIG_ERROR
+    err = capsys.readouterr().err
+    assert "refusing to start" in err
+    assert "undercooked-levels validate" in err
+    assert "Traceback" not in err
+
+
+def test_main_lets_any_other_serving_failure_propagate(monkeypatch: pytest.MonkeyPatch) -> None:
+    """Only the content gate is translated: a real crash stays a real crash."""
+
+    def crashing_run(_target: str, **_kwargs: Any) -> None:
+        raise OSError("port in use")
+
+    monkeypatch.setattr(uvicorn, "run", crashing_run)
+    with pytest.raises(OSError, match="port in use"):
+        main([])
+
+
 def test_unknown_flag_is_an_argument_error() -> None:
     with pytest.raises(SystemExit) as excinfo:
         main(["--nope"])

tests/unit/test_recipe_schema.py

@@ -0,0 +1,311 @@
+"""Unit tests for `undercooked.sim.recipe` (phase 02 task 02).
+
+The recipe book is the content contract every later phase cooks against: strict
+schemas (typos are load errors, not defaults), deterministic helpers, and a
+`min_seconds_to_complete` solver bound the validator trusts. Every failing case
+here is a YAML an author could plausibly write.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Any
+
+import pytest
+from pydantic import ValidationError
+
+from undercooked.sim.model import ItemForm, ItemKind, StationKind
+from undercooked.sim.recipe import (
+    HANDLING_S,
+    ContentError,
+    Ingredient,
+    Recipe,
+    RecipeBook,
+    Step,
+    describe_validation_error,
+    load_recipe_book,
+    read_yaml_mapping,
+)
+
+pytestmark = pytest.mark.unit
+
+
+def soup_dict(**overrides: Any) -> dict[str, Any]:
+    """A minimal valid soup recipe body; overrides replace top-level fields."""
+    body: dict[str, Any] = {
+        "id": "soup",
+        "name": "Tomato Soup",
+        "base_score": 40,
+        "ingredients": [
+            {"kind": "tomato", "form": "raw"},
+            {"kind": "onion", "form": "raw", "count": 2},
+        ],
+        "steps": [
+            {"kind": "chop", "station": "chop_board", "seconds": 0.9},
+            {"kind": "cook", "station": "stove", "seconds": 5.0},
+        ],
+    }
+    body.update(overrides)
+    return body
+
+
+def book_dict(**overrides: Any) -> dict[str, Any]:
+    data: dict[str, Any] = {"recipes": {"soup": soup_dict()}}
+    data.update(overrides)
+    return data
+
+
+def write_book(directory: Path, text: str) -> Path:
+    path = directory / "recipes.yaml"
+    path.write_text(text, encoding="utf-8")
+    return path
+
+
+# --------------------------------------------------------------------- Ingredient / Step
+
+
+def test_handling_constant_is_the_documented_bound() -> None:
+    assert HANDLING_S == 0.6
+
+
+def test_ingredient_defaults_count_to_one() -> None:
+    ingredient = Ingredient(kind=ItemKind.TOMATO, form=ItemForm.RAW)
+    assert ingredient.count == 1
+
+
+def test_ingredient_count_must_be_positive() -> None:
+    with pytest.raises(ValidationError, match="count"):
+        Ingredient(kind=ItemKind.TOMATO, form=ItemForm.RAW, count=0)
+
+
+def test_ingredient_kinds_are_the_enum_not_arbitrary_strings() -> None:
+    with pytest.raises(ValidationError):
+        Ingredient.model_validate({"kind": "durian", "form": "raw"})
+
+
+def test_step_defaults_to_zero_seconds() -> None:
+    step = Step(kind="pick", station=StationKind.SPAWN_SHELF)
+    assert step.seconds == 0.0
+
+
+def test_step_kind_is_a_closed_vocabulary() -> None:
+    with pytest.raises(ValidationError, match="kind"):
+        Step.model_validate({"kind": "throw", "station": "chop_board"})
+
+
+def test_step_seconds_cannot_be_negative() -> None:
+    with pytest.raises(ValidationError, match="seconds"):
+        Step(kind="chop", station=StationKind.CHOP_BOARD, seconds=-0.1)
+
+
+def test_step_station_must_be_a_known_station_kind() -> None:
+    with pytest.raises(ValidationError, match="station"):
+        Step.model_validate({"kind": "cook", "station": "deep_fryer"})
+
+
+# --------------------------------------------------------------------------- Recipe
+
+
+def test_recipe_happy_path_carries_documented_defaults() -> None:
+    recipe = Recipe.model_validate(soup_dict())
+    assert recipe.requires_plating is True
+    assert recipe.tags == []
+    assert [ingredient.count for ingredient in recipe.ingredients] == [1, 2]
+
+
+def test_recipe_rejects_unknown_keys() -> None:
+    with pytest.raises(ValidationError, match=r"extra_forbidden|Extra inputs"):
+        Recipe.model_validate(soup_dict(garnish="parsley"))
+
+
+def test_recipe_requires_ingredients_and_steps() -> None:
+    with pytest.raises(ValidationError, match="ingredients"):
+        Recipe.model_validate(soup_dict(ingredients=[]))
+    with pytest.raises(ValidationError, match="steps"):
+        Recipe.model_validate(soup_dict(steps=[]))
+
+
+def test_recipe_score_and_ids_are_sane() -> None:
+    with pytest.raises(ValidationError):
+        Recipe.model_validate(soup_dict(base_score=-1))
+    with pytest.raises(ValidationError):
+        Recipe.model_validate(soup_dict(id=""))
+
+
+# ------------------------------------------------------------------------ RecipeBook
+
+
+def test_book_reads_keys_ids_and_the_ingredient_catalog() -> None:
+    book = RecipeBook.model_validate(
+        book_dict(ingredients=["tomato", "onion", "plate"])
+    )
+    assert book.ids() == ("soup",)
+    assert book.get("soup").base_score == 40
+    assert book.ingredients == [ItemKind.TOMATO, ItemKind.ONION, ItemKind.PLATE]
+
+
+def test_book_ids_follow_declaration_order() -> None:
+    book = RecipeBook.model_validate(
+        book_dict(recipes={"salad": soup_dict(id="salad"), "soup": soup_dict()})
+    )
+    assert book.ids() == ("salad", "soup")
+
+
+def test_book_key_must_equal_the_recipe_id() -> None:
+    with pytest.raises(ValidationError, match="keys must equal"):
+        RecipeBook.model_validate(book_dict(recipes={"soup": soup_dict(id="chowder")}))
+
+
+def test_get_unknown_recipe_names_the_known_ones() -> None:
+    book = RecipeBook.model_validate(book_dict())
+    with pytest.raises(KeyError, match="unknown recipe 'ghost'"):
+        book.get("ghost")
+
+
+def test_required_station_kinds_for_a_two_step_recipe() -> None:
+    book = RecipeBook.model_validate(book_dict())
+    assert book.required_station_kinds("soup") == frozenset(
+        {StationKind.CHOP_BOARD, StationKind.STOVE}
+    )
+
+
+def four_step_book() -> RecipeBook:
+    recipe = soup_dict(
+        steps=[
+            {"kind": "chop", "station": "chop_board", "seconds": 0.9},
+            {"kind": "cook", "station": "stove", "seconds": 4.0},
+            {"kind": "plate", "station": "prep", "seconds": 1.2},
+            {"kind": "serve", "station": "serve"},
+        ]
+    )
+    return RecipeBook.model_validate(book_dict(recipes={"soup": recipe}))
+
+
+def test_required_station_kinds_for_a_four_step_recipe() -> None:
+    assert four_step_book().required_station_kinds("soup") == frozenset(
+        {StationKind.CHOP_BOARD, StationKind.STOVE, StationKind.PREP, StationKind.SERVE}
+    )
+
+
+def test_required_item_kinds_adds_a_plate_only_when_plating_is_required() -> None:
+    book = RecipeBook.model_validate(book_dict())
+    assert book.required_item_kinds("soup") == frozenset(
+        {ItemKind.TOMATO, ItemKind.ONION, ItemKind.PLATE}
+    )
+    unplated = RecipeBook.model_validate(
+        book_dict(recipes={"soup": soup_dict(requires_plating=False)})
+    )
+    assert unplated.required_item_kinds("soup") == frozenset({ItemKind.TOMATO, ItemKind.ONION})
+
+
+def test_min_seconds_to_complete_is_steps_plus_one_handling_allowance() -> None:
+    book = RecipeBook.model_validate(book_dict())
+    assert book.min_seconds_to_complete("soup") == pytest.approx(0.9 + 5.0 + HANDLING_S)
+
+
+def test_min_seconds_grows_with_more_steps() -> None:
+    minutes = four_step_book().min_seconds_to_complete("soup")
+    assert minutes == pytest.approx(0.9 + 4.0 + 1.2 + 0.0 + HANDLING_S)
+
+
+@pytest.mark.parametrize(
+    "method",
+    ["required_station_kinds", "required_item_kinds", "min_seconds_to_complete"],
+)
+def test_solver_helpers_reject_unknown_recipes(method: str) -> None:
+    book = RecipeBook.model_validate(book_dict())
+    with pytest.raises(KeyError, match="unknown recipe"):
+        getattr(book, method)("ghost")
+
+
+# ------------------------------------------------------------------- load_recipe_book
+
+
+def test_load_recipe_book_reads_a_real_yaml_file(tmp_path: Path) -> None:
+    path = write_book(
+        tmp_path,
+        """
+        ingredients: [tomato, onion, plate]
+        recipes:
+          soup:
+            id: soup
+            name: Tomato Soup
+            base_score: 40
+            ingredients: [{kind: tomato, form: raw}, {kind: onion, form: raw, count: 2}]
+            steps:
+              - {kind: chop, station: chop_board, seconds: 0.9}
+              - {kind: cook, station: stove, seconds: 5}
+        """,
+    )
+    book = load_recipe_book(path)
+    assert book.ids() == ("soup",)
+    assert book.ingredients == [ItemKind.TOMATO, ItemKind.ONION, ItemKind.PLATE]
+
+
+def test_load_recipe_book_reports_every_field_problem(tmp_path: Path) -> None:
+    path = write_book(
+        tmp_path,
+        """
+        recipes:
+          soup:
+            id: soup
+            name: Soup
+            base_score: not-a-number
+            ingredients: [{kind: durian, form: raw}]
+            steps: [{kind: throw, station: chop_board}]
+        """,
+    )
+    with pytest.raises(ContentError) as excinfo:
+        load_recipe_book(path)
+    message = str(excinfo.value)
+    assert "base_score" in message
+    assert "ingredients.0.kind" in message
+    assert "steps.0.kind" in message
+
+
+def test_load_recipe_book_rejects_unknown_top_level_keys(tmp_path: Path) -> None:
+    path = write_book(tmp_path, "recipes: {}\ndesserts: {}\n")
+    with pytest.raises(ContentError, match="desserts"):
+        load_recipe_book(path)
+
+
+def test_load_recipe_book_missing_file_is_a_content_error(tmp_path: Path) -> None:
+    with pytest.raises(ContentError, match="cannot read content file"):
+        load_recipe_book(tmp_path / "nope.yaml")
+
+
+def test_load_recipe_book_invalid_yaml_is_a_content_error(tmp_path: Path) -> None:
+    path = write_book(tmp_path, "recipes: [unclosed\n")
+    with pytest.raises(ContentError, match="invalid YAML"):
+        load_recipe_book(path)
+
+
+def test_load_recipe_book_rejects_a_list_document(tmp_path: Path) -> None:
+    path = write_book(tmp_path, "- just\n- a\n- list\n")
+    with pytest.raises(ContentError, match="expected a YAML mapping"):
+        load_recipe_book(path)
+
+
+def test_empty_file_reads_as_an_empty_mapping_then_fails_the_schema(tmp_path: Path) -> None:
+    path = write_book(tmp_path, "")
+    assert read_yaml_mapping(path) == {}
+    with pytest.raises(ContentError, match="recipes: Field required"):
+        load_recipe_book(path)
+
+
+# ----------------------------------------------------------------- strictness plumbing
+
+
+def test_assignment_revalidates_so_late_edits_cannot_corrupt_a_model() -> None:
+    ingredient = Ingredient(kind=ItemKind.TOMATO, form=ItemForm.RAW)
+    ingredient.count = 3
+    assert ingredient.count == 3
+    with pytest.raises(ValidationError):
+        ingredient.count = 0
+
+
+def test_describe_validation_error_flattens_loc_paths(tmp_path: Path) -> None:
+    with pytest.raises(ValidationError) as excinfo:
+        RecipeBook.model_validate({})
+    described = describe_validation_error(excinfo.value, "book.yaml")
+    assert described.startswith("book.yaml: recipes: Field required")

tests/unit/test_shipped_content.py

@@ -0,0 +1,215 @@
+"""Regression tests for the shipped campaign content (``levels/``, task 04).
+
+These pin what campaign 02 ships, exactly as the phase task specifies: the
+recipe book (ingredients + four named dishes with PLAN §5 step pipelines and
+fixed base scores), the first two levels' tunables (pool, concurrency, spawn
+interval, deadline, time limit, target), the blurb floor, campaign ordering,
+and the level-1 → level-2 escalation contract (≥2 axes up). They read the
+real files from ``levels/`` — not a copy — so editing shipped content that
+a test pins fails here, in review, instead of at a player's startup.
+
+Startup behaviour and the HTTP surface live in ``test_content_startup.py``
+and ``tests/integration/test_content_api.py``; the generic loader and
+validator have their own unit suites.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import pytest
+
+from undercooked.sim.content import ContentBundle, load_content
+from undercooked.sim.level import Level
+from undercooked.sim.recipe import RecipeBook, load_recipe_book
+from undercooked.sim.validation import escalation_axes, validate_bundle
+
+REPO_ROOT = Path(__file__).resolve().parents[2]
+LEVELS_DIR = REPO_ROOT / "levels"
+RECIPES_PATH = LEVELS_DIR / "recipes.yaml"
+LEVEL_FILES = sorted(
+    path for path in LEVELS_DIR.glob("*.yaml") if path.name != "recipes.yaml"
+)
+
+#: The campaign order this task ships (PLAN §8: soup tutorial, then salad).
+SHIPPED_IDS = ["01_first_kitchen", "02_double_shift"]
+
+#: PLAN §5 station pipelines as (interaction, station) pairs:
+#: pick at the shelf -> chop at the board -> cook on the stove -> plate at
+#: prep -> serve at the window; a dish skips the stages its recipe lacks.
+EXPECTED_STEPS = {
+    "soup": [
+        ("pick", "spawn_shelf"),
+        ("chop", "chop_board"),
+        ("cook", "stove"),
+        ("plate", "prep"),
+        ("serve", "serve"),
+    ],
+    "salad": [
+        ("pick", "spawn_shelf"),
+        ("chop", "chop_board"),
+        ("plate", "prep"),
+        ("serve", "serve"),
+    ],
+    "fried_rice": [
+        ("pick", "spawn_shelf"),
+        ("cook", "stove"),
+        ("plate", "prep"),
+        ("serve", "serve"),
+    ],
+    "burger": [
+        ("pick", "spawn_shelf"),
+        ("chop", "chop_board"),
+        ("cook", "stove"),
+        ("plate", "prep"),
+        ("serve", "serve"),
+    ],
+}
+
+
+@pytest.fixture(scope="module")
+def book() -> RecipeBook:
+    return load_recipe_book(RECIPES_PATH)
+
+
+@pytest.fixture(scope="module")
+def bundle() -> ContentBundle:
+    return load_content(LEVELS_DIR, RECIPES_PATH)
+
+
+def level_of(bundle: ContentBundle, level_id: str) -> Level:
+    return bundle.by_id[level_id]
+
+
+# --- the recipe book ---------------------------------------------------------
+
+
+def test_recipe_book_ships_the_named_ingredient_catalog(book: RecipeBook) -> None:
+    assert set(book.ingredients) == {
+        "tomato",
+        "onion",
+        "lettuce",
+        "dough",
+        "patty",
+        "rice",
+        "fish",
+        "plate",
+    }
+
+
+def test_recipe_book_ships_exactly_the_four_dishes(book: RecipeBook) -> None:
+    assert set(book.recipes) == {"soup", "salad", "fried_rice", "burger"}
+
+
+@pytest.mark.parametrize(
+    ("recipe_id", "base_score"),
+    [("soup", 40), ("salad", 30), ("fried_rice", 50), ("burger", 70)],
+)
+def test_dishes_carry_their_pinned_base_scores(
+    book: RecipeBook, recipe_id: str, base_score: int
+) -> None:
+    assert book.get(recipe_id).base_score == base_score
+
+
+@pytest.mark.parametrize("recipe_id", sorted(EXPECTED_STEPS))
+def test_dishes_follow_the_plan_station_pipelines(
+    book: RecipeBook, recipe_id: str
+) -> None:
+    recipe = book.get(recipe_id)
+    assert [(step.kind, step.station.value) for step in recipe.steps] == (
+        EXPECTED_STEPS[recipe_id]
+    )
+
+
+@pytest.mark.parametrize("recipe_id", sorted(EXPECTED_STEPS))
+def test_every_dish_consumes_catalogued_ingredients(
+    book: RecipeBook, recipe_id: str
+) -> None:
+    recipe = book.get(recipe_id)
+    assert recipe.ingredients, recipe_id
+    assert {item.kind for item in recipe.ingredients} <= set(book.ingredients)
+
+
+# --- the level files ----------------------------------------------------------
+
+
+def test_every_yaml_file_in_levels_is_a_shipped_level(bundle: ContentBundle) -> None:
+    stems = {path.stem for path in LEVEL_FILES}
+    assert stems, "levels/ must ship at least one level"
+    assert stems == set(bundle.by_id)
+
+
+def test_shipped_campaign_validates_without_warnings(bundle: ContentBundle) -> None:
+    report = validate_bundle(bundle)
+    assert report.ok, report.render()
+    assert not report.warnings, report.render()
+
+
+def test_campaign_order_is_unique_and_starts_at_one(bundle: ContentBundle) -> None:
+    orders = [level.order for level in bundle.levels]
+    assert orders == list(range(1, len(bundle.levels) + 1))
+
+
+def test_campaign_ships_the_first_two_levels(bundle: ContentBundle) -> None:
+    assert [level.id for level in bundle.levels] == SHIPPED_IDS
+
+
+@pytest.mark.parametrize("level_id", SHIPPED_IDS)
+def test_each_level_has_a_readable_premise(bundle: ContentBundle, level_id: str) -> None:
+    blurb = level_of(bundle, level_id).blurb
+    assert len(blurb) >= 20
+
+
+@pytest.mark.parametrize("level_id", SHIPPED_IDS)
+def test_difficulty_is_within_the_declared_range(
+    bundle: ContentBundle, level_id: str
+) -> None:
+    assert 1 <= level_of(bundle, level_id).difficulty <= 10
+
+
+@pytest.mark.parametrize("level_id", SHIPPED_IDS)
+def test_pool_recipes_exist_in_the_book(
+    bundle: ContentBundle, book: RecipeBook, level_id: str
+) -> None:
+    recipe_ids = level_of(bundle, level_id).orders.pool
+    assert recipe_ids
+    for recipe_id in recipe_ids:
+        assert recipe_id in book.recipes
+
+
+# --- the two levels' pinned tunables (task 04) --------------------------------
+
+
+def test_level_one_is_the_soup_tutorial(bundle: ContentBundle) -> None:
+    level = level_of(bundle, "01_first_kitchen")
+    assert level.order == 1
+    assert level.difficulty == 1
+    assert level.time_limit == 150
+    assert level.target_score == 45
+    assert level.orders.pool == ["soup"]
+    assert level.orders.concurrent == 1
+    assert tuple(level.orders.spawn_interval) == (16, 20)
+    assert level.orders.deadline == 60
+    assert level.hazards == []
+
+
+def test_level_two_adds_salad_and_pressure(bundle: ContentBundle) -> None:
+    level = level_of(bundle, "02_double_shift")
+    assert level.order == 2
+    assert level.difficulty == 2
+    assert level.time_limit == 130
+    assert level.target_score == 80
+    assert level.orders.pool == ["soup", "salad"]
+    assert level.orders.concurrent == 2
+    assert tuple(level.orders.spawn_interval) == (12, 16)
+    assert level.orders.deadline == 52
+    assert level.hazards == []
+
+
+def test_level_two_escalates_at_least_two_axes(bundle: ContentBundle) -> None:
+    before = escalation_axes(level_of(bundle, "01_first_kitchen"))
+    after = escalation_axes(level_of(bundle, "02_double_shift"))
+    moved_up = sorted(
+        name for name, value in after.items() if value > before[name] + 1e-9
+    )
+    assert len(moved_up) >= 2, f"only these axes moved: {moved_up}"