Maze module: generator playground, wall-follower, and maze-solver integration

.claude/settings.json

@@ -0,0 +1,5 @@
+{
+  "enabledPlugins": {
+    "frc2713-skills@frc2713": true
+  }
+}

.github/workflows/deploy-pages.yml

@@ -19,6 +19,26 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v4
+
+      # Build the maze-solver library and vendor its jar into site/public/ so
+      # the Java playground can load the com.frc2713.mazesolver API. Built fresh
+      # each deploy from the library's main; not committed (see site/.gitignore).
+      - uses: actions/checkout@v4
+        with:
+          repository: FRC2713/maze-solver-java
+          ref: main
+          path: maze-solver-java
+      - uses: actions/setup-java@v4
+        with:
+          distribution: temurin
+          java-version: '21'
+          cache: maven
+      - name: Build maze-solver jar
+        run: |
+          mvn -B -q -DskipTests package
+          cp target/maze-solver.jar "$GITHUB_WORKSPACE/site/public/maze-solver.jar"
+        working-directory: maze-solver-java
+
       - uses: actions/setup-node@v4
         with:
           node-version: 20

CLAUDE.md

@@ -166,3 +166,16 @@ folder until it's restarted.
 - **Components**: `LessonCard` (index listing), `PageNav` (in-lesson page
   navigation), `JavaRunner` (the editable/runnable code block UI backed by
   `javaRuntime.ts`).
+- **Maze round-trip** (`site/src/lib/mazeHarness.ts` + the `'java'` mode of
+  `MazePlayground`): the `solver: java` maze fence renders an editor where a
+  student writes a `solve(Robot)` method. `buildMazeHarness` interpolates the
+  current JS maze in as an `int[][]` literal, splices the student's method into a
+  `MazeRun` class, and drives the robot; the **Maze Trail** is emitted as a
+  sentinel-tagged stdout line (`__TRAIL__ [[row,col],…]`), which `parseTrail`
+  reads back (the single `[row,col]→[x,y]` swap) and `MazePlayground` animates.
+  `docs/maze-roundtrip.md` documents the seams. `new GridMaze(grid)` resolves
+  against `site/public/maze-engine.jar` — a concrete engine (GridMaze/GridRobot/
+  GridCell implementing the library interfaces) vendored on the CheerpJ classpath
+  while the `maze-solver` library still ships interfaces only; built by
+  `site/scripts/build-maze-engine.sh` and committed like `tools.jar`. Delete it
+  and the `ENGINE_JAR` classpath entry once the library ships its own `GridMaze`.

CONTEXT.md

@@ -0,0 +1,80 @@
+# Maze Solver Module
+
+Glossary for the maze-solving module of the FRC 2713 training curriculum — a
+series of lessons in which students write Java that solves a maze. This file is
+the shared language for that module only; it is not a spec.
+
+## Language
+
+**Maze**:
+A fixed puzzle the robot must traverse from Start to Goal. The student's program
+is handed the whole maze up front (see _Grid_).
+
+**Grid**:
+The god's-eye representation of a Maze the program receives — the entire layout,
+known in full before the robot moves, as a rectangular 2D array of Cells. Each
+Cell carries its own Wall flags, so the Grid is a `Cell[][]`, not a grid of
+blocked/open squares.
+_Avoid_: map (reserve "map" for a robot-built model, which this module does not use)
+
+**Cell**:
+One square of the Grid — always open and standable. A Cell records a **Wall** on
+each of its four sides (up / down / left / right); those flags, not the Cells
+themselves, are what block movement. Every Cell is reachable to stand on; what
+varies is which of its edges are walled.
+
+**Wall**:
+A **thin** barrier on the edge between two adjacent Cells (or on a Cell's outer
+edge), blocking movement across that edge. Walls live on edges, not whole Cells:
+a Cell is never "a wall". Wall flags are consistent between neighbours — the down
+edge of a Cell is the up edge of the Cell below it.
+
+**Robot**:
+The thing that traverses the Maze by executing a Solution. It has a position on
+the Grid (a Cell) and **no orientation** — it does not face a direction, so every
+Move is an absolute step, not a turn. Orientation is a property of the robot's
+_interface_, not of every algorithm: a wall-follower may choose to *remember*
+which way it last stepped (a `facing` variable it owns), but that state lives in
+the algorithm, not the robot — the robot still only ever takes absolute Moves.
+
+**Start**:
+The cell where the Robot begins.
+
+**Goal**:
+The target cell. A Solution **succeeds** if and only if executing it from Start
+leaves the Robot standing on the Goal.
+
+**Solution**:
+The ordered sequence of Moves the student's program emits. Success is judged
+only by whether it finishes the Maze (reaches the Goal). A Move blocked by a Wall
+is not a modeled failure — it simply does nothing; nothing is penalized. The
+module is about *finishing*, not collision avoidance.
+
+**Maze Trail**:
+The ordered sequence of Cells the Robot has occupied while executing its
+Solution, from Start onward. It is the *consequence* of a Solution, not the same
+thing: a Solution is the Moves the algorithm emits, whereas the Maze Trail is
+where the Robot actually ended up standing after each. A Move blocked by a Wall
+adds no Maze Trail entry, so an algorithm that drives into a wall leaves a Trail
+that simply stays put there.
+_Avoid_: path (ambiguous — could mean the open corridors of the Maze itself)
+
+**Maze Battle**:
+The module's finale. A student submits one program; it is scored by how many of a
+set of *unseen* Mazes it finishes (a gauntlet), tie-broken by total Moves. The
+format rewards a general algorithm (which clears the whole set) over a hard-coded
+Solution (which finishes only the one Maze it was written for).
+
+**Move** (also **Command**):
+One instruction to the Robot. Exactly four exist: **UP**, **DOWN**, **LEFT**,
+**RIGHT** — absolute, screen-relative directions. Each Move steps the Robot
+**one Cell** in that direction *if that edge is open*; a Move blocked by a Wall
+(or the grid edge) does nothing. There are no turns and no orientation to track.
+_Avoid_: turn, heading, facing, North/South/East/West, "drive to the next junction"
+
+**Helper** (injected):
+Convenience methods the harness puts in scope so students query the Maze without
+hand-indexing Wall flags: `robot.moveUp/moveDown/moveLeft/moveRight()`,
+`robot.canGoUp/canGoDown/canGoLeft/canGoRight()`, `robot.atGoal()`,
+`robot.row()/col()`, plus god's-eye access to the whole `Cell[][]` grid for route
+planning.

docs/maze-roundtrip.md

@@ -0,0 +1,226 @@
+# The maze round-trip: JS → Java → animated steps
+
+**Status: built.** The round-trip is wired end to end and drives the interactive
+"write your own solver" playground (the `solver: java` maze fence, last page of
+`lessons/algorithms/README.md`). The glue lives in `site/src/lib/mazeHarness.ts`
+(`buildMazeHarness`, `parseTrail`, `retargetErrorLines`) and the `'java'` mode of
+`site/src/components/MazePlayground.tsx`. This document describes the seams and
+data contracts; a few assumptions in the original spec were corrected against the
+real library and are noted inline below.
+
+**Library shim.** The `maze-solver` library ships only interfaces (Maze/Robot/
+Cell) — no concrete `GridMaze` — so `new GridMaze(grid)` resolves against
+`site/public/maze-engine.jar`, a small concrete engine the site vendors on the
+classpath (built by `site/scripts/build-maze-engine.sh`, sources in
+`site/scripts/maze-engine/`, committed like `tools.jar`). Delete the jar, its
+sources, the build script, and the `ENGINE_JAR` classpath entry in
+`javaRuntime.ts` once the library ships its own `GridMaze`.
+
+Audience: developers working on the site. This is not student-facing lesson
+content.
+
+## What we're building and why
+
+A student writes a maze-solving **algorithm** in Java. They press **Run**, and
+the maze currently shown in the playground — generated in JavaScript — is handed
+to their algorithm, which drives a `Robot` through it. The path the robot took
+is then **animated back** in the same playground.
+
+The motivating win is *removing duplicated logic*. Today
+`site/src/components/MazePlayground.tsx` animates a wall follower that is
+**re-implemented in TypeScript** (`runSolver`, mode `'wall'`) purely so there's
+something to animate. That algorithm already exists in Java and in the
+`maze-solver` library. The round-trip makes the **Java side the single source of
+truth for the path**: JS generates the maze and plays back a result, but never
+re-simulates solving.
+
+The library's job in this is narrow and deliberate: it turns the raw bitmask
+grid into a friendly `Maze`/`Robot`/`Cell` API so the *student* writes an
+algorithm, not file I/O or bitmask arithmetic. See `CONTEXT.md` for the domain
+vocabulary (Maze, Grid, Cell, Robot, Move, Solution, **Maze Trail**).
+
+## Data flow
+
+```mermaid
+sequenceDiagram
+    participant JS as MazePlayground (JS)
+    participant RT as javaRuntime.runJava
+    participant H as Harness (generated Java)
+    participant Stu as student solve(Robot)
+    participant Lib as maze-solver library
+
+    JS->>JS: generateMaze() → int[][] bitmask grid
+    JS->>RT: runJava(harnessSource)   // grid interpolated as literal + solve() spliced in
+    RT->>H: compile + run (CheerpJ)
+    H->>Lib: new GridMaze(grid); maze.robot()
+    H->>Stu: solve(robot)
+    Stu->>Lib: robot.canMove*/move*/atGoal()   // library records the Maze Trail
+    H->>H: print "__TRAIL__ [[row,col],…]" (sentinel line)
+    RT-->>JS: RunOutcome.output (captured stdout)
+    JS->>JS: scan for sentinel, JSON.parse, [row,col]→[x,y]
+    JS->>JS: replay Trail — place robot dot at each cell
+```
+
+## The seams, one at a time
+
+### 1. Generate the maze (JS) — *exists*
+
+`generateMaze(SIZE)` in `MazePlayground.tsx` produces the `number[][]` N/S/E/W
+bitmask grid (`N=1, S=2, E=4, W=8`; a set bit means that side is **open**). This
+is unchanged.
+
+### 2. Pass the maze down (new)
+
+`runJava(code)` accepts **only a source string** — there is no data channel. So
+the current grid is carried in by *interpolating it into the harness source* as
+an `int[][]` literal, reusing the exact format `serializeToJava` already emits
+(the bitmask grid plus `startRow/startCol/goalRow/goalCol`). No VFS file, no
+Java-side parsing.
+
+`serializeToJava` currently targets the clipboard ("Copy for Java"); the new glue
+factors its literal-building out so the harness builder can call it directly.
+
+### 3. Deserialize (Java / library) — *exists in the library*
+
+The harness constructs the friendly API from the literal:
+
+```java
+Maze maze = new GridMaze(grid);
+Robot robot = maze.robot();   // positioned at Start
+```
+
+The student never sees the bitmask grid or `main`.
+
+### 4. Execute the student's algorithm (Java)
+
+The harness calls the student's `solve(Robot)`. The student drives the robot with
+the library's absolute-move API — `robot.canMoveUp/Right/Down/Left()`,
+`robot.moveUp/…()`, `robot.atGoal()`. Each successful move extends the **Maze
+Trail** the library records; a move blocked by a wall does nothing and adds no
+Trail entry (so a buggy algorithm that drives into a wall shows the robot
+*actually stuck there* when animated).
+
+### 5. Emit the steps (new)
+
+After `solve` returns, the **harness** — not the student — prints the Trail as a
+single **sentinel-tagged line**:
+
+```
+__TRAIL__ [[0,0],[0,1],[1,1],[1,2]]
+```
+
+- The `__TRAIL__` prefix makes the line unambiguously extractable even when the
+  student's own `System.out.println` debugging is interleaved in stdout.
+- The payload is a **JSON array of `[row, col]` pairs**, in `row`-major
+  (Java-native) coordinates — see the coordinate note below.
+- It is the full `robot.trail()`, including the Start cell as the first entry, so
+  the animation has a complete path from `trail[0]`.
+
+### 6. Parse the steps back (new)
+
+The JS glue takes `RunOutcome.output`, finds the line beginning with the
+sentinel, and `JSON.parse`s the remainder. `runJava` already returns captured
+stdout in `RunOutcome.output`, and already flags compile/runtime failure via
+`RunOutcome.ok` — the parser only runs on `ok === true` and simply reports "no
+trail found" if the sentinel line is absent.
+
+### 7. Animate the steps (new wiring)
+
+`MazePlayground` replays the parsed Trail by placing the robot dot at each cell
+in sequence (the same interval-driven loop shape as today's `runSolver`, but
+*reading* cells instead of *computing* moves). Because the Trail is exactly what
+happened in Java, the JS side needs **no wall logic at all** — it does not check
+openings, it just plays the cells back.
+
+## The wire contract (authoritative)
+
+| Direction | Payload | Encoding |
+| --- | --- | --- |
+| JS → Java | current maze grid + start/goal | `int[][]` literal interpolated into harness source (`serializeToJava` format) |
+| Java → JS | Maze Trail | one line: `__TRAIL__ ` + JSON `[[row,col],…]` on stdout |
+
+**Coordinate convention — the one conversion point.** Java is row-major
+(`maze[row][col]`, `Cell.row()/col()`), but `MazePlayground`'s robot state is
+`[x, y]` = `[col, row]`. They are transposed. The wire carries **`[row, col]`**
+(Java-native); the JS parser performs the single `[row, col] → [x, y]` swap on
+the way in. Do the swap in exactly one place and nowhere else.
+
+## Harness shape (illustrative)
+
+The app builds a full compilation unit; `prepareSource` in `javaRuntime.ts`
+already runs a declared class with a `main`, so no changes to the runtime's
+wrapping are needed. Roughly:
+
+```java
+import com.frc2713.mazesolver.*;
+
+public class MazeRun {
+    public static void main(String[] args) {
+        int[][] grid = { /* interpolated from the current JS maze */ };
+        Maze maze = new GridMaze(grid);
+        Robot robot = maze.robot();
+
+        solve(robot);   // <-- student's algorithm runs here
+
+        // Emit the Maze Trail as a sentinel-tagged JSON line.
+        StringBuilder sb = new StringBuilder("__TRAIL__ [");
+        Cell[] trail = robot.trail();
+        for (int i = 0; i < trail.length; i++) {
+            if (i > 0) sb.append(',');
+            sb.append('[').append(trail[i].row()).append(',').append(trail[i].col()).append(']');
+        }
+        System.out.println(sb.append(']'));
+    }
+
+    // ===== student-authored, spliced in by the harness builder =====
+    static void solve(Robot robot) {
+        // e.g. wall follower, or whatever the student wrote
+    }
+}
+```
+
+## Edge cases the implementation must handle
+
+- **Algorithm never reaches the Goal.** The Trail is still valid — animate the
+  partial path. `atGoal()` at the end tells you whether it succeeded; the
+  animation can show "stuck"/"gave up" the same way the current playground does.
+- **Student debug output.** Tolerated by design — the sentinel line is found
+  regardless of other stdout. Show the rest of the output as-is if useful.
+- **Compile or runtime error.** Handled by the existing `RunOutcome.ok === false`
+  path (see `simplifyCompileErrors`/`simplifyRuntimeError`); no Trail is parsed.
+- **No sentinel line on success.** Treat as "algorithm produced no trail" and
+  surface a clear message rather than silently animating nothing.
+- **Empty/one-cell Trail.** Robot never left Start — animate nothing / a no-op.
+
+## Exists vs. to-build
+
+**Exists (reuse):**
+- `generateMaze` and the bitmask grid (`MazePlayground.tsx`).
+- `serializeToJava`'s literal format (`MazePlayground.tsx`).
+- `runJava` / `RunOutcome`, stdout capture, compile+run, error simplification,
+  `prepareSource` class-with-`main` handling (`javaRuntime.ts`).
+- The `maze-solver` library API — `GridMaze`, `Robot`, `Cell` (external repo
+  `FRC2713/maze-solver-java`, vendored as `site/public/maze-solver.jar`).
+- `MazePlayground`'s interval-driven robot animation loop.
+
+**To build (the new glue):**
+- A **harness builder** that composes the interpolated grid + spliced `solve`
+  into the `MazeRun` source (factor the literal out of `serializeToJava`).
+- The **`__TRAIL__` sentinel emit** convention in the harness.
+- A **trail parser** on the JS side (sentinel scan → `JSON.parse` → `[row,col]→
+  [x,y]` swap).
+- Wiring `MazePlayground` to call `runJava` and replay the parsed Trail (a new
+  solver "mode" alongside the existing `'random' | 'naive' | 'wall'`).
+
+## Integration details (confirmed against the library)
+
+Resolved against the vendored interface jar's bytecode while wiring this up:
+- `robot.trail()` returns **`int[][]`** — already `[row, col]` pairs, not `Cell[]`.
+  So the harness emits the Trail by iterating the `int[][]` directly (simpler
+  than the `Cell.row()/col()` the earlier draft assumed).
+- `new GridMaze(grid)` returns a `Maze`; `maze.robot()` returns a Start-positioned
+  `Robot`. `GridMaze` is not in the library yet — see the library-shim note at the
+  top; it's supplied by `maze-engine.jar` with Start = (0,0) and Goal =
+  (rows-1, cols-1).
+- The move/query methods are **`canMoveUp/Down/Left/Right`** and
+  **`moveUp/Down/Left/Right`** (the lesson's spelling, not `CONTEXT.md`'s `canGo*`).