@maccesar/titools 4.10.0 → 4.12.0

package/README.md

@@ -435,12 +435,12 @@ When it activates:
 - `tiapp.xml` declares `ti.game`, or any file calls `require('ti.game')`
 - Sprites, sprite sheets, TexturePacker atlases, frame animations, native tweens
 - Physics: gravity and velocity, solid platforms, one-way floors, bouncing, the arcade car model, Newtonian thrust
-- Collision groups, invisible trigger zones, enter/exit events, swept AABB for fast bullets, hitbox tuning
+- Collision groups, invisible trigger zones, enter/exit events, swept AABB for fast bullets, per-axis hitbox tuning with the `debug` overlays
 - Raycasts for line of sight, ledge probes and hitscan shots
 - A* pathfinding (`findPath`) around tagged obstacles, feeding native patrol paths (`followPath`) and animation chaining (`play(name, { then })`)
-- Bitmap-font text sprites and screen-fixed HUDs inside the scene
+- Bitmap-font text sprites with native word wrap, and screen-fixed HUDs inside the scene
 - Particle emitters, Verlet ropes, camera follow/zoom/shake/effects, `scrollFactor` parallax, low-latency sound
-- Game-clock timers that obey pause and slow motion
+- Game-clock timers that obey pause and slow motion, texture `unload()` for level streaming, and what LiveView reloads clean up on their own
 - Drag & drop, pinch, rotate, multitouch controls
 
 Example prompts:

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@maccesar/titools",
-  "version": "4.10.0",
+  "version": "4.12.0",
   "description": "Titanium SDK skills and agents for AI coding assistants (Claude Code, Gemini CLI, Codex CLI)",
   "main": "lib/index.js",
   "type": "module",

package/skills/ti-game/SKILL.md

@@ -1,13 +1,13 @@
 ---
 name: ti-game
-description: "Use when building or reviewing 2D games with the ti.game module for Titanium SDK (Android and iOS) — sprites, sprite sheets, animations, tweens, physics (gravity, solids, carMode drifting, thrust), collision groups, trigger zones, swept AABB, raycast, paths, bitmap text, particles, ropes, camera follow/zoom/shake, parallax, sound, drag & drop, multitouch. AUTO-DETECT: if tiapp.xml declares ti.game or any file calls require('ti.game'), invoke BEFORE writing ANY game code. Also trigger on: game, juego, sprite, spritesheet, platformer, endless runner, top-down, shooter, racer, flappy, breakout, rhythm, cards, puzzle, collision, hitbox, raycast, line of sight, pathfinding, maze, patrol, parallax, particles, HUD, score, game loop, 60 fps, GameView. ti.game is NOT Phaser, Godot or Box2D: you never write a game loop, never move a sprite from setInterval, and several familiar features (tilemap layer, sprite parenting, joystick, word wrap) do not exist yet — check references/roadmap.md before inventing an API."
+description: "Use when building or reviewing 2D games with the ti.game module for Titanium SDK (Android and iOS) — sprites, sprite sheets, animations, tweens, physics (gravity, solids, carMode drifting, thrust), collision groups, trigger zones, swept AABB, raycast, paths, bitmap text, particles, ropes, camera follow/zoom/shake, parallax, sound, drag & drop, multitouch. AUTO-DETECT: if tiapp.xml declares ti.game or any file calls require('ti.game'), invoke BEFORE writing ANY game code. Also trigger on: game, juego, sprite, spritesheet, platformer, endless runner, top-down, shooter, racer, flappy, breakout, rhythm, cards, puzzle, collision, hitbox, raycast, line of sight, pathfinding, maze, patrol, parallax, particles, HUD, score, game loop, 60 fps, GameView. ti.game is NOT Phaser, Godot or Box2D: you never write a game loop, never move a sprite from setInterval, and several familiar features (tilemap layer, sprite parenting, joystick, gamepads) do not exist yet — check references/roadmap.md before inventing an API."
 argument-hint: "[genre or feature]"
 allowed-tools: Read, Grep, Glob, Edit, Write, Bash(node *)
 ---
 
 # ti.game — 2D game engine for Titanium SDK
 
-`ti.game` is a native OpenGL ES 2.0 sprite engine exposed to JavaScript, written and maintained by **Michael Gangolf (m1ga)** — upstream is https://github.com/m1ga/ti.game. This skill tracks upstream `main` as of **2026-08-20**, manifest **`0.4.0`** — that bump is the first in a long run of feature landings (the text engine, `screenFixed`, `swept` collisions, `collisionend`, `followPath`, animation chaining, `raycast`, `findPath`, game-clock timers, `scrollFactor` parallax and the `multiply`/`screen` blend modes all shipped while it still said `0.3.0`), so anything still calling itself 0.3.0 has none of them. When upstream and any fork disagree, upstream wins. Same JS API on Android and iOS — the iOS side is a class-per-class Obj-C twin of the Android engine.
+`ti.game` is a native OpenGL ES 2.0 sprite engine exposed to JavaScript, written and maintained by **Michael Gangolf (m1ga)** — upstream is https://github.com/m1ga/ti.game. This skill tracks upstream `main` as of **2026-08-23**, manifest **`0.4.0`** — that bump is the first in a long run of feature landings (the text engine, `screenFixed`, `swept` collisions, `collisionend`, `followPath`, animation chaining, `raycast`, `findPath`, game-clock timers, `scrollFactor` parallax and the `multiply`/`screen` blend modes all shipped while it still said `0.3.0`), so anything still calling itself 0.3.0 has none of them. The number lags the code in the other direction too: per-axis hitboxes (`hitboxScaleX`/`hitboxScaleY`), text word wrap (`maxWidth`), `unload()` on sheets and fonts, and the LiveView renderer lifecycle all landed on 2026-08-23 with the manifest still at `0.4.0`, so date the build rather than trusting the version. When upstream and any fork disagree, upstream wins. Same JS API on Android and iOS — the iOS side is a class-per-class Obj-C twin of the Android engine.
 
 Works in **both Classic and Alloy** projects. Nothing here depends on PurgeTSS.
 
@@ -33,6 +33,7 @@ The practical consequence, and the single biggest mistake to avoid:
 | Parallax layers under a moving camera | hand-scrolled layers | `scrollFactor` on each layer |
 | Camera that follows the player | timer writing `cameraX` | `gameView.follow(sprite, options)` |
 | A score, a title, a label | `Ti.UI.Label` overlays | `Game.createText({ text, screenFixed: true })` |
+| A dialog box that fits the screen | breaking the lines with `\n` after measuring in JS | `maxWidth` on the text sprite |
 | A bullet fast enough to skip past a thin wall | shrinking the time step | `swept: true` on the bullet |
 | To know the player *left* a zone | polling on a timer | the `collisionend` event |
 | An enemy on a fixed patrol route | chained `animate()` legs | `sprite.followPath(points, { loop })` |
@@ -140,7 +141,7 @@ And when the question is "how do I get over there?", `gameView.findPath(from, to
 
 **A sprite with `width`/`height` but no `sheet` renders nothing and works as an invisible trigger.** Score zones, goals, checkpoints, ceilings, kill floors and walls are all built this way in the official demos. This is the idiomatic way to add logic to a scene without art.
 
-Tune fairness with `hitboxScale` (art rarely fills its frame; slightly small hitboxes feel better) and `hitboxShape: 'circle'` for balls and rocks — circles also resolve against solids along the contact normal, so they bounce off corners diagonally instead of like a box. Turn on `debug: true` per sprite, or on the GameView for everything, to see green collision AABBs, blue touch bounds and the orange anchor dot.
+Tune fairness with `hitboxScale` (art rarely fills its frame; slightly small hitboxes feel better), `hitboxScaleX`/`hitboxScaleY` when the drawing fills its frame by a different fraction on each axis (a 20×44 hero in a 32×48 frame wants `0.62` wide and `0.92` tall — the two multiply `hitboxScale` rather than replace it), and `hitboxShape: 'circle'` for balls and rocks — circles also resolve against solids along the contact normal, so they bounce off corners diagonally instead of like a box. Turn on `debug: true` per sprite, or on the GameView for everything, to see green collision AABBs, blue touch bounds and the orange anchor dot.
 
 ## Performance rules that actually matter
 
@@ -163,6 +164,8 @@ gameView.add(score);
 score.text = 'SCORE 10';   // re-lays out natively
 ```
 
+One long string wraps itself: `maxWidth` breaks lines on word boundaries and re-wraps every time `text` changes, so a dialog box needs no hand-broken `\n`. It is measured in font-space px, *before* `scale` — divide the screen width you want by the scale you are drawing at.
+
 `screenFixed: true` works on **any** sprite: `x`/`y` become surface coordinates and camera position, zoom and shake are ignored, while touch still maps back. That is the HUD pattern — the score stays put while the camera follows the player. Without it, text is world-space and scrolls past like scenery (signposts, floating damage numbers).
 
 Overlaid `Ti.UI.Label` views still work and are still right for anything the engine cannot draw: native inputs, system fonts, accented text the built-in font lacks, scrollable menus. Just do not reach for them for a score any more.
@@ -173,6 +176,8 @@ Standard Titanium touch events do not reach the GameView, but ordinary Titanium
 
 The render loop follows the activity lifecycle automatically, but **your JS timers and sounds do not**. Every demo that starts an interval or plays looping music stops it on window close. In Alloy, wire this into the controller's `cleanup()`.
 
+Two things the engine now cleans up on its own, so do not hand-roll them: a **LiveView reload** retires the previous runtime's render loops, `SoundPool`, `MediaPlayer`s and audio proxies, so reloads stop stacking game loops; and releasing a sheet or font frees its GL texture. What you *can* now do explicitly is `spriteSheet.unload()` / `font.unload()` when streaming levels — permanent, so only after nothing draws from them.
+
 Game-clock timers (`gameView.after` / `every`) are the exception that proves the rule: they run on the engine's clock, so they freeze at `timeScale: 0` and pause with the render loop instead of firing into a dead scene. That is why spawn waves and AI ticks belong there. A `setInterval` keeps running through a pause, which is right for a real-time countdown and wrong for everything else.
 
 ```javascript
@@ -196,16 +201,28 @@ Read the one that matches the task instead of guessing — the API surface below
 | Install the module, tiapp.xml, Alloy vs Classic wiring, assets, sprite sheets, atlases, lifecycle, testing | [project-setup.md](references/project-setup.md) |
 | What does NOT exist yet, what is coming, and how to work around it today | [roadmap.md](references/roadmap.md) |
 
+When this skill runs out — a brand-new feature, a demo's exact code, a build question — read upstream itself at https://github.com/m1ga/ti.game. Each file there is authoritative for something different:
+
+| Upstream file | What it is authoritative for |
+| --- | --- |
+| `README.md` | The module's own API documentation, and the demo index. Canonical, but it has lagged the code: verify a defaulted value against the source before relying on it |
+| `example/` | 26 runnable demos, one per feature area — the reference implementations this skill's recipes are distilled from |
+| `TODO.md` | What the maintainer has planned, with priorities. The source of [roadmap.md](references/roadmap.md) |
+| `tutorial.md` | A step-by-step first scene (sprite, animation, tap-to-move). For a human learning the module, not for looking things up |
+| `ios/README.md` | The iOS port's threading, coordinates and touch constants. Mostly for module developers; the app-facing facts are already here |
+| `AGENTS.md` | Rules for changing the module itself — irrelevant to building a game with it |
+| `android/src/ti/game/`, `ios/Classes/` | The truth. Defaults, clamping and edge cases in this skill were read here, not from prose |
+
 ## Before writing game code, check these
 
 1. Is the level built inside a guarded `resize` handler, using `e.width`/`e.height`?
 2. Does any timer move a sprite? Replace it with velocity, gravity, a tween, `carMode` or `thrust`.
 3. Are all the level's objects added in one `gameView.add([...])` call?
 4. Does every moving body have the right model — `solidWith` to be blocked, `collidesWith` to be notified?
-5. Is `hitboxScale` set on sprites whose art does not fill the frame?
+5. Is `hitboxScale` set on sprites whose art does not fill the frame — with `hitboxScaleX`/`hitboxScaleY` when the fit differs per axis?
 6. Do triggers (scores, goals, kill zones) use sheet-less sprites instead of visible art?
 7. Is every `createSound`, `setInterval` and `setTimeout` stopped when the window closes?
 8. Is the HUD `Game.createText` with `screenFixed: true`, rather than a stack of overlaid labels?
 9. Does anything travel far enough per frame to tunnel? Give it `swept: true`.
 10. Does a timer drive game logic (spawn waves, AI ticks, respawns)? Move it to `gameView.every`/`after` so it obeys pause and slow motion.
-11. Are you using an API that actually exists? Anything involving tilemaps, sprite parenting, joysticks, gamepads or word-wrapped text — check [roadmap.md](references/roadmap.md) first.
+11. Are you using an API that actually exists? Anything involving tilemaps, sprite parenting, joysticks, gamepads or slopes — check [roadmap.md](references/roadmap.md) first.

package/skills/ti-game/references/api.md

@@ -1,6 +1,6 @@
 # ti.game API reference
 
-Complete JS surface of `ti.game` as of upstream `main` on 2026-08-20, verified against the module source (`android/src/ti/game/*.java` and `engine/*.java`, mirrored by `ios/Classes/TG*`). Defaults come from the engine fields, not from prose. The manifest reads **`0.4.0`** since 2026-08-20, and that bump matters: the text engine (`createFont`/`createText`), `screenFixed`, `swept`, `collisionend`, `followPath`, animation chaining, `raycast`, `findPath`, the game-clock timers, `scrollFactor` and the `multiply`/`screen` blend modes all landed while the manifest still said `0.3.0`, so a build calling itself 0.3.0 has none of them.
+Complete JS surface of `ti.game` as of upstream `main` on 2026-08-23, verified against the module source (`android/src/ti/game/*.java` and `engine/*.java`, mirrored by `ios/Classes/TG*`). Defaults come from the engine fields, not from prose. The manifest reads **`0.4.0`** since 2026-08-20, and that bump matters: the text engine (`createFont`/`createText`), `screenFixed`, `swept`, `collisionend`, `followPath`, animation chaining, `raycast`, `findPath`, the game-clock timers, `scrollFactor` and the `multiply`/`screen` blend modes all landed while the manifest still said `0.3.0`, so a build calling itself 0.3.0 has none of them. The version number is not a feature list either: `hitboxScaleX`/`hitboxScaleY`, text `maxWidth`, `SpriteSheet.unload()`/`Font.unload()` and the LiveView renderer lifecycle all landed on 2026-08-23 with the manifest still at `0.4.0`, so two builds calling themselves 0.4.0 can differ — check the build date, or feature-detect by reading the property **before** ever writing it (`typeof sprite.hitboxScaleX === 'undefined'` means the build predates it). Writing first proves nothing: an unknown property is stored on the Kroll proxy and reads back exactly as it was set.
 
 Every property listed is **live**: reading returns the current native value, even mid-drag or mid-tween. Every property can also be passed to its `create*` factory. All durations crossing the JS boundary are **milliseconds** (the engine converts to seconds internally).
 
@@ -128,8 +128,9 @@ Create one sheet per texture and share it across every sprite that uses it — s
 | `frameCount` | Number of frames. `0` until the texture has loaded, for grid sheets |
 | `frameNames` | Sorted frame names — atlas sheets only |
 | `frameIndex(name)` | Index for an atlas frame name, `-1` if unknown |
+| `unload()` | Frees the GL texture on the next rendered frame, from the render thread. **Permanent** — sprites still using the sheet stop drawing. For level streaming: unload the finished level's atlases instead of accumulating GPU memory. Since 2026-08-23 |
 
-Textures upload to the GPU lazily on first use, from the render thread, and are re-created automatically after an EGL context loss.
+Textures upload to the GPU lazily on first use, from the render thread, and are re-created automatically after an EGL context loss. Releasing the proxy unloads its texture too, so a sheet that goes out of scope does not strand GPU memory.
 
 ## Sprite
 
@@ -206,7 +207,8 @@ Hit-testing runs against the transformed shape (rotation and scale included), to
 | `collisionGroup` | — | The tag *this* sprite carries |
 | `collidesWith` | — | Array of groups this sprite reports overlaps with |
 | `hitboxScale` | `1` | Shrinks the collision box around the anchor |
-| `hitboxShape` | `'rect'` | `'circle'` — radius = half the smaller drawn side × `hitboxScale`. Circles resolve against solids along the contact normal (corner bounces) and get a round touch area |
+| `hitboxScaleX`, `hitboxScaleY` | `1` | Per-axis corrections **multiplied on top of** `hitboxScale`, for art that fills its frame by a different fraction on each axis. A 20×44 drawing in a 32×48 frame needs `0.62` wide and `0.92` tall — no single `hitboxScale` describes it. Ignored by circle hitboxes. Since 2026-08-23 |
+| `hitboxShape` | `'rect'` | `'circle'` — radius = half the smaller drawn side × `hitboxScale` (the per-axis scales are skipped: a circle has no axes). Circles resolve against solids along the contact normal (corner bounces) and get a round touch area |
 | `swept` | `false` | Test this sprite's movement as a **path** (swept AABB), not just at the end position, so a fast mover cannot tunnel between frames. Applies to both `collidesWith` events and `solidWith` blocking (the sprite is clamped to the impact point, then resolved by the normal static pass). Circle hitboxes sweep as their bounding box. Set it on the *mover* — the bullet, not the wall |
 | `debug` | `false` | Draw this sprite's shapes: green = collision AABB, blue = touch bounds, orange dot = anchor |
 
@@ -306,6 +308,7 @@ const mono = Game.createFont({ image: 'assets/mono.png', charWidth: 9, charHeigh
 | `characters` | ASCII 32..126 | Grid mode: which characters the cells map to, row-major |
 | `smoothing` | `true` | Filters the glyph texture like a sheet. The built-in font is always crisp |
 | `lineHeight` | — | Read-only: the font's natural line height in px |
+| `unload()` | Method, not an option: frees the glyph texture exactly like `SpriteSheet.unload()`, and just as permanently. Since 2026-08-23 |
 
 A descriptor that fails to parse falls back to the built-in font and logs the error — the text still renders, in the wrong face, which is the symptom to recognize.
 
@@ -335,8 +338,22 @@ score.text = 'SCORE 10';   // re-lays out natively on the next frame
 | `align` | `'left'` | `'center'` / `'right'` — how multiple lines align against each other. Unknown values fall back to `'left'` |
 | `letterSpacing` | `0` | Extra px between glyphs; negative tightens |
 | `lineSpacing` | `1` | Multiplier on the font's line height |
+| `maxWidth` | `0` | Wrap width in **font-space px** (`0` = no wrap). Lines break on word boundaries, re-wrapping whenever `text` is written. Since 2026-08-23 |
 | `width`, `height` | laid-out size | **Derived, not settable** — the glyph layout drives the size, and with it the anchor, hit test, AABB and `ySort` bottom edge |
 
+`maxWidth` wraps a single long string so dialog boxes stop needing hand-broken `\n` lines — hard `\n` breaks still apply on top of it. Measurement reuses the layout pen (kerning, `letterSpacing`, the missing-glyph advance), so a wrapped line never renders wider than it measured, and the spaces around a soft break are dropped.
+
+```javascript
+const dialog = Game.createText({
+	text: 'A WISE FROG ONCE SAID: THE POND LOOKS SMALL UNTIL YOU TRY TO HOP ACROSS IT.',
+	maxWidth: Math.round(W * 0.55 / UNIT),   // font-space px, so divide by the scale
+	align: 'center',
+	scale: UNIT,
+	lineSpacing: 1.3
+});
+dialog.text = nextLine;   // re-wraps natively
+```
+
 There is no font size: scale the sprite. With the built-in pixel font, use **integer** `scale` values (`Math.max(1, Math.round(W / 240))` is the demos' idiom) so texels stay square.
 
 Every glyph is a quad in the same batch, so one label costs one draw call — a screen of labels sharing a font costs one too.
@@ -438,7 +455,14 @@ There is `collision` (enter) and `collisionend` (exit), but deliberately **no**
 - **`swept` belongs on the fast mover, not on the target.** A stationary thin wall with `swept: true` changes nothing; the bullet is the sprite whose path needs testing.
 - **A swept bullet can enter and leave in consecutive frames.** If it crosses a thin target entirely, you get `collision` and then `collisionend` almost immediately — react on the enter.
 - **Text `width`/`height` are read-only in practice.** They report the laid-out block; to make text bigger use `scale`.
+- **`unload()` is permanent, not a cache eviction.** There is no reload: sprites and text still pointing at an unloaded sheet or font simply stop drawing. Unload a level's atlases after its sprites are gone, never to "free memory" on something still on screen.
+- **A tap is a press released within 300 ms**, inside the touch slop (the platform's scaled slop on Android, 8 pt × screen scale on iOS). Hold a button longer and you get `press` and `release` but no `tap` — which is why the demos drive held controls from `press`/`release` and reserve `tap` for discrete hits.
 - **`hitboxShape: 'circle'` also changes the touch area**, not just collisions.
+- **The hitbox scales do not change the touch area.** `hitTest` runs against the full drawn frame, so a sprite with `hitboxScale: 0.6` still takes taps out to its art's edges. Only the hitbox *shape* reaches touch. Shrink `width`/`height` if a tap target must match the collision box.
+- **`hitboxScaleX`/`hitboxScaleY` multiply `hitboxScale`, they do not replace it.** `hitboxScale: 0.85` with `hitboxScaleX: 0.66` gives an effective 0.561 on X and 0.85 on Y. Both default to `1`, so sprites that never set them are unaffected — and `0` still means zero, as everywhere else in the engine.
+- **Text `maxWidth` is in font-space px, before `scale`.** The wrap width on screen is `maxWidth * scale`, so a label at `scale: 3` wrapping at 55% of the screen wants `maxWidth: W * 0.55 / 3`. Passing screen pixels straight in produces a block three times too wide.
+- **`align` centers against the widest wrapped line, not against `maxWidth`.** The block's `width` is what the glyphs actually occupied, so a short wrapped line is not centered inside the wrap column — anchor the sprite where the block should sit.
+- **A word longer than `maxWidth` overflows rather than breaking mid-word.** There is no hyphenation and no character-level break; a long unbroken token widens the block past the limit.
 - **Every blend-mode change costs a batch flush.** Group same-blend sprites on their own `zIndex` band; alternating modes sprite by sprite degrades toward one draw call each.
 - **`glowColor` alone draws nothing.** The glow pass only runs when `glowBlur > 0` *and* `glowOpacity > 0`; both default in a way that makes `glowColor` on its own a no-op (`glowBlur` is `0`). Always set the blur radius with the color.
 - **`followPath` overrides position absolutely.** The path is applied after velocity and gravity have been integrated and writes `x`/`y` outright, so a path-driven sprite ignores its own physics for placement. Do not expect `velocityX` and a path to add up. A tween on `x`/`y` runs *after* the path in the same frame and wins outright — pick one.
@@ -447,7 +471,7 @@ There is `collision` (enter) and `collisionend` (exit), but deliberately **no**
 - **`raycast` groups: pass an array, not varargs.** Android accepts both `raycast(x0, y0, x1, y1, ['enemy'])` and loose arguments; iOS reads only the array. Write the array and it works on both.
 - **`raycast` only sees `visible`, non-`screenFixed` sprites carrying a `collisionGroup`.** A hidden pooled sprite is invisible to the ray, which is usually what you want; an untagged wall is invisible too, which usually is not. HUD sprites are skipped by design, so a screen-fixed score never blocks a world-space line of sight.
 - **`raycast` is a discrete query, not a sensor.** Calling it from a game-clock timer or a tap handler is the intended use. Calling it every frame from JS puts exactly the bridge traffic in the loop that the whole engine exists to avoid.
-- **`findPath` walks the sprite's center, not its silhouette.** The waypoints are a line through free grid cells, so a body wider than a cell clips corners unless you pass `clearance` (roughly half the walker's width) or shrink `cellSize`. Obstacles are read from the **hitbox**, so `hitboxScale` and `hitboxShape: 'circle'` shape the walkable space too.
+- **`findPath` walks the sprite's center, not its silhouette.** The waypoints are a line through free grid cells, so a body wider than a cell clips corners unless you pass `clearance` (roughly half the walker's width) or shrink `cellSize`. Obstacles are read from the **hitbox**, so `hitboxScale`, the per-axis scales and `hitboxShape: 'circle'` shape the walkable space too.
 - **`findPath` returning `null` is not always "no route".** Degenerate `bounds`, `cellSize <= 0` and a grid over ~1M cells all return `null` as well — a `bounds` rect of the whole world at `cellSize: 4` fails silently that way. Check the numbers before blaming the maze.
 - **A path of one or two points is the straight-line case, not a bug.** With nothing in the way `findPath` returns just `[start, goal]`; `path.length < 2` is the guard the demos use to detect a tap outside `bounds`.
 - **Game-clock timers are not `setTimeout`.** `after`/`every` freeze at `timeScale: 0` and pause with the render loop — which is why spawn waves belong there — but a pause menu that needs a real countdown still needs `setTimeout`.

package/skills/ti-game/references/project-setup.md

@@ -41,7 +41,16 @@ Then declare it in `tiapp.xml`. The iOS platform key is `iphone`, not `ios`:
 
 Pinning `version` is worth the extra attribute: with two module versions installed side by side, an unpinned entry silently picks the highest one.
 
-The manifest went `0.3.0` → **`0.4.0`** on 2026-08-20, after months of feature landings under the old number — so `0.4.0` is the first version string that actually means something: it carries everything this skill documents. `0.3.0` does not, and there are several different `0.3.0` builds in the wild. One built before 2026-08-18 has no `createText`, no `createFont`, no `screenFixed`, no `swept` and no `collisionend`; one built before 2026-08-19 also lacks `followPath`, `play(name, { then })`, `raycast`, `after`/`every`, `scrollFactor` and the `multiply`/`screen` blend modes; one built before 2026-08-19 16:00 UTC+2 additionally bleeds at grid-frame edges on smoothed sheets. `findPath` exists only in `0.4.0`. If a call that this skill documents is `undefined` at runtime, check the manifest version first and rebuild from upstream `main` rather than hunting for a typo.
+**The version string does not identify a feature set — the build date does.** The manifest went `0.3.0` → **`0.4.0`** on 2026-08-20, after months of feature landings under the old number, and then kept accepting features without moving again. So both numbers cover several incompatible builds:
+
+| A build from | Missing, compared to what this skill documents |
+| --- | --- |
+| `0.3.0` before 2026-08-18 | `createText`, `createFont`, `screenFixed`, `swept`, `collisionend` |
+| `0.3.0` before 2026-08-19 | also `followPath`, `play(name, { then })`, `raycast`, `after`/`every`, `scrollFactor`, `multiply`/`screen` blends |
+| `0.3.0` before 2026-08-19 16:00 UTC+2 | additionally bleeds at grid-frame edges on smoothed sheets |
+| `0.4.0` — including the 2026-08-20 release zip | `hitboxScaleX`/`hitboxScaleY`, text `maxWidth`, `SpriteSheet.unload()`/`Font.unload()` and the LiveView renderer lifecycle, all added on 2026-08-23 with the manifest untouched |
+
+The published releases are pre-releases and lag `main`: the newest one is `0.4.0` (2026-08-20). If a call that this skill documents is `undefined` at runtime, check the manifest version *and the build date* before hunting for a typo, and rebuild from upstream `main`. More reliable than either is feature-detecting in code: read the property **before** ever writing it (`typeof sprite.hitboxScaleX === 'undefined'` means the build predates it). Write-then-read always says yes, because an unknown property is stored on the Kroll proxy and reads back exactly as it was set.
 
 Building the module from source:
 
@@ -287,9 +296,32 @@ function listen(target, eventName, handler) {
 
 For restart-without-teardown (a retry button), `removeAllSprites()` plus rebuilding is simplest, but pooling and repositioning is cheaper: reset positions, zero velocities, `clearTweens()`, and re-`play()` the animations.
 
+### GPU textures: `unload()`
+
+Sheets and fonts hold GL textures, and until 2026-08-23 nothing but losing the context ever freed one. `spriteSheet.unload()` and `font.unload()` release the texture on the next rendered frame, from the render thread. Releasing the proxy does the same, so a sheet that goes out of scope no longer strands its memory.
+
+This is for **level streaming** — a game that loads world 2's atlases while world 1's are still resident. It is not a cache: the unload is permanent, and any sprite still pointing at that sheet stops drawing.
+
+```javascript
+function unloadLevel(level) {
+	level.sprites.forEach(s => gameView.remove(s));
+	level.sheets.forEach(sheet => sheet.unload());   // after the sprites are gone
+}
+```
+
+A single-level game with a handful of sheets should not call it at all: the textures die with the window.
+
+### LiveView reloads
+
+A LiveView reload replaces Titanium's JS runtime while the native side survives, which used to leave the previous `GameView`'s render loop running — every reload added another GL thread, and the audio engine handed the new runtime sample ids the old one still owned.
+
+Since 2026-08-23 the module retires the previous generation itself: stale renderers are shut down (the `GLSurfaceView` is detached, which stops its `GLThread`), the old `SoundPool`, `MediaPlayer`s and audio proxies are released, and on Android `GameViewProxy.releaseViews()` drops the view reference so the `GLSurfaceView` — and the Activity it holds — can be collected after a window close. iOS does the same by hooking the LiveView restart and shutting down the views registered to the retiring runtime context.
+
+**Nothing in your code needs to change**, and this is not a substitute for the cleanup above: your own timers, sounds and listeners still belong to you on a normal window close. It does mean that a reload no longer accumulates game loops, so a stuttering scene after ten reloads is a symptom that dates the build.
+
 ## Debugging and tuning
 
-- **`debug: true` on the GameView** draws collision shapes for the whole scene; on a single sprite, only that one. Green = collision AABB with `hitboxScale` applied, blue = sprite/touch bounds, orange dot = anchor. This is the fastest way to explain a hit that should not have registered.
+- **`debug: true` on the GameView** draws collision shapes for the whole scene; on a single sprite, only that one. Green = collision AABB with `hitboxScale` and the per-axis `hitboxScaleX`/`hitboxScaleY` applied, blue = sprite/touch bounds (the full drawn frame, which is also the touch area), orange dot = anchor. This is the fastest way to explain a hit that should not have registered.
 - **Physics that feels wrong is usually a units problem.** Speeds are px/s and accelerations px/s², both in surface pixels — which is why the demos express them as fractions of `W` and `H`. A gravity of `20` is imperceptible; `H * 2.2` is a snappy platformer.
 - **Sprite is invisible?** Check in order: was it added to the GameView, does it have a sheet, is `width`/`height` non-zero, is it inside the surface bounds, and is its `zIndex` below a background sprite.
 - **Collision never fires?** `collisionGroup` goes on the *target*, `collidesWith` on the *mover*. Both need to be set, and `visible: false` disables collision entirely.

package/skills/ti-game/references/recipes.md

@@ -1,13 +1,14 @@
 # ti.game recipes by genre
 
-Working patterns distilled from the 25 official demos in `example/` and from a shipped Alloy game (Titanium Lander). Each recipe shows the part that is specific to the genre; the shared scaffolding is in [The scaffolding](#the-scaffolding) and is not repeated.
+Working patterns distilled from the 26 official demos in `example/` and from a shipped Alloy game (Titanium Lander). Each recipe shows the part that is specific to the genre; the shared scaffolding is in [The scaffolding](#the-scaffolding) and is not repeated.
 
-Every snippet uses only API that exists in upstream `main` as of 2026-08-20 (manifest `0.4.0`). Check [roadmap.md](roadmap.md) before reaching for anything not shown here.
+Every snippet uses only API that exists in upstream `main` at `df66122`, 2026-08-23 (manifest `0.4.0`, unchanged since 2026-08-20 — the version number stopped tracking the code). Check [roadmap.md](roadmap.md) before reaching for anything not shown here.
 
 ## Contents
 
+- [The 26 demos, and what each one is the reference for](#the-26-demos-and-what-each-one-is-the-reference-for)
 - [The scaffolding](#the-scaffolding)
-- [Cross-cutting patterns](#cross-cutting-patterns) — pooling, parallax, multitouch controls, HUD text, invisible triggers, enter/exit zones, patrol routes, line of sight, pathfinding, game-clock timers, hit-stop
+- [Cross-cutting patterns](#cross-cutting-patterns) — pooling, freeing textures, parallax, multitouch controls, HUD text, wrapped dialog, invisible triggers, seeing the hitbox, enter/exit zones, patrol routes, line of sight, pathfinding, game-clock timers, hit-stop
 - [Tap-to-flap](#tap-to-flap) (`flappy.js`)
 - [Platformer](#platformer) (`platformer.js`)
 - [Top-down / Zelda](#top-down--zelda) (`topdown.js`)
@@ -24,6 +25,39 @@ Every snippet uses only API that exists in upstream `main` as of 2026-08-20 (man
 - [Camera work](#camera-work) (`camera.js`)
 - [Visual effects](#visual-effects) (`blend.js`, `flip.js`, `timescale.js`, `demoscene.js`)
 
+## The 26 demos, and what each one is the reference for
+
+`example/` is the module's real documentation: every demo is a runnable app, and upstream keeps them current with the engine. When a recipe here is not enough, open the demo — it compiles today.
+
+| Demo | The reference implementation of | Covered here |
+| --- | --- | --- |
+| `basic.js` | Sheets, animations, drag/pinch/rotate, tween chaining | [The scaffolding](#the-scaffolding) |
+| `puzzle.js` | Drag & drop with snapping, press-to-lift, one piece per finger | [Cards and board games](#cards-and-board-games) |
+| `flappy.js` | Gravity + tap impulse, score zones, wrapping parallax | [Tap-to-flap](#tap-to-flap) |
+| `platformer.js` | `solidWith`, `land`, `oneWay` stairs, a moving platform that carries you, d-pad | [Platformer](#platformer) |
+| `volley.js` | `restitution`, JS hit response, an AI timer | [Ball sports / breakout physics](#ball-sports--breakout-physics) |
+| `racing.js` | `carMode` drift, skid marks, lap and checkpoint logic | [Top-down racer](#top-down-racer) |
+| `cards.js` | Deck dealing, fanned hand, selection tweens, idle wobble | [Cards and board games](#cards-and-board-games) |
+| `asteroids.js` | `thrust`/`angularVelocity`, `wrapAround`, bullet pooling, additive bolts | [Asteroids / space shooter](#asteroids--space-shooter) |
+| `topdown.js` | Tile map from a string array, `ySort` depth, follower NPC | [Top-down / Zelda](#top-down--zelda) |
+| `skate.js` | Endless runner: pooled obstacles, jump button, pixel-art parallax | [Endless runner](#endless-runner) |
+| `pointclick.js` | Tap-to-walk with `findPath` + `followPath`, verb coin, hotspots | [Point & click adventure](#point--click-adventure) |
+| `particles.js` | Fountain, tap bursts, a smoke trail following a dragged sprite | [Particles](#particles) |
+| `rhythm.js` | Pooled notes on native velocity, `press` pads, timing windows | [Rhythm game](#rhythm-game) |
+| `camera.js` | Dead-zone follow, `cameraBounds`, zoom, shake, `cameraEffect`, `scrollFactor` | [Camera work](#camera-work) |
+| `rope.js` | Verlet ropes from a dragged head and from a fixed anchor | [Ropes and chains](#ropes-and-chains) |
+| `flip.js` | `flipX`/`flipY` driven by movement; tap inverts gravity | [Visual effects](#visual-effects) |
+| `hitbox.js` | `debug: true` overlays, and why the per-axis hitbox scales exist | [Seeing the hitbox](#seeing-the-hitbox--hitboxjs) |
+| `blend.js` | The four blend modes side by side, `flash()` variants | [Visual effects](#visual-effects) |
+| `text.js` | Screen-fixed HUD, world-space labels, text buttons, wrapped dialog | [HUD](#hud--textjs-flappyjs) |
+| `swept.js` | `swept: true` vs tunneling, same bullets, rising speed | [Object pooling](#object-pooling) |
+| `path.js` | Smoothed circuits, sharp patrols, `play(name, { then })` chains | [Patrol routes and animation chains](#patrol-routes-and-animation-chains--pathjs) |
+| `raycast.js` | Line of sight, ledge probes, tap-fired hitscan | [Line of sight, ledge probes and hitscan](#line-of-sight-ledge-probes-and-hitscan--raycastjs) |
+| `zones.js` | The `collision`/`collisionend` lifecycle, including despawned partners | [Zones you can be inside of](#zones-you-can-be-inside-of-enter--exit--zonesjs) |
+| `demoscene.js` | Sine scrollers and copper bars built from `followPath` circles | [Visual effects](#visual-effects) |
+| `maze.js` | A* over a tile maze, `simplify: false` route visualization, a re-pathing hound | [Pathfinding around obstacles](#pathfinding-around-obstacles--mazejs-pointclickjs) |
+| `timescale.js` | `timeScale` slow motion and freeze; game clock vs `setInterval` | [Timers on the game clock](#timers-on-the-game-clock) |
+
 ## The scaffolding
 
 Every game repeats this shape. Sheets and sounds are created once, outside `resize`, because they are texture handles, not scene objects.
@@ -135,10 +169,23 @@ function fire(x, y, vx, vy) {
 
 The `generation` counter matters: without it, a timeout fired for an old shot deactivates the *new* bullet occupying that slot.
 
-`swept: true` (compared side by side in `swept.js`) is what keeps a fast bolt from crossing a rock between two frames without ever overlapping it. The asteroids demo predates it and inflates the hitbox instead (`hitboxScale: 1.4`) — that works, at the price of counting near misses as hits. Prefer `swept`, and reserve `hitboxScale` for fairness tuning.
+`swept: true` (compared side by side in `swept.js`) is what keeps a fast bolt from crossing a rock between two frames without ever overlapping it. The asteroids demo predates it and inflates the hitbox instead (`hitboxScale: 1.4`) — that works, at the price of counting near misses as hits. Prefer `swept`, and reserve `hitboxScale` — with `hitboxScaleX`/`hitboxScaleY` when the art's fit differs per axis — for fairness tuning.
 
 Recycling a sprite with `visible = false` fires `collisionend` on anything it was touching. Handlers that treat separation as "the player walked away" need to tolerate a despawn.
 
+### Freeing a level's textures
+
+Pooling recycles sprites; it does not free the **texture** behind them. A game that streams worlds accumulates every atlas it has ever loaded until the GL context dies. `unload()` releases one on the next rendered frame.
+
+```javascript
+function unloadLevel(level) {
+	level.sprites.forEach((s) => gameView.remove(s));
+	level.sheets.forEach((sheet) => sheet.unload());   // only once nothing draws from them
+}
+```
+
+It is permanent — not an eviction that reloads on demand. A sprite still pointing at an unloaded sheet stops drawing, silently. A single-level game should not call it: the textures die with the window.
+
 ### Parallax
 
 Two ways, and the choice is decided by whether the **camera** moves or the **world** does.
@@ -266,6 +313,28 @@ Because a label is a sprite, the HUD gets the whole toolbox for free — `animat
 
 Drop `screenFixed` and the same label becomes world furniture: signposts, floating damage numbers, a name over an NPC — it scrolls with everything else.
 
+A dialog box is one long string with `maxWidth`: the engine breaks the lines on word boundaries and re-wraps every time `text` is written, so the source keeps no hand-placed `\n`.
+
+```javascript
+const UNIT = Math.max(1, Math.round(W / 240));
+const dialog = Game.createText({
+	text: 'TAP FOR POND WISDOM',
+	maxWidth: Math.round(W * 0.55 / UNIT),   // font-space px: divide by the scale
+	align: 'center',
+	lineSpacing: 1.3,
+	scale: UNIT,
+	x: W / 2,
+	y: H * 0.78,
+	zIndex: 3
+});
+dialog.addEventListener('tap', () => {
+	dialog.text = lines[i++ % lines.length];   // re-wraps natively
+	dialog.flash('#fff', 150);
+});
+```
+
+`maxWidth` is measured before `scale`, which is the only trap here — pass screen pixels straight in and the block comes out `scale` times too wide. `align` positions the lines against each other (the block's own width), not inside the wrap column, so anchor the sprite where the text should sit. A word longer than `maxWidth` overflows instead of breaking mid-word.
+
 Overlaid `Ti.UI.Label` views remain the answer for text the engine's font cannot draw (accents, system fonts, right-to-left) and for real UI: menus, dialogs, text inputs.
 
 ### Invisible triggers
@@ -298,6 +367,19 @@ car.addEventListener('collision', (e) => {
 });
 ```
 
+### Seeing the hitbox — `hitbox.js`
+
+`debug: true` on the GameView is the whole tuning workflow, and the demo exists to make the overlays legible: two identical adventurers walk into the same wall, one untuned and one with `hitboxScaleX: 0.62` / `hitboxScaleY: 0.92`. The untuned one stops a body's width short and hovers above the floor on his frame's padding; the tuned one goes flush and lands his feet. Tapping toggles the tuning live, so the green box snaps between frame and drawing.
+
+```javascript
+const gameView = Game.createGameView({ debug: true });   // green = collision AABB, blue = drawn frame, orange = anchor
+
+hero.hitboxScaleX = 0.62;   // both are live: write them and watch the box move
+hero.hitboxScaleY = 0.92;
+```
+
+Read the two boxes as a pair: green is what collides, blue is what draws **and what takes taps**. Green much smaller than blue on one axis only is exactly the case the per-axis scales exist for; green and blue identical on a sprite whose art has padding is the bug the demo dramatizes.
+
 ### Zones you can be inside of (enter / exit) — `zones.js`
 
 `collision` and `collisionend` are the two halves of a trigger: entering water, standing on a pressure plate, holding a capture point. Nothing fires while you stay inside — the state lives in JS between the two events, and that is deliberate, since a per-frame "stay" would be bridge traffic every frame.
@@ -505,6 +587,7 @@ const pig = Game.createSprite({
 	height: PIG_SIZE,
 	zIndex: 10,
 	hitboxScale: 0.7,                 // the art does not fill the frame
+	hitboxScaleY: 0.85,               // ...and the pig is wider than he is tall
 	collidesWith: ['pipe', 'ground', 'score', 'ceiling'],
 	animations: { fly: { frames: [0, 1, 2, 1], fps: 10, loop: true } }
 });
@@ -554,6 +637,7 @@ const player = Game.createSprite({
 	zIndex: 10,
 	gravity: GRAVITY,                        // H * 2.2
 	hitboxScale: 0.85,
+	hitboxScaleX: 0.66,                      // the blob is 38px wide in a 64px frame, but nearly full height
 	solidWith: ['solid', 'trampoline'],
 	animations: {
 		idle: { frames: [0], fps: 1, loop: true },

package/skills/ti-game/references/roadmap.md

@@ -1,6 +1,6 @@
 # What does not exist yet — and what to do instead
 
-The most expensive mistake with `ti.game` is writing code against a feature that sounds like it should be there because Phaser, Godot or Cocos has it. This file is the guard rail: everything below is **absent from upstream `main` as of 2026-08-20** (manifest `0.4.0`), with the workaround that is idiomatic today.
+The most expensive mistake with `ti.game` is writing code against a feature that sounds like it should be there because Phaser, Godot or Cocos has it. This file is the guard rail: everything below is **absent from upstream `main` as of 2026-08-23** (`df66122`, manifest `0.4.0`), with the workaround that is idiomatic today.
 
 Planned items come from the module's `TODO.md`. Planned is not shipped — when a game needs one of these, use the workaround and keep the code isolated enough to swap later.
 
@@ -11,7 +11,6 @@ Planned items come from the module's `TODO.md`. Planned is not shipped — when
 | A tilemap layer, Tiled JSON import, collision layer | Planned (priority 4) | One sprite per tile. Fine at 12×15, dead at 200×200. For large ground planes use a single `tileRepeat` sprite instead of a grid, and add sparse invisible solids for collision. `findPath` already rasterizes those per-tile sprites, so a tile maze pathfinds today — the planned collision layer is meant to feed the same grid |
 | `sprite.parent`, attaching a turret to a tank, a hat to a hero, multi-part bosses | Planned (priority 7), the biggest structural item | Move the parts together from JS on the same coarse tick, or bake the combination into sheet frames. There is no transform inheritance |
 | `gameView.panTo(x, y, { duration, easing })` for a cutscene camera move | Planned (priority 1) | `follow()` an invisible sprite and tween that sprite, then `follow()` the player again |
-| `maxWidth` word wrap on text sprites | Planned (priority 5) | Break the lines yourself with `\n`. Measure with the font's `charWidth` on a monospace grid font; a proportional font needs a per-glyph tally |
 | Ping-pong animation playback, a per-sprite animation speed multiplier, a random start offset | Planned (priority 9) | Author the ping-pong into the sheet as a longer animation. Desync a field of torches by giving each one its own `idleAnimation` timing, or start them from staggered timers |
 | Native virtual joystick or d-pad bound to a sprite | Planned (priority 6) | Overlaid Titanium views, one per button, writing `velocityX`/`steering`/`throttle` on `touchstart`/`touchend`. Sibling views get simultaneous pointers, so multitouch works |
 | Gamepad support | Planned (priority 6) | Nothing native. Touch controls only |
@@ -57,7 +56,11 @@ Recent additions that are easy to miss and easy to reimplement badly in JS:
 - **`scrollFactor`** (2026-08-19): per-sprite parallax against camera travel. A background layer is one property, not a hand-scrolled pair of copies.
 - **`blend: 'multiply'` and `'screen'`** (2026-08-19) join `'normal'` and `'add'`, on sprites and emitters alike.
 - **`gameView.findPath(from, to, options)`** (2026-08-20): grid A* around the sprites carrying a `collisionGroup`, returning waypoints ready for `followPath` — the Godot `AStar2D` / GameMaker `mp_grid` equivalent. Tap-to-walk and chasing AI route around obstacles natively now; hand-authored waypoint arrays and straight-line walking are the old workaround. Discrete like `raycast`: taps and AI timers, not per frame.
+- **Text word wrap** (2026-08-23): `maxWidth` on a text sprite breaks lines on word boundaries and re-wraps whenever `text` changes, so a dialog box is one long string instead of hand-broken `\n` lines. The width is in font-space px, before `scale`. Do not write a JS line-breaker any more.
+- **`hitboxScaleX` / `hitboxScaleY`** (2026-08-23): per-axis corrections multiplied on top of `hitboxScale`, for art that fills its frame by a different fraction on each axis. The old workaround — settling for one compromise scale, or padding the frame so a single number fits — is obsolete.
+- **`spriteSheet.unload()` / `font.unload()`** (2026-08-23): frees a GL texture on the next frame instead of holding it until the context dies. Level streaming no longer needs a fresh window to reclaim GPU memory. Permanent — it is not a cache eviction.
+- **LiveView reloads retire the previous generation** (2026-08-23): stale render loops, the old `SoundPool`, `MediaPlayer`s and audio proxies are shut down by the module itself. Do not write your own reload guard, and do not read a scene stuttering after several reloads as an engine limit — it dates the build.
 
 ## Checking the current state
 
-This file describes upstream `main` on 2026-08-20 — a moving target: nine items on this list shipped within four days of the skill being written, four of them in a single day, and the manifest only caught up at the end (`0.3.0` → `0.4.0` on 2026-08-20, after all of them). Before relying on any "planned" item being still absent — or on a workaround still being necessary — read `TODO.md` and `README.md` at https://github.com/m1ga/ti.game, the module's upstream repo. `README.md` there is the canonical API documentation; the `documentation/` folder only points at it. The priorities quoted above are the maintainer's, not a roadmap this skill decides.
+This file describes upstream `main` at `df66122` on 2026-08-23 — a moving target: thirteen items on this list shipped within a week of the skill being written, four of them in a single day, and the manifest tracks none of it precisely (`0.3.0` → `0.4.0` on 2026-08-20, after nine of them; word wrap, the per-axis hitbox, `unload()` and the LiveView lifecycle all landed on 2026-08-23 with the number unchanged). Before relying on any "planned" item being still absent — or on a workaround still being necessary — read `TODO.md` and `README.md` at https://github.com/m1ga/ti.game, the module's upstream repo. `README.md` there is the canonical API documentation; the `documentation/` folder only points at it. The priorities quoted above are the maintainer's, not a roadmap this skill decides.