@genex-ai/cli-demo 0.71.0 → 0.74.0-dev.190

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (42) hide show
  1. package/dist/index.js +354 -8
  2. package/package.json +2 -1
  3. package/templates/controllers/character/follow-camera.ts +15 -4
  4. package/templates/controllers/character/vrm/vrm-loader.ts +74 -11
  5. package/templates/controllers/quality/governor.ts +147 -0
  6. package/templates/controllers/quality/pick-asset.ts +57 -0
  7. package/templates/controllers/quality/tier.ts +170 -0
  8. package/templates/skills/genex-ai-hud/SKILL.md +11 -2
  9. package/templates/skills/genex-ai-menu/SKILL.md +18 -2
  10. package/templates/skills/genex-ai-skybox/SKILL.md +15 -4
  11. package/templates/skills/genex-ai-texture/SKILL.md +1 -1
  12. package/templates/skills/genex-ai-video/SKILL.md +1 -1
  13. package/templates/skills/genex-explore/SKILL.md +1 -1
  14. package/templates/skills/genex-getting-started/SKILL.md +2 -2
  15. package/templates/skills/genex-threejs-adaptive-quality/SKILL.md +141 -0
  16. package/templates/skills/genex-threejs-adaptive-quality/references/adaptive-quality.md +105 -0
  17. package/templates/skills/genex-threejs-bloom/SKILL.md +4 -1
  18. package/templates/skills/genex-threejs-bloom/references/bloom.md +1 -1
  19. package/templates/skills/genex-threejs-camera-direction/SKILL.md +62 -12
  20. package/templates/skills/genex-threejs-camera-direction/references/camera-rigs.md +62 -0
  21. package/templates/skills/genex-threejs-character-controller/references/wiring.md +9 -3
  22. package/templates/skills/genex-threejs-embed-auth/SKILL.md +4 -1
  23. package/templates/skills/genex-threejs-game-feel/SKILL.md +4 -1
  24. package/templates/skills/genex-threejs-game-ui/SKILL.md +113 -30
  25. package/templates/skills/genex-threejs-game-ui/references/style-capsules.md +4 -1
  26. package/templates/skills/genex-threejs-image-pipeline/SKILL.md +5 -0
  27. package/templates/skills/genex-threejs-image-pipeline/references/image-pipeline.md +1 -1
  28. package/templates/skills/genex-threejs-lighting-design/SKILL.md +5 -1
  29. package/templates/skills/genex-threejs-multiplayer/SKILL.md +7 -1
  30. package/templates/skills/genex-threejs-multiplayer/references/host-physics.md +6 -3
  31. package/templates/skills/genex-threejs-physics-rapier/references/colliders-from-assets.md +1 -0
  32. package/templates/skills/genex-threejs-screen-space-ambient-occlusion/references/ambient-occlusion.md +1 -1
  33. package/templates/skills/genex-threejs-shadow-systems/SKILL.md +6 -0
  34. package/templates/skills/genex-threejs-shadow-systems/references/shadow-systems.md +1 -1
  35. package/templates/skills/genex-threejs-skill-router/SKILL.md +26 -5
  36. package/templates/skills/genex-threejs-skill-router/references/routing-map.md +25 -9
  37. package/templates/skills/genex-threejs-spectral-ocean/references/spectral-ocean.md +1 -1
  38. package/templates/skills/genex-threejs-touch-controls/SKILL.md +11 -0
  39. package/templates/skills/genex-threejs-vehicle-controllers/references/enter-exit.md +7 -0
  40. package/templates/skills/genex-threejs-visual-validation/SKILL.md +36 -12
  41. package/templates/skills/genex-threejs-water-optics/references/water-optics.md +1 -1
  42. package/templates/skills/genex-updates/SKILL.md +1 -1
@@ -0,0 +1,141 @@
1
+ ---
2
+ name: genex-threejs-adaptive-quality
3
+ description: Make a Genex Three.js game phone-survivable with the vendored adaptive-quality kit — device-tier detection, tier-budgeted renderer settings, a runtime governor that steps quality down before phones run out of memory, per-tier asset rungs for generated skyboxes/textures, and a Quality picker in settings. Load for every game at boot wiring time, and whenever a game is heavy, crashes on phones, or gets flagged desktop-only.
4
+ ---
5
+
6
+ # Genex Three.js Adaptive Quality
7
+
8
+ Phones enforce a hard GPU-memory ceiling desktops don't have: iOS silently
9
+ kills the page when a game allocates too much, and the kill arrives at BOOT —
10
+ exactly when a skybox, models, and the post stack all decode at once. This
11
+ skill wires the vendored quality kit so the game boots conservatively on
12
+ phones, steps quality UP when the device proves smooth, and never gets uglier
13
+ on desktop. **You can recover from ugly; you cannot recover from a killed
14
+ page.**
15
+
16
+ This is a completion gate like the post stack: every game wires the tier at
17
+ boot before it is called done. It costs three lines, not a testing burden —
18
+ you still verify on desktop only.
19
+
20
+ ## Install
21
+
22
+ ```bash
23
+ npx genex controller quality
24
+ ```
25
+
26
+ Installs `src/controllers/quality/{tier.ts, governor.ts, pick-asset.ts}` —
27
+ game-owned code, edit freely.
28
+
29
+ ## Wire the tier at boot (before renderer construction)
30
+
31
+ ```ts
32
+ import { detectTier } from "./controllers/quality/tier.ts";
33
+ import { QualityGovernor } from "./controllers/quality/governor.ts";
34
+
35
+ const tier = detectTier(); // phone-low | phone | desktop; manual Quality setting wins
36
+ const renderer = new THREE.WebGLRenderer({ antialias: tier.antialias });
37
+ renderer.setPixelRatio(Math.min(window.devicePixelRatio, tier.dprCap));
38
+ ```
39
+
40
+ The tier owns every budget decision: `dprCap` (1.5 on phones — the single
41
+ biggest framebuffer lever), `antialias` (off on phones; it is fixed at context
42
+ creation and can never change live), `shadowMapSize` (1024 phone / 2048
43
+ desktop), `postLevel` (`'light'` = tone mapping + cheap passes on phones;
44
+ `'full'` = the named stack on desktop), `particleScale`, `drawDistanceScale`,
45
+ `frameCap`, and `remoteAvatarCap` for multiplayer. The exact ladder and which
46
+ knobs may change at runtime vs load time vs never: [references/adaptive-quality.md](references/adaptive-quality.md).
47
+
48
+ ## Wire the governor (runtime — phones throttle over minutes)
49
+
50
+ ```ts
51
+ const governor = new QualityGovernor(tier, {
52
+ setDprScale: (m) => renderer.setPixelRatio(Math.min(window.devicePixelRatio, tier.dprCap * m)),
53
+ setPostEnabled: (on) => (composerEnabled = on),
54
+ setDrawDistanceScale: (m) => (scene.fog!.far = baseFogFar * tier.drawDistanceScale * m),
55
+ }, renderer);
56
+
57
+ // In the render loop, with the same performance.now() delta the loop already computes:
58
+ governor.frame(deltaMs);
59
+ ```
60
+
61
+ Sustained slow frames step down (DPR ×0.8 → post off → draw distance →
62
+ 30 fps cap); twenty smooth seconds step back up; a knob that failed twice
63
+ stays down for the session. It keeps governing forever — thermal throttling
64
+ arrives at minute eight, not second thirty. It also pauses judgment when the
65
+ tab is hidden and publishes memory counts the platform's crash telemetry
66
+ reads; pause your own loop and audio on `visibilitychange` too.
67
+
68
+ ## Generated assets: load through the rungs
69
+
70
+ Generated skyboxes are 8192×4096 — about 178 MB decoded, over half a phone's
71
+ whole budget in one texture. Every generated image asset ships with downscale
72
+ rungs; phones must load through them:
73
+
74
+ ```ts
75
+ import { detectTier } from "./controllers/quality/tier.ts";
76
+ import { loadTextureWithFallback } from "./controllers/quality/pick-asset.ts";
77
+
78
+ const texture = await loadTextureWithFallback(SKYBOX_URL, tier, (u) =>
79
+ new THREE.TextureLoader().loadAsync(u),
80
+ );
81
+ ```
82
+
83
+ Desktop loads the original, phones the `@2048` rung (~11 MB), and a missing
84
+ rung falls back to the original — never a broken boot. The `$genex-ai-skybox`
85
+ and `$genex-ai-texture` skills show the wiring in place.
86
+
87
+ ## Quality picker in settings
88
+
89
+ The pause/settings screen (see `$genex-threejs-game-ui`) always carries a
90
+ Quality entry: **Auto / Low / Medium / High**, wired to
91
+ `setQualitySetting(...)` from `tier.ts` + a reload or re-tier. Persisted
92
+ per-device in localStorage on purpose — quality is a property of the phone,
93
+ not the player's account. Default Auto.
94
+
95
+ ## Budgets that ride the tier (not separate rules)
96
+
97
+ - Shadows: `tier.shadowMapSize`, `shadow.autoUpdate = false` for static scenes,
98
+ at most 2 cascades on phones (`$genex-threejs-shadow-systems`).
99
+ - Post: phone floor is a BUILT tone-mapping/output pass (`postLevel: 'light'`
100
+ adds FXAA/vignette); SSAO, volumetrics, and DoF are desktop-tier only
101
+ (`$genex-threejs-skill-router` owns the floor wording).
102
+ - Particles/scatter: multiply counts by `tier.particleScale`; render heavy
103
+ transparency at half resolution and upsample.
104
+ - Animation: distant mixers update at 1/2–1/4 rate; multiplayer remotes above
105
+ `tier.remoteAvatarCap` billboard instead of animating
106
+ (`$genex-threejs-multiplayer`).
107
+ - Physics: phone tiers prefer hull/cuboid colliders for props — trimesh only
108
+ where gameplay demands it (`$genex-threejs-physics-rapier`).
109
+ - Shaders: default `mediump` (iOS exception: float-texture sampling needs
110
+ `highp sampler2D`); precompile with `renderer.compileAsync(scene, camera)`
111
+ during the loader screen so first-frame jank doesn't read as a stall; skip
112
+ max anisotropy on phones.
113
+ - Disposal: level swaps traverse the outgoing scene and dispose geometry,
114
+ materials, AND textures (three never frees them for you); watch
115
+ `renderer.info.memory` while testing — if textures/geometries climb across
116
+ swaps, you leak toward the kill.
117
+
118
+ ## WebGPU games
119
+
120
+ The scaffold ships WebGL and stays the default; if this project already uses
121
+ `WebGPURenderer`, keep it (never switch renderers mid-project). Context loss
122
+ differs: WebGL fires `webglcontextlost` events; WebGPU exposes a
123
+ `device.lost` promise — attach a handler that pauses the loop and rebuilds,
124
+ mirroring the scaffold's WebGL pattern. All tier knobs apply identically
125
+ except `antialias` (WebGPU MSAA is per-render-target and CAN change at
126
+ runtime).
127
+
128
+ ## Failure conditions
129
+
130
+ - Renderer constructed before `detectTier()` → context-creation knobs
131
+ (antialias) are locked wrong for the session. Tier first, renderer second.
132
+ - Governor wired to context-creation flags → no-op at best. Runtime knobs
133
+ only: DPR, post toggles, distances, frame cap.
134
+ - Skybox loaded with a bare `TextureLoader.loadAsync(SKYBOX_URL)` on a game
135
+ that targets phones → ~178 MB decoded; route it through
136
+ `loadTextureWithFallback`.
137
+ - Quality stepping on every spike → shader compiles read as slowness. The
138
+ governor requires SUSTAINED slow windows; do not shorten them.
139
+ - Testing quality tiers by resizing the desktop window → tiers key off touch +
140
+ OS, not viewport. Trust desktop verification plus the preflight report
141
+ `genex preview` prints.
@@ -0,0 +1,105 @@
1
+ # Genex adaptive quality — tiers, knobs, and the governor in depth
2
+
3
+ Use this reference when tuning the tier ladder, deciding which knob may change
4
+ when, or teaching the governor a game-specific step.
5
+
6
+ ## Why boot-conservative
7
+
8
+ The phone budget is a hard ceiling that includes GPU memory (textures,
9
+ framebuffers), and the OS kill arrives with no catchable event. Boot is the
10
+ danger window: skybox + models + post targets decode together. So phone tiers
11
+ START one notch below what the heuristics suggest and the governor steps UP
12
+ after ~20 smooth seconds. The cost of guessing low is moments of softness; the
13
+ cost of guessing high is a dead page.
14
+
15
+ ## The tier ladder
16
+
17
+ | Knob | phone-low | phone | desktop |
18
+ |---|---|---|---|
19
+ | DPR cap | 1.0 | 1.5 | 2 |
20
+ | antialias (context) | off | off | on |
21
+ | Shadow map | 512 (static-cached) | 1024 | 2048 |
22
+ | Post level | tone map only | + FXAA/vignette | full named stack |
23
+ | Skybox rung | @2048 (~11 MB) | @4096 (~45 MB) | original |
24
+ | Texture rung (props) | @1024 | @2048 | original |
25
+ | Particles/scatter | 0.25× | 0.5× | 1× |
26
+ | Draw distance | 0.5× | 0.75× | 1× |
27
+ | Frame target | stable 30 | 60 | 60 |
28
+ | Remote avatars animated | 4 | 8 | all |
29
+ | Prop colliders | hull/cuboid | hull | as designed |
30
+
31
+ A DPR drop from 3 (raw iPhone) to 1.5 cuts every full-screen surface — color,
32
+ depth, and each post target — to a quarter of the bytes. It is the single
33
+ strongest lever the tier owns.
34
+
35
+ ## The knob split — what may change when
36
+
37
+ Getting this wrong produces silent no-ops or a session stuck ugly:
38
+
39
+ - **Context-creation-fixed (never changes live):** `antialias`, `alpha`,
40
+ `stencil`, `powerPreference` on WebGL. Changing them means a new context and
41
+ a full re-init — the tier must decide them BEFORE the renderer exists.
42
+ (WebGPU differs: MSAA is per-render-target sample count and is runtime-
43
+ changeable.)
44
+ - **Load-time (fixed for the session once fetched):** asset rungs (skybox and
45
+ texture resolutions), model LOD sets. `pickAsset` decides them from the tier
46
+ at load; switching later means a re-fetch — treat as fixed.
47
+ - **Runtime-free (the governor's domain):** `setPixelRatio`, post passes on/off
48
+ and their target resolutions, shadow map size (realloc), draw distance and
49
+ fog, LOD bias, particle counts, mixer update rates, frame cap, remote-avatar
50
+ animation count.
51
+
52
+ ## Governor mechanics
53
+
54
+ - Slow = frame delta over budget for a SUSTAINED window (4 s) — never single
55
+ spikes, which are usually shader compiles or GC. Precompiling with
56
+ `renderer.compileAsync` during the loader screen removes most spikes at the
57
+ source (and keeps first-frame jank from reading as a stall to the platform's
58
+ telemetry).
59
+ - Step-down order: DPR ×0.8 → post off → draw distance ×0.6 → 30 fps cap.
60
+ Each step is the cheapest remaining lever with the biggest headroom return.
61
+ - Step-up needs 20 smooth seconds (hysteresis), and a step that had to be
62
+ re-applied twice is pinned for the session — oscillating quality reads worse
63
+ than stable-low.
64
+ - The governor never stops: thermal throttling degrades phones after minutes
65
+ of play, so a boot-time benchmark alone always ends up wrong.
66
+ - A stable 30 fps cap beats a stuttery 40–50: consistent frame pacing reads
67
+ smoother and halves GPU work per second (heat, battery, memory bandwidth).
68
+ - Backgrounded tab (`visibilitychange`): pause the render loop and audio, not
69
+ just the governor — a hidden game burning GPU is pure thermal debt on the
70
+ device class that can least afford it.
71
+
72
+ ## Detection honesty
73
+
74
+ - Apple devices mask the GPU renderer string ("Apple GPU") — screen dims + DPR
75
+ + iOS major version are the usable signals there, and the governor corrects
76
+ the rest from measured frames.
77
+ - Android exposes real renderer strings (Adreno/Mali/Xclipse); the vendored
78
+ lookup in `tier.ts` promotes strong GPUs to the `phone` tier. It is a
79
+ heuristic on purpose — extend the regex when field data shows a
80
+ misclassified family, and let the governor absorb the rest.
81
+ - Never burn a probe context on a memory-strapped phone at play time; the one
82
+ probe in `tier.ts` runs at boot and frees its context immediately.
83
+
84
+ ## Memory discipline that rides the tier
85
+
86
+ - Dispose on every level swap: traverse the outgoing scene and call
87
+ `.dispose()` on geometry, material, AND each material's textures — material
88
+ dispose does not free textures, and three frees nothing automatically.
89
+ - Watch `renderer.info.memory.{textures,geometries}` across swaps in dev; a
90
+ monotonic climb is a leak marching toward the OS kill. The governor
91
+ publishes these counts for the platform's field telemetry.
92
+ - Prefer meshopt/instanced geometry for repeats; `BatchedMesh` batches
93
+ HETEROGENEOUS static meshes into one draw where instancing (identical
94
+ meshes only) can't.
95
+ - Half-resolution transparency: render heavy particle/transparency passes to a
96
+ half-size target and composite up — fill-rate is the phone bottleneck.
97
+
98
+ ## Multiplayer at tier
99
+
100
+ Remotes are visual-only; with the shared avatar file, `loadVrmClone` gives N
101
+ remotes one set of GPU geometry/textures. Animate and fully draw only the
102
+ nearest `tier.remoteAvatarCap`; beyond it, freeze the mixer and billboard or
103
+ hide. Matchmade games can also declare a lower `maxPlayers` in
104
+ `genex.matchmaking` for phone-heavy audiences — capacity is a server-owned
105
+ knob.
@@ -32,7 +32,10 @@ reference before adding selective bloom to a composed scene.
32
32
  - selective bloom requires mutating scene materials every frame without restoration guarantees;
33
33
  - transparent particles disappear from extraction because pass ownership is unclear;
34
34
  - bloom radius changes wildly with resolution;
35
- - highlights become gray because energy is clamped too early.
35
+ - highlights become gray because energy is clamped too early;
36
+ - bloom ships un-tiered: phone tiers run the light post level
37
+ (`$genex-threejs-adaptive-quality`) — bloom is a desktop-tier pass, and its
38
+ full-res HDR target is exactly the allocation phones get killed for.
36
39
 
37
40
  ## Routing boundary
38
41
 
@@ -2,7 +2,7 @@
2
2
 
3
3
  Use this reference to choose bloom ownership, signal order, selective contribution, and scene-relative emissive ranges without making bloom responsible for the underlying form.
4
4
 
5
- > **Renderer note:** this reference assumes `WebGPURenderer` + TSL node materials. Check the project's actual renderer first — the Genex scaffold ships vanilla WebGL three.js. On WebGL, adapt the technique with standard materials / `EffectComposer` passes or pick a simpler alternative; never switch renderers mid-project.
5
+ > **Renderer note:** this reference assumes `WebGPURenderer` + TSL node materials. Check the project's actual renderer first — the Genex scaffold ships vanilla WebGL three.js. On WebGL, adapt the technique with standard materials / `EffectComposer` passes or pick a simpler alternative; never switch renderers mid-project. Either way the effect obeys the device tier (`$genex-threejs-adaptive-quality`): expensive passes are desktop-tier, and on WebGPU the per-target MSAA sample count is a runtime knob the governor may drive.
6
6
 
7
7
  ## Contents
8
8
 
@@ -35,9 +35,12 @@ rules, floating-origin shot, pointer controls, and implementation limits.
35
35
 
36
36
  ## Aiming and pointer lock
37
37
 
38
- On the bundled `FollowCamera`, **pointer-lock aim is ON by default** on desktop:
39
- a click locks the pointer and raw mouse movement drives the view. You rarely turn
40
- it on you decide whether to turn it OFF. Name the bucket in the build plan:
38
+ The rule is binary: during play, **the cursor is either a gameplay tool or it is
39
+ locked away**. An OS arrow parked over the action in a game that never uses it is
40
+ a shipped defect, not a default. On the bundled `FollowCamera`, pointer-lock aim
41
+ is ON by default on desktop — a click locks the pointer and raw mouse movement
42
+ drives the view — so mostly you decide whether to turn it OFF. Name the bucket in
43
+ the build plan:
41
44
 
42
45
  - **MANDATORY** — first-person of any kind (FPS, walking sim, horror) and any
43
46
  mouse-aimed action (third-person shooter, turret/range). Lock is on by default;
@@ -47,11 +50,16 @@ it on — you decide whether to turn it OFF. Name the bucket in the build plan:
47
50
  `genex controller character` game). On by default; leave it on. Opt out with
48
51
  `pointerLockAim: false` only for a stated reason (a cursor-heavy UI at the core
49
52
  of play).
50
- - **NEVER** — cursor-core games (top-down click-to-move, tower defense, builders,
51
- card/board/puzzle) and spectator/orbit showcases. These **must pass
52
- `pointerLockAim: false`** otherwise the bundled camera grabs the cursor on the
53
- first click. (Touch needs nothing: pointer lock doesn't exist there and the mode
54
- no-ops on coarse pointers.)
53
+ - **Keyboard-driven games lock too** — a racer, platformer, or runner that never
54
+ reads the mouse still locks the pointer on the play/Start click: the cursor is
55
+ not a tool there, so lock it away (hidden cursor, no stray clicks, Esc = pause
56
+ as usual). Bundled-controller games get this for free; a hand-rolled game uses
57
+ the minimal lock recipe in `$genex-threejs-game-ui` (~6 lines).
58
+ - **NEVER** — cursor-core games where the pointer IS the gameplay tool (top-down
59
+ click-to-move, tower defense, builders, card/board/puzzle) and spectator/orbit
60
+ showcases. These **must pass `pointerLockAim: false`** — otherwise the bundled
61
+ camera grabs the cursor on the first click. (Touch needs nothing: pointer lock
62
+ doesn't exist there and the mode no-ops on coarse pointers.)
55
63
 
56
64
  **Mechanism — games on the bundled controller (most games):** do NOT hand-roll
57
65
  lock handling. Aim is already enabled; the kit ships a ready-made cue overlay —
@@ -65,6 +73,7 @@ character`) to `onAimChange` and you get the reticle + "click to aim" cue for fr
65
73
  domElement: renderer.domElement,
66
74
  onAimChange: cue.onAimChange,
67
75
  });
76
+ followCam.setPaused(getPhase() !== "playing"); // park aim if the game boots on a menu/loader
68
77
 
69
78
  Cursor-core game? Pass `pointerLockAim: false` instead and skip the cue. Want a
70
79
  custom HUD? Read `onAimChange` yourself — `{ state }` is one of `locked` /
@@ -86,12 +95,17 @@ grants the permission.
86
95
 
87
96
  1. Two states only: locked = playing, unlocked = menu/paused. "Unlocked but
88
97
  gameplay continues" is the imprecise-aim defect in disguise.
89
- 2. Opening any menu (inventory/shop/dialog): exit the lock
98
+ 2. Opening any menu — the BOOT/main menu included — parks aim: exit the lock
90
99
  (`followCam.setPaused(true)`), cursor returns, gameplay input pauses. Menu
91
- keys are Tab/I/E — never Esc.
100
+ keys are Tab/I/E — never Esc. The one true wiring is the phase binding
101
+ `followCam.setPaused(phase !== "playing")` applied on phase transitions,
102
+ never per frame: an unguarded per-frame resume once re-requested the lock
103
+ every frame, and any menu click's ~5s of transient activation then locked
104
+ the pointer over the open menu.
92
105
  3. Closing a menu re-locks INSIDE the close click/keypress handler
93
- (`followCam.setPaused(false)`) — the browser requires a user gesture, so
94
- menus close by click/keypress, never by timeout.
106
+ (`followCam.setPaused(false)` — the phase binding's `setPhase("playing")`
107
+ in the Play/Resume click does exactly this) — the browser requires a user
108
+ gesture, so menus close by click/keypress, never by timeout.
95
109
  4. Esc is the browser's release valve (you can't intercept it; Chrome enforces
96
110
  a re-lock cooldown) → treat Esc as pause: show the overlay with a
97
111
  "click to resume" button.
@@ -108,11 +122,45 @@ grants the permission.
108
122
  Chrome's post-Esc cooldown) fires `onAimChange` with reason `needs-gesture` and
109
123
  keeps the "click to aim" cue up; the next click or keypress re-locks. Never
110
124
  retry on a timer.
125
+ 10. The bundled camera OWNS the lock: never call `document.exitPointerLock()` or
126
+ `canvas.requestPointerLock()` yourself alongside it — raw calls desync
127
+ `aimState` and the cue. Everything routes through `setPaused` (which is
128
+ idempotent — a same-value call is a no-op).
111
129
 
112
130
  **Not an OS setting:** pointer `movementX/movementY` is never inverted by the OS
113
131
  (trackpad "natural scrolling" only flips the wheel). If look feels inverted it's a
114
132
  sign bug in the rig, not a device quirk — fix the sign, don't sniff the trackpad.
115
133
 
134
+ ## The screen-direction contract
135
+
136
+ Every input axis has one correct on-screen direction, for every rig, bundled or
137
+ hand-rolled. These four invariants are testable and non-negotiable:
138
+
139
+ 1. Mouse/touchpad RIGHT turns the view right; mouse UP looks up (down only behind
140
+ an explicit invert option the player chose).
141
+ 2. `KeyD`/ArrowRight moves or turns the player toward screen-RIGHT; `KeyA`/
142
+ ArrowLeft toward screen-left. Same for a touch stick's +x.
143
+ 3. Drag-pan picks ONE convention — grab-the-world (terrain follows the pointer)
144
+ or move-the-camera — and BOTH axes obey it. One axis each is the
145
+ "diagonals feel twisted" bug.
146
+ 4. See the OS-setting note above: inversion is always your sign, never the device.
147
+
148
+ The formula that settles every sign argument: `screenRight = cross(cameraForward, worldUp)`.
149
+ For a Y-up world and forward `(sin yaw, 0, cos yaw)`, screen-right is
150
+ `(-cos yaw, 0, sin yaw)`. **Warning — `(cos yaw, 0, -sin yaw)` is the LEFT
151
+ vector** (that's `cross(worldUp, cameraForward)`), and writing it as "right" is
152
+ the single most-shipped direction bug in generated games: two independent
153
+ projects inverted their A/D exactly this way. Related trap: positive
154
+ `rotation.y` turns a +Z-facing object toward +X, which is screen-LEFT from a
155
+ chase camera behind it — so "positive yaw = turn right" is false in this basis.
156
+
157
+ Never derive signs by intuition — intuition about right-handed frames is wrong
158
+ about half the time and has been wrong in every shipped instance. Copy a
159
+ verified pair (sign AND basis together) from
160
+ [references/camera-rigs.md](references/camera-rigs.md), then confirm with the
161
+ input-direction part of the smoke check: hold D and watch which way the world
162
+ answers.
163
+
116
164
  ## Non-negotiable rules
117
165
 
118
166
  - Use subject dimensions to derive offsets; do not tune one fixed distance for
@@ -122,6 +170,8 @@ sign bug in the rig, not a device quirk — fix the sign, don't sniff the trackp
122
170
  - During an explicit handoff, use one interpolation stage. Do not stack a
123
171
  transition blend and a second follow smoother over the same interval.
124
172
  - Re-sync yaw/pitch from the camera when pointer lock is acquired.
173
+ - Hand-rolled steering/pan/look math copies a verified basis from the reference
174
+ and passes the input-direction check — signs are never derived by intuition.
125
175
  - Update the projection matrix whenever FOV, near, far, or aspect changes.
126
176
  - Keep stars or infinite backgrounds camera-relative when large translation
127
177
  would create false parallax or precision loss.
@@ -11,6 +11,7 @@ Use this reference for scale-aware chase, side, orbit, authored-shot, pointer-lo
11
11
  - Explicit camera handoffs
12
12
  - cinematic implementation shot ownership
13
13
  - Pointer-look and movement constraints
14
+ - Verified screen-direction bases
14
15
  - Floating origin and background handling
15
16
  - Projection and lifecycle ownership
16
17
  - Failure modes and diagnostics
@@ -166,6 +167,9 @@ yaw -= mouseDeltaX * 0.0022
166
167
  pitch -= mouseDeltaY * 0.0018
167
168
  ```
168
169
 
170
+ Expected on screen: mouse-right orbits the view right, mouse-up tilts it up —
171
+ verify both axes against the screen-direction contract before retuning the scales.
172
+
169
173
  Pitch bounds vary by flight mode. The implementation also enforces camera height above
170
174
  the ship:
171
175
 
@@ -281,6 +285,9 @@ distance = movementSpeed * dt
281
285
 
282
286
  Default speed is `9`, sensitivity `0.0023`.
283
287
 
288
+ Expected on screen: mouse-right turns the view right, mouse-up looks up — assert
289
+ both axes (yaw-only evidence has let inverted pitch ship).
290
+
284
291
  Keys are cleared on:
285
292
 
286
293
  - pointer-lock exit;
@@ -295,6 +302,61 @@ Scene-specific constraints then run after controls:
295
302
 
296
303
  Input control and spatial constraint are separate layers.
297
304
 
305
+ ## Verified screen-direction bases
306
+
307
+ Copy these pairs whole — the sign and the basis are only correct TOGETHER. Each
308
+ was derived from `screenRight = cross(cameraForward, worldUp)` and verified
309
+ against the on-screen result; if you change one half, re-verify with the
310
+ input-direction check instead of reasoning about it.
311
+
312
+ **Chase-cam steering** (vehicle/character heading, camera behind):
313
+
314
+ ```text
315
+ heading = (sin yaw, 0, cos yaw) // matches rotation.y for a +Z-front model
316
+ yaw -= steer * rate * dt // steer: D/right = +1, A/left = -1
317
+ position += heading * speed * dt
318
+ camera at position - heading * dist, lookAt(position)
319
+ ```
320
+
321
+ Why the minus: the camera looks along `heading`, so screen-right is
322
+ `cross(heading, up) = (-cos yaw, 0, sin yaw)`, while `d(heading)/d(yaw) =
323
+ (cos yaw, 0, -sin yaw)` — exactly screen-LEFT. Increasing yaw always veers the
324
+ nose left on screen, so "D turns right" needs `yaw -=`. (Equivalently
325
+ `yaw += steer` is correct only with heading `(-sin yaw, 0, cos yaw)` — a pair,
326
+ never a lone sign.)
327
+
328
+ **RTS / overhead pan camera** (fixed pitch, yaw-orbiting):
329
+
330
+ ```text
331
+ right = (-cos yaw, 0, sin yaw) // pitch-independent screen-right on the ground
332
+ forwardGround = (sin yaw, 0, cos yaw) // into the screen along the ground
333
+ D / ArrowRight: target += right * pan A / ArrowLeft: target -= right * pan
334
+ W / ArrowUp: target += forwardGround * pan S: target -= forwardGround * pan
335
+ ```
336
+
337
+ Drag-pan, grab-the-world (terrain follows the pointer; both axes, one
338
+ convention — `movementY` is positive DOWNWARD):
339
+
340
+ ```text
341
+ target -= right * movementX * k
342
+ target += forwardGround * movementY * k
343
+ ```
344
+
345
+ Move-the-camera convention = flip BOTH signs, never one. The classic shipped bug
346
+ writes `right = (cos yaw, 0, -sin yaw)` — that is `cross(up, forward)`, the LEFT
347
+ vector — inverting A/D and the horizontal drag while W/S stay correct.
348
+
349
+ **Pointer-look** (locked mouse driving yaw/pitch, Euler order `YXZ`):
350
+
351
+ ```text
352
+ yaw -= movementX * sensitivity // mouse-right -> view turns RIGHT
353
+ pitch -= movementY * sensitivity // mouse-up -> view tilts UP
354
+ ```
355
+
356
+ Expected on screen, both axes, before any tuning: a straight-ahead landmark
357
+ slides LEFT when the mouse moves right (the view pans right) and slides DOWN
358
+ when the mouse moves up (the view tilts up).
359
+
298
360
  ## Floating origin and background handling
299
361
 
300
362
  The Saturn scene first computes a virtual camera pose, stores its orientation
@@ -149,9 +149,15 @@ Want a custom HUD? Read `onAimChange: ({ state }) => …` yourself instead of th
149
149
  - **Opt out:** cursor-core games (RTS, tower defense, card, builder) pass
150
150
  `pointerLockAim: false` — the camera then stays drag-orbit and never grabs the
151
151
  cursor.
152
- - **Menus / vehicles:** `followCam.setPaused(true)` when a menu opens or the
153
- player starts driving; `followCam.setPaused(false)` INSIDE the closing
154
- click/keypress handler (re-lock needs a user gesture). Aim is on-foot only.
152
+ - **Menus / vehicles:** the BOOT/main menu counts as a menu park aim from the
153
+ first frame. The one true wiring is the phase binding
154
+ `followCam.setPaused(phase !== "playing")` applied on phase TRANSITIONS
155
+ (inside `setPhase()`), never driven from the render loop: `setPaused` is
156
+ idempotent, but a per-frame resume once locked the pointer over an open menu
157
+ the moment a menu click granted transient activation. The Play/Resume click's
158
+ transition doubles as the required user gesture for the re-lock. Vehicles:
159
+ combine conditions (`phase !== "playing" || activeId !== CHARACTER_ID`) — aim
160
+ is on-foot only.
155
161
  - **First-person:** the same mode plus `minDistance`/`maxDistance` ≈ 0.1, an
156
162
  eye-height target fed to `moveTo`, `avatar.visible = false`, and
157
163
  `controller.setLockForward(true)`.
@@ -162,7 +162,10 @@ From `@genex-ai/embed-sdk/sentry` (crash reporting; exactly these two):
162
162
  - `sentryCanvasSnapshot(canvas)` — session replay records the DOM, not the 3D
163
163
  canvas; call this once per frame at the END of the render loop so replays
164
164
  show actual gameplay. Works for BOTH WebGL and WebGPU renderers; internally
165
- throttled, so calling at 60fps is fine:
165
+ throttled, so calling at 60fps is fine. On TOUCH devices it is a deliberate
166
+ no-op (and session replay/tracing sample down): each capture is a full-canvas
167
+ GPU readback, exactly the overhead phones get memory-killed for — mobile
168
+ replays are DOM-only by design, on-error replays still record everywhere:
166
169
 
167
170
  ```ts
168
171
  function animate() {
@@ -37,7 +37,10 @@ no matter how good it looks.
37
37
  unlocked drag-to-turn camera feels imprecise no matter how tight the numbers
38
38
  are. The bucket rule + the bundled `FollowCamera` aim mode live in
39
39
  `$genex-threejs-camera-direction`; on the bundled controller it's ON by default
40
- (with a ready-made cue), not hand-rolled events.
40
+ (with a ready-made cue), not hand-rolled events. Direction is half of it: a
41
+ movement key or look axis whose on-screen direction contradicts its label is a
42
+ defect, not a tuning issue — the screen-direction contract and verified bases
43
+ are in `$genex-threejs-camera-direction`.
41
44
 
42
45
  ## Movement: snappy beats realistic
43
46