From bc2821adb11a30244fc4663f4fafb756877d9508 Mon Sep 17 00:00:00 2001
From: Yasutake Yohei <61961825+yasutakeyohei@users.noreply.github.com>
Date: Wed, 7 Oct 2026 23:09:40 +0900
Subject: bluebey-studio: public/ の外へ移動し非公開化
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 bluebey-studio/src/main.js | 3056 ++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 3056 insertions(+)
 create mode 100644 bluebey-studio/src/main.js

(limited to 'bluebey-studio/src/main.js')

diff --git a/bluebey-studio/src/main.js b/bluebey-studio/src/main.js
new file mode 100644
index 0000000..20c6c7d
--- /dev/null
+++ b/bluebey-studio/src/main.js
@@ -0,0 +1,3056 @@
+import * as THREE from 'three';
+import { OrbitControls } from 'three/addons/controls/OrbitControls.js';
+import { GLTFLoader } from 'three/addons/loaders/GLTFLoader.js';
+import { loadModel } from './model.js';
+import { Face } from './face.js';
+import { Styles, isLineStyle } from './styles.js';
+import { Rig } from './rig.js';
+import { Animator, createRecorder, startRecording, stopRecording } from './animation.js';
+import { defaultState, applyPatch, THEMES } from './presets.js';
+import { buildPanel } from './panel.js';
+import * as exporter from './exporter.js';
+import { toast } from './ui.js';
+import { History } from './history.js';
+import { createLook } from './look.js';
+import { createClipper } from './clip.js';
+import { buildHat } from './hats.js';
+import { createBackdrop, backdropStyle } from './background.js';
+import { drawCaption, layoutCaption } from './caption.js';
+import { loadFont, captionFontStack } from './textOutlines.js';
+import { buildProp, PROP_DEFAULTS, disposeProp, applyPropText } from './props.js';
+import { rollAll, decodeSeed } from './gacha.js';
+import { encodeState, decodeState } from './urlState.js';
+import { createMouthFlap, levelToMouth } from './mouthFlap.js';
+import { createScreenOutline, BEHIND_LABEL, SOLID_LABEL, LEAF_LABEL, NOSE_LABEL } from './outline.js';
+import { gionSheetUrl, stampRect, imageAspect, imageReady, drawGion, DEFAULT_STAMP_WIDTH } from './gion.js';
+
+const DEG = Math.PI / 180;
+/** The other direction: what three's Spherical reports is radians. */
+const RAD_TO_DEG = 180 / Math.PI;
+const MODEL_URL = 'assets/bluebey.glb';
+
+/** The caption font is only fetched when a bubble is first used. */
+const CAPTION_FONT_TTF = 'assets/fonts/bluebey-caption.ttf';
+const CAPTION_FONT_WOFF2 = 'assets/fonts/bluebey-caption.woff2';
+/** How far the eyes can swing for a look-at target, in radians. */
+const MAX_LOOK_YAW = 0.5;
+const MAX_LOOK_PITCH = 0.34;
+
+/** Undo steps are labelled by the part of the state that changed. */
+const SCOPE_LABELS = {
+  all: '変更', render: '見た目', face: '表情', view: 'カメラ・背景', pose: 'ポーズ', caption: 'セリフ', gion: '擬音',
+};
+
+/**
+ * The single-file build inlines the fonts and the backdrop library, so the same
+ * code reads a data URL there and a file URL in the normal build.
+ */
+function assetUrl(file, kind) {
+  const inlined = globalThis.__BLUEBEY_FONTS__ ?? {};
+  if (kind === 'background') {
+    const backdrops = globalThis.__BLUEBEY_BACKGROUNDS__ ?? {};
+    return backdrops[file] ?? `assets/backgrounds/${file}.webp`;
+  }
+  if (kind === 'beard') {
+    const beards = globalThis.__BLUEBEY_BEARDS__ ?? {};
+    return beards[file] ?? `assets/beards/${file}`;
+  }
+  if (kind === 'brow') {
+    const brows = globalThis.__BLUEBEY_BROWS__ ?? {};
+    return brows[file] ?? `assets/brows/${file}`;
+  }
+  return inlined[file] ?? `assets/fonts/${file}`;
+}
+
+/**
+ * Where the model comes from. The normal build fetches `assets/bluebey.glb`;
+ * the single-file build (tools/build-standalone.mjs) inlines the same bytes as
+ * base64 in `window.__BLUEBEY_MODEL__`, which also lets the page be opened
+ * straight from disk without any web server.
+ */
+function resolveModelSource() {
+  const injected = globalThis.__BLUEBEY_MODEL__;
+  if (!injected) return MODEL_URL;
+  if (typeof injected !== 'string') return injected;
+  try {
+    const binary = atob(injected);
+    const bytes = new Uint8Array(binary.length);
+    for (let i = 0; i < binary.length; i += 1) bytes[i] = binary.charCodeAt(i);
+    return bytes.buffer;
+  } catch (error) {
+    console.error('inline model could not be decoded, falling back to the file', error);
+    return MODEL_URL;
+  }
+}
+
+const state = defaultState();
+
+const app = {
+  state,
+  model: null,
+  container: null,
+  ready: false,
+  needsRender: true,
+  /** The 擬音 stamp the panel is editing, so the viewport can outline it. */
+  gionSelected: null,
+  /** Set by init(), used by the panel. */
+  actions: {},
+};
+
+/* ------------------------------------------------------------------ startup */
+
+// The view-rig classes are declared below this point; deferring by a microtask
+// guarantees the whole module has finished evaluating before anything is built.
+queueMicrotask(() => {
+  init().catch((error) => {
+    console.error(error);
+    showLoadError(error);
+  });
+});
+
+async function init() {
+  const canvas = document.getElementById('view');
+  // #stage owns the layout; the canvas only fills it (see style.css).
+  const stage = document.getElementById('stage') ?? canvas;
+
+  const renderer = new THREE.WebGLRenderer({
+    canvas,
+    antialias: true,
+    alpha: true,
+    preserveDrawingBuffer: true,
+  });
+  renderer.setPixelRatio(Math.min(window.devicePixelRatio || 1, 2));
+  renderer.outputColorSpace = THREE.SRGBColorSpace;
+  renderer.toneMapping = THREE.ACESFilmicToneMapping;
+  renderer.shadowMap.enabled = true;
+  renderer.shadowMap.type = THREE.PCFShadowMap;
+
+  const scene = new THREE.Scene();
+  const environment = makeEnvironment(renderer);
+
+  const key = new THREE.DirectionalLight(0xffffff, 2.1);
+  key.castShadow = true;
+  key.shadow.mapSize.set(1536, 1536);
+  key.shadow.bias = -0.0008;
+  key.shadow.normalBias = 0.02;
+  const fill = new THREE.DirectionalLight(0xffffff, 0.5);
+  const ambient = new THREE.HemisphereLight(0xffffff, 0xd9d0f2, 0.9);
+  scene.add(key, key.target, fill, ambient);
+
+  // A shadow catcher, not a solid floor: `depthWrite: false` keeps the plane from
+  // ever hiding the character, so moving ぶるべー below it no longer makes the body
+  // (or its textures) vanish - the shadow still falls on it, but it never occludes.
+  const ground = new THREE.Mesh(
+    new THREE.PlaneGeometry(200, 200),
+    new THREE.ShadowMaterial({ opacity: 0.22, transparent: true, depthWrite: false }),
+  );
+  ground.rotation.x = -Math.PI / 2;
+  ground.receiveShadow = true;
+  scene.add(ground);
+
+  // The camera lives in two places: `state.view` (what the panel, the saved file
+  // and a shared link use) and the orbit controls (what the mouse moves). Capture
+  // the controls' own position back into the state on every change, so the two
+  // never disagree - otherwise any refresh of the view yanked the camera back to
+  // wherever the panel had last put it.
+  const view = new ViewRig(canvas, () => {
+    view.captureInto(state.view);
+    app.needsRender = true;
+  });
+
+  /**
+   * The screen-space outline (see src/outline.js). The normal pass must not see
+   * the ground, the contact shadow or the gizmo, so they are excluded once those
+   * objects exist (see below).
+   */
+  const outline = createScreenOutline({
+    renderer,
+    scene,
+    camera: view.camera,
+    width: canvas.clientWidth || 1280,
+    height: canvas.clientHeight || 800,
+  });
+  outline.exclude([ground]);
+
+  /**
+   * The drawing buffer must follow the canvas' CSS size. Without this the
+   * browser stretches the default 300x150 buffer across the window and the
+   * whole picture looks coarse (and every exported trace is built from those
+   * few pixels).
+   *
+   * But it must not be *unlimited* either: on a large or high-density display
+   * the buffer can reach many millions of pixels, and with antialiasing, a
+   * shadow map and preserveDrawingBuffer on top, a weak GPU can take seconds
+   * per frame - which looks exactly like "nothing is showing". So the pixel
+   * count is capped and the ratio is lowered automatically on slow hardware.
+   */
+  const MAX_DRAWING_PIXELS = 2_600_000;
+  const QUALITY_FLOOR = 0.5;
+  let quality = Math.min(window.devicePixelRatio || 1, 2);
+  let autoReductions = 0;
+
+  function fitQuality(width, height, wanted) {
+    let ratio = wanted;
+    while (ratio > QUALITY_FLOOR && width * height * ratio * ratio > MAX_DRAWING_PIXELS) {
+      ratio = Math.round((ratio - 0.1) * 100) / 100;
+    }
+    return Math.max(QUALITY_FLOOR, ratio);
+  }
+
+  let lastResize = 0;
+  let resizeBurst = 0;
+
+  /** Declared early: the viewport can resize before the model has loaded. */
+  let captionCanvas = null;
+  /** Set while undo/redo replays a snapshot, so it is not recorded again. */
+  let suspendHistory = false;
+
+  // ------------------------------------------------------------------- 擬音
+  // `redrawCaption` runs from the very first `resizeViewport`, so the sheet cache
+  // and the drag handles have to exist before it does (a `const` cannot be read
+  // before it is evaluated). The drawing itself lives further down, next to the
+  // caption overlay it is layered under.
+  const gionImages = new Map();   // sheet name -> loaded Image
+  const gionLoading = new Map();  // sheet name -> in-flight Promise
+  const gionHandles = new Map();  // stamp id -> transparent drag div
+  let gionDrag = null;
+  /** Hands out stamp ids; a counter so removing one never reuses a live id. */
+  let gionSeq = 0;
+
+  function resizeViewport() {
+    const width = stage.clientWidth || window.innerWidth;
+    const height = stage.clientHeight || window.innerHeight;
+
+    // Safety net: if something ever feeds the viewport size back into itself,
+    // stop flipping the drawing buffer on and off and say so.
+    const now = performance.now();
+    if (now - lastResize < 60) {
+      resizeBurst += 1;
+      if (resizeBurst > 30 && resizeBurst % 30 === 0) {
+        console.warn('[bluebey] too many resizes in a row; ignoring them', resizeBurst);
+      }
+      if (resizeBurst > 30) return;
+    } else {
+      resizeBurst = 0;
+    }
+    lastResize = now;
+
+    const wanted = Math.min(window.devicePixelRatio || 1, 2);
+    quality = fitQuality(width, height, wanted);
+    renderer.setPixelRatio(quality);
+    renderer.setSize(width, height, false);
+    outline.setSize(width * quality, height * quality);
+    view.resize();
+    redrawCaption();
+    app.needsRender = true;
+  }
+  resizeViewport();
+  new ResizeObserver(() => resizeViewport()).observe(stage);
+
+  // Losing the WebGL context leaves a permanently blank canvas, so say so
+  // instead of leaving the user staring at nothing.
+  canvas.addEventListener('webglcontextlost', (event) => {
+    event.preventDefault();
+    showNotice('描画が止まりました（WebGLコンテキストを失いました）。\n'
+      + 'ブラウザのウィンドウを小さくするか、再読み込みしてください。');
+  });
+  canvas.addEventListener('webglcontextrestored', () => {
+    app.needsRender = true;
+  });
+
+  // -------------------------------------------------------------- the model
+  const model = await loadModel(resolveModelSource(), {
+    onProgress: (ratio) => setLoadingProgress(ratio),
+  });
+  app.model = model;
+
+  // The character hangs off a mirror group of its own, so "左右反転" flips the
+  // model without touching the pose offset (`container.position`) or the props.
+  const mirrorGroup = new THREE.Group();
+  mirrorGroup.add(model.root);
+  const container = new THREE.Group();
+  container.add(mirrorGroup);
+  scene.add(container);
+  app.container = container;
+
+  // 小物 live outside the mirror group: they are scenery, not part of the body.
+  const propRoot = new THREE.Group();
+  scene.add(propRoot);
+
+  // --- 鼻ちょうちん: a modelled bubble hung under the nose --------------------
+  // The nose is a *skinned* part of the body, so a child of the nose mesh would
+  // sit at the origin instead of on the face. The whole body (nose included)
+  // rides the `master` bone, so the bubble is parented there, placed at the nose's
+  // rest position in that bone's local frame, and read from the face state each
+  // frame in the render loop.
+  function applySnotBubble(group, snot) {
+    if (!group) return;
+    group.visible = snot?.enabled === true;
+    if (!group.visible) return;
+    // The offsets are applied in the *bone's* frame, along the world axes measured
+    // once at rest, so the bubble travels with the body. Recomputing them against
+    // the live world matrix made it snap back whenever the body had moved.
+    const k = 0.01; // artwork px -> world units, enough to nudge the bubble
+    const { localBase, axes } = group.userData;
+    group.position.copy(localBase)
+      .addScaledVector(axes.x, (snot.offsetX ?? 0) * k)
+      .addScaledVector(axes.y, (snot.offsetY ?? 0) * k)
+      .addScaledVector(axes.z, (snot.offsetZ ?? 0) * k);
+    group.userData.baseScale = Number.isFinite(snot.size) ? snot.size : 1;
+    group.scale.setScalar(group.userData.baseScale);
+    group.userData.material.color.set(snot.color ?? '#dfe8ff');
+  }
+
+  const snotBubble = (() => {
+    const nose = model.parts.noseMesh;
+    const master = model.bones.find((entry) => entry.name.replace(/[.\s]/g, '').toLowerCase() === 'master');
+    if (!nose || !master || !nose.isSkinnedMesh) return null;
+    model.root.updateMatrixWorld(true);
+    // The nose is a *skinned* mesh, so its raw geometry sits in bind space (at the
+    // back of the model, in fact) - sample the skinned positions instead, which is
+    // where the nose actually ends up.
+    const attr = nose.geometry.attributes.position;
+    const step = Math.max(1, Math.floor(attr.count / 400));
+    const point = new THREE.Vector3();
+    const centre = new THREE.Vector3();
+    let samples = 0;
+    for (let i = 0; i < attr.count; i += step) {
+      point.fromBufferAttribute(attr, i);
+      nose.applyBoneTransform(i, point);
+      centre.add(point);
+      samples += 1;
+    }
+    if (!samples) return null;
+    centre.multiplyScalar(1 / samples);
+    nose.localToWorld(centre);
+    let radius = 0;
+    for (let i = 0; i < attr.count; i += step) {
+      point.fromBufferAttribute(attr, i);
+      nose.applyBoneTransform(i, point);
+      nose.localToWorld(point);
+      radius = Math.max(radius, point.distanceTo(centre));
+    }
+    if (!Number.isFinite(radius) || radius <= 0) radius = 0.25;
+    const material = new THREE.MeshStandardMaterial({
+      color: 0xdfe8ff, roughness: 0.45, metalness: 0, transparent: true, opacity: 0.92,
+    });
+    const group = new THREE.Group();
+    group.name = 'snot-bubble';
+    const bubble = new THREE.Mesh(new THREE.SphereGeometry(radius * 1.15, 28, 20), material);
+    group.add(bubble);
+    const bone = master.bone;
+    const baseWorld = centre.clone().add(new THREE.Vector3(radius * 1.9, -radius * 0.85, radius * 0.7));
+    const localBase = bone.worldToLocal(baseWorld.clone());
+    // World-axis directions expressed in the bone's frame, captured at rest: the
+    // sliders keep meaning "right / up / towards the viewer" while the bubble rides
+    // with the body.
+    const boneQuat = new THREE.Quaternion();
+    bone.getWorldQuaternion(boneQuat);
+    const invQuat = boneQuat.clone().invert();
+    const axes = {
+      x: new THREE.Vector3(1, 0, 0).applyQuaternion(invQuat),
+      y: new THREE.Vector3(0, 1, 0).applyQuaternion(invQuat),
+      z: new THREE.Vector3(0, 0, 1).applyQuaternion(invQuat),
+    };
+    group.position.copy(localBase);
+    group.userData = { bubble, material, bone, localBase, axes, baseScale: 1 };
+    group.visible = false;
+    bone.add(group);
+    return group;
+  })();
+  const snotMeshes = snotBubble ? [...snotBubble.children] : [];
+  applySnotBubble(snotBubble, state.face.eyes.snot);
+
+  // The hat rides the `master` bone like the snot bubble does: it is parented
+  // there, placed at the top of the head in that bone's local frame, and turned
+  // so it is world-aligned at rest (the bone's own rest rotation cancelled once).
+  // The character's front is +z, so the hats are modelled facing +z.
+  const hatMount = (() => {
+    const master = model.bones.find((entry) => entry.name.replace(/[.\s]/g, '').toLowerCase() === 'master');
+    if (!master) return null;
+    model.root.updateMatrixWorld(true);
+    const bone = master.bone;
+    const group = new THREE.Group();
+    group.name = 'hat';
+    group.position.copy(bone.worldToLocal(new THREE.Vector3(0, model.bounds.max.y - 0.06, 0)));
+    group.quaternion.copy(bone.getWorldQuaternion(new THREE.Quaternion()).invert());
+    bone.add(group);
+    return group;
+  })();
+
+  let hatKindCurrent = null;
+  let hatObject = null;
+  /** A GLB hat the user loaded; kept so it can be re-chosen without re-importing. */
+  let customHatObject = null;
+  const hatMeshes = [];
+  /** Last transform applied, so the render loop only wakes when it changes. */
+  const hatXform = { x: null, z: null, height: null, tiltX: null, tiltZ: null };
+  /** Set once the outline pass exists, so a hat change can refresh the exclude set. */
+  let refreshOutlineExclusion = null;
+  /** Set once the style system exists, so a hat can register for styles + outline. */
+  let hatStyleRegister = null;
+  let hatStyleUnregister = null;
+
+  /**
+   * Swap the mounted hat for `built` (or clear it), keeping the style system and
+   * the outline exclusion in step.
+   *
+   * The old meshes are collected BEFORE they are unregistered: `removeMesh`
+   * deletes the outline hull (a child of the hat group) while we traverse, which
+   * used to shorten `children` mid-loop and call `.traverse` on `undefined`.
+   */
+  function mountHat(built) {
+    if (hatObject) {
+      const old = [];
+      hatObject.traverse((o) => { if (o.isMesh) old.push(o); });
+      hatMount.remove(hatObject);
+      for (const mesh of old) hatStyleUnregister?.(mesh);
+      // The custom model is kept so it can be re-chosen; a built hat is dropped.
+      if (hatObject !== customHatObject) {
+        for (const mesh of old) {
+          mesh.geometry?.dispose();
+          if (Array.isArray(mesh.material)) mesh.material.forEach((m) => m.dispose());
+          else mesh.material?.dispose();
+        }
+      }
+      hatObject = null;
+    }
+    hatMeshes.length = 0;
+    if (built) {
+      hatMount.add(built);
+      hatObject = built;
+      const meshes = [];
+      built.traverse((o) => { if (o.isMesh) meshes.push(o); });
+      for (const mesh of meshes) {
+        hatMeshes.push(mesh);
+        hatStyleRegister?.(mesh);
+      }
+    }
+    refreshOutlineExclusion?.();
+  }
+
+  function applyHat(kind) {
+    if (!hatMount) return;
+    const custom = state.face.hat?.custom === true && customHatObject;
+    const key = custom ? 'custom' : kind;
+    if (key !== hatKindCurrent) {
+      hatKindCurrent = key;
+      mountHat(custom ? customHatObject : buildHat(kind));
+    }
+
+    // 高さ and 傾き. The mount is world-aligned (see above), so +y is up, +z is the
+    // character's front and +x its right: a plain translation and rotation read
+    // the way the sliders say.
+    const hat = state.face.hat ?? {};
+    const x = hat.x ?? 0;
+    const z = hat.z ?? 0;
+    const height = hat.height ?? 0;
+    const tiltX = hat.tiltX ?? 0;
+    const tiltZ = hat.tiltZ ?? 0;
+    if (hatObject) {
+      hatObject.position.set(x, height, z);
+      hatObject.rotation.set((tiltX * Math.PI) / 180, 0, (tiltZ * Math.PI) / 180);
+    }
+    if (x !== hatXform.x || z !== hatXform.z || height !== hatXform.height || tiltX !== hatXform.tiltX || tiltZ !== hatXform.tiltZ) {
+      hatXform.x = x;
+      hatXform.z = z;
+      hatXform.height = height;
+      hatXform.tiltX = tiltX;
+      hatXform.tiltZ = tiltZ;
+      app.needsRender = true;
+    }
+  }
+  /**
+   * Import a GLB hat. It is centred sideways, dropped so its base sits at y = 0
+   * (the mount sits on the head), and scaled to about the head's width, then
+   * mounted like any built hat.
+   */
+  async function loadHatModel(file) {
+    if (!hatMount) return;
+    const buffer = await exporter.readFileAsArrayBuffer(file);
+    const loader = new GLTFLoader();
+    const gltf = await new Promise((resolve, reject) => loader.parse(buffer, '', resolve, reject));
+    const root = gltf.scene;
+    const box = new THREE.Box3().setFromObject(root);
+    const size = box.getSize(new THREE.Vector3());
+    const target = Math.max(0.2, model.size.x * 0.95);
+    const scale = target / Math.max(size.x, size.z, 0.001);
+    root.scale.setScalar(scale);
+    box.setFromObject(root);
+    const center = box.getCenter(new THREE.Vector3());
+    root.position.x -= center.x;
+    root.position.z -= center.z;
+    root.position.y -= box.min.y;
+    root.traverse((o) => { if (o.isMesh) { o.castShadow = true; o.receiveShadow = false; } });
+    customHatObject = root;
+    state.face.hat.custom = true;
+    state.face.hat.kind = 'none';
+    hatKindCurrent = null;
+    refresh('all');
+    app.panel?.sync();
+  }
+
+  applyHat(state.face.hat?.kind ?? 'none');
+
+  const halfHeight = model.size.y * 0.5;
+  view.frame(model.size, halfHeight);
+
+  // shadow camera big enough for the whole character
+  const radius = Math.max(model.size.x, model.size.y, model.size.z) * 1.6;
+  const shadowCamera = key.shadow.camera;
+  shadowCamera.left = -radius;
+  shadowCamera.right = radius;
+  shadowCamera.top = radius;
+  shadowCamera.bottom = -radius;
+  shadowCamera.near = 0.5;
+  shadowCamera.far = radius * 9;
+  shadowCamera.updateProjectionMatrix();
+  key.target.position.set(0, halfHeight, 0);
+
+  // The light's offset from the character, so the key light (and therefore its
+  // shadow camera) can follow the character around: moving far from the origin
+  // used to leave the shadow behind and cut it off at the shadow camera's edge.
+  const keyOffset = new THREE.Vector3(0, model.size.y * 3, model.size.y * 3);
+  function updateKeyLight() {
+    key.target.position.set(container.position.x, halfHeight + container.position.y, container.position.z);
+    key.position.copy(key.target.position).add(keyOffset);
+  }
+
+  // ------------------------------------------------------------- controllers
+  const styles = new Styles({ meshes: model.parts.body });
+  // Hats are built at runtime, so they are not in the GLB's mesh list. Register
+  // them with the style system (and the current hat) so they follow flat/line art
+  // and get an outline hull like every other part.
+  hatStyleRegister = (mesh) => styles.addMesh(mesh, { tone: true, lineOnly: true });
+  hatStyleUnregister = (mesh) => styles.removeMesh(mesh);
+  for (const mesh of hatMeshes) styles.addMesh(mesh, { tone: true, lineOnly: true });
+  const face = new Face({
+    eyeMesh: model.parts.eyeMesh,
+    mouthMesh: model.parts.mouthMesh,
+    hairMesh: model.parts.hairMesh,
+    originals: model.originals,
+  });
+  const animator = new Animator({ state });
+
+  // Beard drawings supplied as files (`assets/beards/<kind>.png`): a kind that has
+  // one is drawn from the file instead of the built-in strokes. A missing file is
+  // skipped, so the app works before any are added.
+  void (async () => {
+    // Each entry: a beard kind and the drawing in `assets/beards/`. A file ending
+    // `-right.png` is one side of a moustache; it is drawn twice (mirrored) with a
+    // `spacing` gap between them. A kind with no entry uses the built-in strokes.
+    const files = [
+      { kind: 'scotch', file: 'scotch.png' },
+      { kind: 'kaiser', file: 'kaiser-right.png' },
+    ];
+    const loaded = {};
+    await Promise.all(files.map(async ({ kind, file }) => {
+      try {
+        const response = await fetch(assetUrl(file, 'beard'));
+        if (!response.ok) return;
+        loaded[kind] = {
+          image: await createImageBitmap(await response.blob()),
+          half: file.endsWith('-right.png'),
+        };
+      } catch {
+        /* a missing drawing falls back to the built-in strokes */
+      }
+    }));
+    if (Object.keys(loaded).length) {
+      face.setBeardImages(loaded);
+      app.needsRender = true;
+    }
+  })();
+
+  // Brow drawings supplied as files (`assets/brows/<kind>.png`). A file ending
+  // `-right.png` is one side; it is drawn twice (mirrored) at each eye. Missing
+  // files are skipped, so the built-in ゴル風 strokes are the fallback.
+  void (async () => {
+    const files = [{ kind: 'gol', file: 'gol-right.png' }];
+    const loaded = {};
+    await Promise.all(files.map(async ({ kind, file }) => {
+      try {
+        const response = await fetch(assetUrl(file, 'brow'));
+        if (!response.ok) return;
+        loaded[kind] = {
+          image: await createImageBitmap(await response.blob()),
+          half: file.endsWith('-right.png'),
+        };
+      } catch {
+        /* a missing drawing falls back to the built-in strokes */
+      }
+    }));
+    if (Object.keys(loaded).length) {
+      face.setBrowImages(loaded);
+      app.needsRender = true;
+    }
+  })();
+
+  // ------------------------------------------------------------------- look
+  const look = createLook({
+    renderer,
+    scene,
+    styles,
+    model,
+    ground,
+    key,
+    character: mirrorGroup,
+    gradientEnvironment: environment,
+  });
+
+  // --------------------------------------------------------------- 見えない壁
+  // A clip plane that hides part of the character, so it can look half buried
+  // in a wall (see src/clip.js). Applied in `refresh`, with the render settings.
+  const clip = createClipper({ scene, renderer, model });
+  // A wall's position is stored relative to the character, so every time the
+  // walls are applied they are shifted by the body's own offset - that is what
+  // makes them travel with ぶるべー when it is moved.
+  const characterOffset = () => {
+    const root = state.pose?.root;
+    return { x: root?.[0] ?? 0, y: root?.[1] ?? 0, z: root?.[2] ?? 0 };
+  };
+  const applyWalls = () => clip.apply(
+    [state.render.wall, state.render.wall2],
+    characterOffset(),
+    wallYaw(),
+  );
+  // The character's own turn about Y: the `master` bone's Y (Shift+drag) plus any
+  // body yaw the 見る先 mode added. The walls take this too, so turning the
+  // character with Shift+drag no longer slides the hidden region around.
+  const wallYaw = () => {
+    const masterY = state.pose?.bones?.master?.[1] ?? 0;
+    return (masterY + (container.rotation.y * 180) / Math.PI) * DEG;
+  };
+
+  // --------------------------------------------------------------- backdrop
+  const backdrop = createBackdrop({
+    stage,
+    resolveUrl: (name) => assetUrl(name, 'background'),
+    onNeedsRender: () => { app.needsRender = true; },
+  });
+  /** The bitmap the exports paint behind the model (an Image or the camera video). */
+  let backdropImage = null;
+
+  /**
+   * Story-panel previews (keyed by panel id) and the counter that hands out
+   * those ids.
+   *
+   * WHY the previews live outside the state: a thumbnail is an image, and the
+   * state is exactly what a shared link carries - a data URL per panel would
+   * push the link past what a URL can hold. They are keyed by id and rebuilt as
+   * panels are added, so a restored link simply shows panels without previews.
+   * (Declared up here because `buildPanel` runs - and syncs - before the panel
+   * functions below are reached.)
+   */
+  const storyThumbs = new Map();
+  let storySeq = 0;
+
+  // ------------------------------------------------------- caption overlay
+  // The bubble is drawn into its own 2D canvas laid over the WebGL one, with the
+  // very same routine the PNG export uses - so the preview cannot lie.
+  captionCanvas = document.createElement('canvas');
+  captionCanvas.id = 'caption-layer';
+  captionCanvas.style.cssText =
+    'position:absolute;inset:0;width:100%;height:100%;pointer-events:none;z-index:2';
+  stage.append(captionCanvas);
+
+  /** opentype font for outline export and text layout (lazily fetched). */
+  let captionFont = null;
+  let captionFontPromise = null;
+  /** Where each bubble was last drawn, in CSS pixels (in state order). */
+  const CAPTION_KEYS = ['caption', 'caption2', 'narration'];
+  const captionBoxes = new Map();
+  const captionHandles = new Map();
+  /** Set while a bubble is being dragged. */
+  let captionDrag = null;
+
+  // The bubbles are draggable, but the caption canvas has to stay
+  // `pointer-events: none`: it covers the whole viewport, and the orbit controls
+  // and the bone picking live underneath it. So each bubble gets a separate
+  // transparent drag box, parked exactly over it by `positionCaptionHandles`.
+  for (const key of CAPTION_KEYS) {
+    const handle = document.createElement('div');
+    handle.id = `caption-handle-${key}`;
+    handle.style.cssText = 'position:absolute;display:none;cursor:move;'
+      + 'touch-action:none;pointer-events:auto;z-index:3';
+    handle.addEventListener('pointerdown', (event) => startCaptionDrag(event, key));
+    handle.addEventListener('pointermove', moveCaptionDrag);
+    handle.addEventListener('pointerup', endCaptionDrag);
+    handle.addEventListener('pointercancel', endCaptionDrag);
+    handle.dataset.captionHandle = '1';
+    stage.append(handle);
+    captionHandles.set(key, handle);
+  }
+
+  /** Park each drag target over its bubble (or hide it when it is not shown). */
+  function positionCaptionHandles() {
+    for (const key of CAPTION_KEYS) {
+      const handle = captionHandles.get(key);
+      const box = captionBoxes.get(key);
+      if (!handle) continue;
+      if (!box) {
+        handle.style.display = 'none';
+        continue;
+      }
+      // A few pixels of slack, so the rounded corners are still easy to grab.
+      handle.style.display = 'block';
+      handle.style.left = `${box.x - 4}px`;
+      handle.style.top = `${box.y - 4}px`;
+      handle.style.width = `${box.w + 8}px`;
+      handle.style.height = `${box.h + 8}px`;
+    }
+  }
+
+  function startCaptionDrag(event, key) {
+    const caption = state[key];
+    const box = captionBoxes.get(key);
+    if (caption?.enabled !== true || !box) return;
+    // Touching a bubble also selects it in the panel, so the sliders edit the one
+    // you just grabbed.
+    app.panel?.selectCaption?.(key);
+    event.preventDefault();
+    event.stopPropagation();
+    // Not all pointers can be captured (and a synthetic one cannot), so a
+    // failure here must not stop the drag from starting.
+    try {
+      captionHandles.get(key)?.setPointerCapture(event.pointerId);
+    } catch {
+      /* capture is a nicety, not a requirement */
+    }
+    captionDrag = {
+      key,
+      pointerId: event.pointerId,
+      fromX: event.clientX,
+      fromY: event.clientY,
+      boxX: box.x,
+      boxY: box.y,
+    };
+  }
+
+  function moveCaptionDrag(event) {
+    if (!captionDrag || event.pointerId !== captionDrag.pointerId) return;
+    const caption = state[captionDrag.key];
+    if (!caption) return;
+    event.preventDefault();
+    const width = stage.clientWidth || 1280;
+    const height = stage.clientHeight || 800;
+    caption.x = (captionDrag.boxX + (event.clientX - captionDrag.fromX)) / width;
+    caption.y = (captionDrag.boxY + (event.clientY - captionDrag.fromY)) / height;
+    // Store back what the layout will actually use, so pushing past an edge does
+    // not pile up an out-of-range value that has to be undone before the bubble
+    // moves again. `refresh` redraws the bubble and records one undo step (the
+    // history coalesces a run of the same label).
+    const layout = layoutCaption(caption, { width, height, scale: 1, font: captionFont });
+    if (width > 0) caption.x = layout.box.x / width;
+    if (height > 0) caption.y = layout.box.y / height;
+    refresh('caption');
+  }
+
+  function endCaptionDrag(event) {
+    if (!captionDrag || (event && event.pointerId !== captionDrag.pointerId)) return;
+    captionDrag = null;
+    // The 横位置 / 縦位置 sliders catch up when the drag ends, not every pixel.
+    app.actions.syncPanel?.();
+  }
+
+  async function ensureCaptionFont() {
+    if (captionFontPromise) return captionFontPromise;
+    captionFontPromise = (async () => {
+      try {
+        const face = new FontFace('M PLUS Rounded 1c', `url(${assetUrl('bluebey-caption.woff2')})`);
+        document.fonts.add(await face.load());
+      } catch (error) {
+        console.warn('[bluebey] caption font face failed', error);
+      }
+      try {
+        captionFont = await loadFont(assetUrl('bluebey-caption.ttf'));
+      } catch (error) {
+        console.warn('[bluebey] caption outlines unavailable', error);
+      }
+      redrawCaption();
+      return captionFont;
+    })();
+    return captionFontPromise;
+  }
+
+  function redrawCaption() {
+    if (!captionCanvas) return;
+    const width = stage.clientWidth || 1280;
+    const height = stage.clientHeight || 800;
+    const ratio = Math.min(2, window.devicePixelRatio || 1);
+    const wantW = Math.round(width * ratio);
+    const wantH = Math.round(height * ratio);
+    if (captionCanvas.width !== wantW || captionCanvas.height !== wantH) {
+      captionCanvas.width = wantW;
+      captionCanvas.height = wantH;
+    }
+    const ctx = captionCanvas.getContext('2d');
+    ctx.setTransform(ratio, 0, 0, ratio, 0, 0);
+    ctx.clearRect(0, 0, width, height);
+    // 擬音はセリフより背面に置く: the bubbles stay readable on top of them.
+    const gionItems = state.gion?.items ?? [];
+    for (const item of gionItems) ensureGionSheet(item.sheet);
+    drawGion(ctx, gionItems, gionImages, { width, height, scale: 1 });
+    for (const key of CAPTION_KEYS) {
+      const caption = state[key];
+      captionBoxes.delete(key);
+      if (!caption?.enabled) continue;
+      // The vendored font is only fetched once a bubble is actually used; when it
+      // arrives `ensureCaptionFont` redraws this canvas.
+      if (!captionFontPromise) void ensureCaptionFont();
+      captionBoxes.set(key, drawCaption(ctx, caption, { width, height, scale: 1, font: captionFont }));
+    }
+    positionCaptionHandles();
+    ensureGionHandles(gionItems);
+    positionGionHandles(gionItems);
+  }
+
+  // ----------------------------------------------------------------- 擬音
+  // A stamp is a crop of a sheet, drawn on the caption canvas *under* the bubbles.
+  // The canvas is `pointer-events: none`, so - exactly like a caption - each stamp
+  // gets its own transparent drag box, parked over it by `positionGionHandles`.
+
+  /** Load a sheet once, on first use, and redraw when it arrives. */
+  function ensureGionSheet(name) {
+    if (gionImages.has(name)) return Promise.resolve(gionImages.get(name));
+    if (gionLoading.has(name)) return gionLoading.get(name);
+    const image = new Image();
+    const promise = new Promise((resolve) => {
+      image.onload = () => resolve(image);
+      image.onerror = () => resolve(null);
+    }).then((loaded) => {
+      gionLoading.delete(name);
+      if (loaded) {
+        gionImages.set(name, loaded);
+        redrawCaption();
+      }
+      return loaded;
+    });
+    gionLoading.set(name, promise);
+    image.src = gionSheetUrl(name);
+    return promise;
+  }
+
+  /** Create one drag target per stamp, and drop the ones whose stamp is gone. */
+  function ensureGionHandles(items) {
+    const live = new Set();
+    for (const item of items) {
+      live.add(item.id);
+      if (gionHandles.has(item.id)) continue;
+      const handle = document.createElement('div');
+      handle.className = 'gion-handle';
+      handle.addEventListener('pointerdown', (event) => startGionDrag(event, item.id));
+      handle.addEventListener('pointermove', moveGionDrag);
+      handle.addEventListener('pointerup', endGionDrag);
+      handle.addEventListener('pointercancel', endGionDrag);
+      handle.dataset.gionHandle = '1';
+      stage.append(handle);
+      gionHandles.set(item.id, handle);
+    }
+    for (const [id, handle] of gionHandles) {
+      if (live.has(id)) continue;
+      handle.remove();
+      gionHandles.delete(id);
+    }
+  }
+
+  /** Park each drag target over its stamp (or hide it while the sheet loads). */
+  function positionGionHandles(items) {
+    const width = stage.clientWidth || 1280;
+    const height = stage.clientHeight || 800;
+    for (const item of items) {
+      const handle = gionHandles.get(item.id);
+      if (!handle) continue;
+      const image = gionImages.get(item.sheet);
+      if (!imageReady(image)) {
+        handle.style.display = 'none';
+        continue;
+      }
+      const rect = stampRect(item, { width, height, scale: 1 }, imageAspect(image));
+      handle.style.display = 'block';
+      handle.style.left = `${rect.x}px`;
+      handle.style.top = `${rect.y}px`;
+      handle.style.width = `${rect.w}px`;
+      handle.style.height = `${rect.h}px`;
+      // Match the drawn stamp, which is rotated about its centre.
+      handle.style.transformOrigin = 'center';
+      handle.style.transform = item.rot ? `rotate(${item.rot}deg)` : 'none';
+      handle.classList.toggle('selected', item.id === app.gionSelected);
+    }
+  }
+
+  function startGionDrag(event, id) {
+    const item = (state.gion?.items ?? []).find((entry) => entry.id === id);
+    if (!item) return;
+    // Touching a stamp also selects it in the panel, so the sliders edit the one
+    // you just grabbed.
+    app.panel?.selectGion?.(id);
+    event.preventDefault();
+    event.stopPropagation();
+    try {
+      gionHandles.get(id)?.setPointerCapture(event.pointerId);
+    } catch {
+      /* capture is a nicety, not a requirement */
+    }
+    gionDrag = {
+      id,
+      pointerId: event.pointerId,
+      fromX: event.clientX,
+      fromY: event.clientY,
+      x: item.x ?? 0.5,
+      y: item.y ?? 0.5,
+    };
+  }
+
+  function moveGionDrag(event) {
+    if (!gionDrag || event.pointerId !== gionDrag.pointerId) return;
+    const item = (state.gion?.items ?? []).find((entry) => entry.id === gionDrag.id);
+    if (!item) return;
+    event.preventDefault();
+    const width = stage.clientWidth || 1280;
+    const height = stage.clientHeight || 800;
+    item.x = clamp01(gionDrag.x + (event.clientX - gionDrag.fromX) / width);
+    item.y = clamp01(gionDrag.y + (event.clientY - gionDrag.fromY) / height);
+    refresh('gion');
+  }
+
+  function endGionDrag(event) {
+    if (!gionDrag || (event && event.pointerId !== gionDrag.pointerId)) return;
+    gionDrag = null;
+    app.actions.syncPanel?.();
+  }
+
+  /** Add a stamp from a picker crop, centred, and hand it to the panel. */
+  function addGionStamp({ sheet, sx, sy, sw, sh }) {
+    gionSeq += 1;
+    const item = {
+      id: `g${gionSeq}`,
+      sheet,
+      sx, sy, sw, sh,
+      x: 0.5,
+      y: 0.5,
+      w: DEFAULT_STAMP_WIDTH,
+      rot: 0,
+      flip: false,
+    };
+    state.gion = state.gion ?? { items: [] };
+    state.gion.items = [...(state.gion.items ?? []), item];
+    app.gionSelected = item.id;
+    void ensureGionSheet(sheet);
+    applyState({}, { scope: 'gion', sync: true });
+    toast('擬音を追加しました');
+  }
+
+  // ------------------------------------------------------------------ props
+  let propSignature = '';
+
+  function applyProps() {
+    const items = state.props?.items ?? [];
+    const signature = JSON.stringify([items, state.render.colors]);
+    if (signature === propSignature) return;
+    propSignature = signature;
+    for (const child of [...propRoot.children]) {
+      // Hand every mesh back to the style system before the prop is disposed.
+      // `removeMesh` restores each mesh's own material (a prop under 線画 is
+      // wearing the shared paper material, which must not be disposed). Collect
+      // the meshes first: `removeMesh` deletes the outline hull, a child of the
+      // prop group.
+      const meshes = [];
+      child.traverse((object) => {
+        if (styles.originals.has(object)) meshes.push(object);
+      });
+      for (const mesh of meshes) styles.removeMesh(mesh);
+      propRoot.remove(child);
+      disposeProp(child);
+    }
+    for (const item of items) {
+      const def = PROP_DEFAULTS[item.kind] ?? { x: 0, y: 0, z: 0, rotX: 0, rotY: 0, rotZ: 0, scale: 1 };
+      let group;
+      try {
+        group = buildProp(item.kind, { colors: state.render.colors, scale: item.scale ?? def.scale ?? 1 });
+      } catch (error) {
+        console.warn('[bluebey] unknown prop', item.kind, error);
+        continue;
+      }
+      group.position.set(item.x ?? def.x ?? 0, item.y ?? def.y ?? 0, item.z ?? def.z ?? 0);
+      group.rotation.set(
+        (item.rotX ?? def.rotX ?? 0) * DEG,
+        (item.rotY ?? def.rotY ?? 0) * DEG,
+        (item.rotZ ?? def.rotZ ?? 0) * DEG,
+      );
+      propRoot.add(group);
+      // Register with the style system, so フラット and 線画 reach the props too
+      // and each part gets an outline hull like the body and the hats. `lineOnly`
+      // keeps that outline out of リアル/フラット, where the props read by shading.
+      group.traverse((object) => {
+        if (object.isMesh) styles.addMesh(object, { lineOnly: true });
+      });
+      // 看板 only: paint the saved text onto the freshly built face (no-op for
+      // every other prop, which has no writing surface).
+      applyPropText(group, item.text);
+    }
+    refreshOutlineExclusion?.();
+    app.needsRender = true;
+  }
+
+  // ------------------------------------------------------------ prop dragging
+  // A prop can be dragged with the mouse. The pointer picks one, then moves it in
+  // the plane that faces the camera, so it follows the cursor at whatever depth
+  // it already sits at. (The ground plane is the other obvious choice, but the
+  // camera sits almost level with the floor, so the floor is nearly edge-on and
+  // a prop dragged across it would fly off.)
+  //
+  // The listener is on `stage` in the CAPTURE phase: that runs before the canvas
+  // and `view`'s container, so stopping the event there keeps the orbit controls
+  // and the bone picking out of a prop drag.
+  const propRaycaster = new THREE.Raycaster();
+  const propPointer = new THREE.Vector2();
+  const propPlane = new THREE.Plane();
+  const propPoint = new THREE.Vector3();
+  const propNormal = new THREE.Vector3();
+  let propDrag = null;
+  // The 見えない壁 uses the same machinery: `wallDrag` keeps the grabbed point's
+  // offset from the wall's anchor so grabbing a corner does not make it jump.
+  const wallDrag = { pointerId: null, key: 'wall', offset: new THREE.Vector3() };
+
+  /** The pointer in normalised device coordinates, i.e. what a raycaster wants. */
+  function pointerNdc(event) {
+    const rect = canvas.getBoundingClientRect();
+    return propPointer.set(
+      ((event.clientX - rect.left) / Math.max(1, rect.width)) * 2 - 1,
+      -((event.clientY - rect.top) / Math.max(1, rect.height)) * 2 + 1,
+    );
+  }
+
+  /** The prop under the pointer as `{ index, group, point }`, or null. */
+  function propAt(event) {
+    if (!propRoot.children.length) return null;
+    propRaycaster.setFromCamera(pointerNdc(event), view.camera);
+    const hits = propRaycaster.intersectObjects(propRoot.children, true);
+    if (!hits.length) return null;
+    // The meshes hang off the group `propRoot` holds, so walk up to that group.
+    let node = hits[0].object;
+    while (node.parent && node.parent !== propRoot) node = node.parent;
+    const items = state.props?.items ?? [];
+    const index = propRoot.children.indexOf(node);
+    if (index < 0 || index >= items.length) return null;
+    return { index, group: node, point: hits[0].point.clone() };
+  }
+
+  /**
+   * The point on the wall's guide under the pointer, or null.
+   *
+   * Only while the wall is on AND its guide is switched on: the guide is both
+   * what there is to grab and the switch that says "I am placing the wall now",
+   * so with it off a huge invisible plane cannot swallow every orbit drag.
+   */
+  function wallAt(event) {
+    propRaycaster.setFromCamera(pointerNdc(event), view.camera);
+    let best = null;
+    for (let index = 0; index < clip.guides.length; index += 1) {
+      const key = index === 0 ? 'wall' : 'wall2';
+      const wall = state.render[key];
+      if (wall?.on !== true || wall?.guide !== true) continue;
+      const hits = propRaycaster.intersectObject(clip.guides[index], false);
+      if (!hits.length) continue;
+      if (!best || hits[0].distance < best.distance) {
+        best = { index, key, point: hits[0].point.clone() };
+      }
+    }
+    return best;
+  }
+
+  function onPropPointerDown(event) {
+    if (event.button !== 0) return;
+    if (event.target?.dataset?.captionHandle) return;   // a bubble drags itself
+    if (rig?.controls?.dragging) return;          // the rotation gizmo is in charge
+    const hit = propAt(event);
+    if (hit) {
+      // Claim the pointer, so nothing underneath sees it.
+      event.stopPropagation();
+      event.preventDefault();
+      view.camera.getWorldDirection(propNormal);
+      propPlane.setFromNormalAndCoplanarPoint(propNormal.negate(), hit.point);
+      propDrag = {
+        pointerId: event.pointerId,
+        index: hit.index,
+        group: hit.group,
+        // Grabbing a corner rather than the middle must not make it jump.
+        offset: hit.group.position.clone().sub(hit.point),
+      };
+      stage.style.cursor = 'grabbing';
+      app.needsRender = true;
+      return;
+    }
+
+    const onWall = wallAt(event);
+    if (!onWall) {
+      // Nothing to grab. Remember the press, and settle on pointer-up whether it
+      // was a tap (set the 見る先 target) or the start of an orbit.
+      if (state.lookAt?.enabled === true && !rig?.controls?.dragging) {
+        lookTap = { pointerId: event.pointerId, x: event.clientX, y: event.clientY };
+      }
+      return;
+    }
+    // A prop wins if it is under the pointer; otherwise the wall takes it.
+    event.stopPropagation();
+    event.preventDefault();
+    view.camera.getWorldDirection(propNormal);
+    propPlane.setFromNormalAndCoplanarPoint(propNormal.negate(), onWall.point);
+    const wall = state.render[onWall.key] ?? (state.render[onWall.key] = defaultState().render[onWall.key]);
+    const offset = characterOffset();
+    const yaw = wallYaw();
+    const cos = Math.cos(yaw);
+    const sin = Math.sin(yaw);
+    const px = onWall.point.x - offset.x;
+    const pz = onWall.point.z - offset.z;
+    wallDrag.pointerId = event.pointerId;
+    wallDrag.key = onWall.key;
+    // Keep the grab point's offset *in the wall's own (character-relative) frame*,
+    // so grabbing a corner does not make the wall jump.
+    wallDrag.offset.set(
+      (wall.x ?? 0) - (cos * px - sin * pz),
+      (wall.y ?? 0) - (onWall.point.y - offset.y),
+      (wall.z ?? 0) - (sin * px + cos * pz),
+    );
+    stage.style.cursor = 'grabbing';
+    app.needsRender = true;
+  }
+
+  function onPropPointerMove(event) {
+    if (lookTap && event.pointerId === lookTap.pointerId
+      && Math.hypot(event.clientX - lookTap.x, event.clientY - lookTap.y) > LOOK_TAP_SLOP) {
+      lookTap = null;   // that was an orbit drag, not a tap
+    }
+    if (wallDrag.pointerId != null && event.pointerId === wallDrag.pointerId) {
+      event.stopPropagation();
+      propRaycaster.setFromCamera(pointerNdc(event), view.camera);
+      if (!propRaycaster.ray.intersectPlane(propPlane, propPoint)) return;
+      const offset = characterOffset();
+      const yaw = wallYaw();
+      const cos = Math.cos(yaw);
+      const sin = Math.sin(yaw);
+      const px = propPoint.x - offset.x;
+      const pz = propPoint.z - offset.z;
+      const wall = state.render[wallDrag.key] ?? (state.render[wallDrag.key] = defaultState().render[wallDrag.key]);
+      // Back into the wall's character-relative frame, then the same ranges the
+      // panel's sliders offer, so a drag and a number always describe the same
+      // place.
+      wall.x = round2(clampNumber(cos * px - sin * pz + wallDrag.offset.x, -4, 4));
+      wall.y = round2(clampNumber(propPoint.y - offset.y + wallDrag.offset.y, -2, 4));
+      wall.z = round2(clampNumber(sin * px + cos * pz + wallDrag.offset.z, -4, 4));
+      applyWalls();
+      app.needsRender = true;
+      return;
+    }
+    if (!propDrag || event.pointerId !== propDrag.pointerId) return;
+    event.stopPropagation();
+    propRaycaster.setFromCamera(pointerNdc(event), view.camera);
+    if (!propRaycaster.ray.intersectPlane(propPlane, propPoint)) return;
+    const next = propPoint.add(propDrag.offset);
+    const item = state.props?.items?.[propDrag.index];
+    if (!item) return;
+    // Clamp to the ranges the panel's sliders offer, so a drag can never put a
+    // prop somewhere the numbers cannot describe (and never through the floor).
+    item.x = round2(clampNumber(next.x, -4, 4));
+    item.y = round2(clampNumber(next.y, 0, 2));
+    item.z = round2(clampNumber(next.z, -4, 4));
+    propDrag.group.position.set(item.x, item.y, item.z);
+    // Keep `applyProps`' signature in step with what is on screen, so a refresh
+    // during the drag does not tear the group down and rebuild it from the state
+    // (which would re-create every geometry on every pointer move).
+    propSignature = JSON.stringify([state.props?.items ?? [], state.render.colors]);
+    app.needsRender = true;
+  }
+
+  function onPropPointerUp(event) {
+    if (lookTap && (!event || event.pointerId === lookTap.pointerId)) {
+      lookTap = null;
+      if (event) applyLookTap(event);
+      return;
+    }
+    if (wallDrag.pointerId != null && (!event || event.pointerId === wallDrag.pointerId)) {
+      wallDrag.pointerId = null;
+      stage.style.cursor = '';
+      history.push(state, '見えない壁');
+      app.panel?.sync();
+      return;
+    }
+    if (!propDrag || (event && event.pointerId !== propDrag.pointerId)) return;
+    propDrag = null;
+    stage.style.cursor = '';
+    // One undo step for the whole drag, and the panel catches up with the value.
+    history.push(state, '小物');
+    app.panel?.sync();
+  }
+
+  // A tap (a press that does not turn into a drag) in the viewport sets the
+  // 見る先 target, when that mode is on. Cancelled as soon as the pointer moves,
+  // so an orbit drag never moves the target by accident.
+  let lookTap = null;
+  const LOOK_TAP_SLOP = 5;
+
+  stage.addEventListener('pointerdown', onPropPointerDown, { capture: true });
+  window.addEventListener('pointermove', onPropPointerMove);
+  window.addEventListener('pointerup', onPropPointerUp);
+  window.addEventListener('pointercancel', onPropPointerUp);
+
+  // ---------------------------------------------------------------- look-at
+  const lookScratch = new THREE.Vector3();
+  const headScratch = new THREE.Vector3();
+  const lookPlane = new THREE.Plane();
+  const floorPlane = new THREE.Plane(new THREE.Vector3(0, 1, 0), 0);
+  const lookPlaneNormal = new THREE.Vector3();
+  const lookPivot = new THREE.Vector3();
+  const lookHit = new THREE.Vector3();
+  const lookFloor = new THREE.Vector3();
+  /** Where the eyes should aim this frame, from `state.lookAt` (a world point). */
+  let aimOffset = { x: 0, y: 0 };
+
+  /**
+   * The world point under a viewport tap, for the 見る先 target.
+   *
+   * The tap is projected onto the first thing it can sensibly mean: a vertical
+   * plane through the character that faces the camera (so a tap on the body or
+   * the backdrop keeps the character's own depth), or the floor, for a tap that
+   * lands on the ground in front. The nearer of the two wins.
+   */
+  function lookAtPointFrom(event) {
+    propRaycaster.setFromCamera(pointerNdc(event), view.camera);
+    const ray = propRaycaster.ray;
+    const headY = container.position.y + model.size.y * 0.72;
+    lookPlaneNormal.set(ray.direction.x, 0, ray.direction.z);
+    if (lookPlaneNormal.lengthSq() < 1e-8) lookPlaneNormal.set(0, 0, 1);
+    lookPlane.setFromNormalAndCoplanarPoint(
+      lookPlaneNormal.normalize(),
+      lookPivot.set(container.position.x, headY, container.position.z),
+    );
+    const onPlane = ray.intersectPlane(lookPlane, lookHit);
+    const onFloor = ray.direction.y < -1e-4 ? ray.intersectPlane(floorPlane, lookFloor) : null;
+    if (!onPlane) return onFloor;
+    if (!onFloor) return onPlane;
+    return onPlane.distanceToSquared(ray.origin) <= onFloor.distanceToSquared(ray.origin) ? onPlane : onFloor;
+  }
+
+  /** A tap in the viewport: aim the eyes at whatever was under the pointer. */
+  function applyLookTap(event) {
+    if (state.lookAt?.enabled !== true) return;
+    const point = lookAtPointFrom(event);
+    if (!point) return;
+    const cfg = state.lookAt;
+    // The same ranges the panel's sliders offer, so a tap and a number describe
+    // the same place.
+    cfg.x = round2(clampNumber(point.x, -6, 6));
+    cfg.y = round2(clampNumber(point.y, 0, 5));
+    cfg.z = round2(clampNumber(point.z, -6, 6));
+    history.push(state, '見る先');
+    refresh('lookAt');
+    app.panel?.sync();
+  }
+
+  function applyLookAt() {
+    const cfg = state.lookAt;
+    if (!cfg?.enabled) {
+      // Switching the mode off leaves the character where it was: the body turn
+      // and the gaze are remembered, not re-derived, so nothing snaps back.
+      container.rotation.y = (cfg?.bodyYawDeg ?? 0) * DEG;
+      const frozen = cfg?.freeze;
+      return frozen
+        ? { x: clamp(frozen.x ?? 0, -1, 1), y: clamp(frozen.y ?? 0, -1, 1) }
+        : { x: 0, y: 0 };
+    }
+
+    container.updateMatrixWorld(true);
+    lookScratch.set(cfg.x ?? 0, cfg.y ?? 0, cfg.z ?? 0);
+    mirrorGroup.worldToLocal(lookScratch);
+    headScratch.set(0, model.size.y * 0.72, 0);
+    lookScratch.sub(headScratch);
+
+    const flat = Math.hypot(lookScratch.x, lookScratch.z) || 1e-6;
+    const yaw = Math.atan2(lookScratch.x, lookScratch.z);
+    const pitch = Math.atan2(lookScratch.y, flat);
+    const amount = clamp(cfg.amount ?? 1, 0, 1);
+    const aim = {
+      x: clamp((yaw / MAX_LOOK_YAW) * amount, -1, 1),
+      y: clamp((pitch / MAX_LOOK_PITCH) * amount, -1, 1),
+    };
+
+    // Turning the whole body is what a person does to look behind themselves; a
+    // half-strength turn keeps the feet planted while the body leans round. The
+    // turn is *stored*, so switching 体も向ける off stops updating it and the body
+    // stays where it is instead of snapping back to the front.
+    if (cfg.turnBody) {
+      cfg.bodyYawDeg = round2(clamp(yaw, -MAX_LOOK_YAW * 1.6, MAX_LOOK_YAW * 1.6) * amount * 0.7 / DEG);
+    }
+    container.rotation.y = (cfg.bodyYawDeg ?? 0) * DEG;
+
+    // Keep the gaze too, for the moment the mode is switched off.
+    cfg.freeze = aim;
+    return aim;
+  }
+
+  // -------------------------------------------- history, 口パク, gacha
+  const history = new History({ limit: 120 });
+  /** How far the mouth is open right now because of the 口パク animation. */
+  let flapMouth = 0;
+  const mouthFlap = createMouthFlap({
+    onLevel: (level) => {
+      flapMouth = levelToMouth(level, state.mouthFlap?.mouthGain ?? 1);
+      app.needsRender = true;
+    },
+    onEnd: () => { flapMouth = 0; app.needsRender = true; },
+  });
+
+  let rig = null;
+  // The gizmo always belongs to whichever camera is currently active.
+  view.onCameraChange = (camera, controls) => {
+    if (!rig) return;
+    rig.controls.camera = camera;
+    rig.orbit = controls;
+  };
+
+  rig = new Rig({
+    bones: model.bones,
+    scene,
+    camera: view.camera,
+    domElement: canvas,
+    orbit: view.controls,
+    pickTargets: () => [model.parts.eyeMesh, model.parts.mouthMesh, ...model.parts.body],
+    onChange: () => {
+      capturePose();
+      app.panel?.onRigChanged?.();
+      app.needsRender = true;
+    },
+  });
+
+  // The ground, the contact shadow and the gizmo are scenery, not character, and
+  // the two face plates are the *front half* of the body: their open edge would be
+  // picked up as a silhouette and draw a line across the face, while contributing
+  // nothing, since they sit exactly on the body.
+  //
+  // The outline hulls must stay out too, and that one is easy to miss: they are
+  // extra copies of every mesh sitting in the scene graph, and this pass swaps the
+  // material of every visible mesh. A hull is an expanded *back-face* shell, so
+  // handing it a plain front-side label material puts an expanded copy of the
+  // whole character in front of itself - it then paints its own label over
+  // everything, and every leaf edge against the body is inked as if the body were
+  // empty paper. That is exactly the faint dotted line along the leaves' bases.
+  function outlineExclusion() {
+    return [
+      ground,
+      look.contact,
+      ...clip.guides,
+      rig.helper,
+      model.parts.eyeMesh,
+      model.parts.mouthMesh,
+      ...(model.parts.hairMesh ? [model.parts.hairMesh] : []),
+      ...snotMeshes,
+      ...hatMeshes,
+      ...styles.outlineMeshes,
+    ];
+  }
+  outline.exclude(outlineExclusion());
+  refreshOutlineExclusion = () => outline.exclude(outlineExclusion());
+  // The wall is *not* excluded: it is handed to the pass as an occluder, so the
+  // labels behind it are culled as well. Excluded, the leaves and the nose
+  // buried in the wall kept their outlines, because the labels are read before
+  // its depth is applied.
+  outline.occlude(clip.occluders);
+
+  /**
+   * The parts the screen-space pass draws, with the label each one writes.
+   *
+   * The leaves are the hard case: a hull cannot outline a shell that thin. They
+   * are drawn with the LEAF label, which means the pass inks them where they meet
+   * the paper and where they fold against each other, but *not* where they run
+   * into the body - a line there reads as the leaf sinking into the body, and the
+   * original artwork has none (the leaf simply passes behind the body).
+   *
+   * The nose joins them in the line-art styles, with the NOSE label, so its ring
+   * against the body *is* drawn: it is the only thing that shows the nose in a
+   * line drawing. In the shaded styles it is left out entirely, and its hull then
+   * keeps only the part that pokes out of the silhouette - which is what the
+   * original artwork does (the nose reads by its own colour there).
+   *
+   * @param {string} style
+   * @returns {Array<[import('three').Object3D, number]>}
+   */
+  function screenParts(style) {
+    const parts = (model.parts.leaves ?? []).map((mesh) => [mesh, LEAF_LABEL]);
+    if (isLineStyle(style) && model.parts.noseMesh) parts.push([model.parts.noseMesh, NOSE_LABEL]);
+    return parts;
+  }
+
+  /**
+   * The label every part writes (see src/outline.js).
+   *
+   * The leaves and the nose are the outlined ones. The body's label is the one
+   * real choice here, and it is exposed as `render.leafBodyLine`:
+   *
+   *  - `BEHIND` (the default) makes the body count as paper, so each leaf is
+   *    outlined where it emerges from the body. In a line drawing the two are
+   *    paper on paper, so without this the leaves and the body merge into one
+   *    white shape with no way to tell them apart - which is the greater evil.
+   *  - `SOLID` makes the body block that line, so a leaf simply passes behind it.
+   *    Cleaner where it works, but the leaves lose their outline there.
+   *
+   * Everything else - the feet, the hands, the props - is *merely behind*, so it
+   * stays BEHIND either way: it still hides what is behind it, but a leaf lying
+   * across it keeps its outline.
+   */
+  function outlineLabels(parts) {
+    const labels = new Map(parts);
+    const bodyLabel = state.render.leafBodyLine === false ? SOLID_LABEL : BEHIND_LABEL;
+    for (const mesh of model.parts.body ?? []) {
+      if (labels.has(mesh)) continue;
+      labels.set(mesh, model.parts.kinds?.get(mesh) === 'blb' ? bodyLabel : BEHIND_LABEL);
+    }
+    return [...labels];
+  }
+
+  /** Point both outline systems at the right parts for `method`. */
+  function applyOutlineMethod(style, method = state.render.outlineMethod) {
+    const hybrid = method === 'screen';
+    const parts = hybrid ? screenParts(style) : [];
+    styles.setHullHidden(parts.map(([mesh]) => mesh));
+    outline.only(parts.length ? outlineLabels(parts) : null);
+  }
+
+  /**
+   * Run `fn` with the outline pass set up for another style, then put the
+   * on-screen configuration back. Both the hull/screen split and which parts the
+   * screen pass may see are shared by the live view and every offscreen render,
+   * so a pass that draws another style has to say so - the SVG and the "lines
+   * only" PNG are ink through and through, so they always want the line-art
+   * arrangement (a hull cannot outline the leaves at all).
+   */
+  async function withOutlineFor({ style, method = 'screen' }, fn) {
+    applyOutlineMethod(style, method);
+    try {
+      return await fn();
+    } finally {
+      applyOutlineMethod(state.render.style);
+    }
+  }
+
+  Object.assign(app, {
+    renderer, scene, camera: view.camera, view, styles, face, rig, animator,
+    key, ambient, ground, environment, halfHeight, outline, clip,
+    look, backdrop, history, mouthFlap, propRoot, mirrorGroup, container,
+    get captionFont() { return captionFont; },
+    get backdropImage() { return backdropImage; },
+  });
+
+  const recorder = createRecorder(canvas);
+
+  // -------------------------------------------------------------- the panel
+  const panel = buildPanel(app, document.getElementById('panel'));
+  app.panel = panel;
+
+  // On a phone the panel is a bottom sheet. Its grab handle drags the sheet's top
+  // edge up and down, so the menu can be pulled out of the way of the character.
+  const panelEl = document.getElementById('panel');
+  const panelGrip = document.createElement('div');
+  panelGrip.className = 'panel-grip';
+  panelGrip.setAttribute('aria-hidden', 'true');
+  panelEl.prepend(panelGrip);
+  installSheetDrag(panelEl, panelGrip);
+
+  /** Drag on the phone sheet's handle; a no-op anywhere else. */
+  function installSheetDrag(sheet, grip) {
+    const phone = window.matchMedia('(max-width: 768px), (max-height: 500px) and (pointer: coarse)');
+    let drag = null;
+    const minHeight = () => Math.round(window.innerHeight * 0.16);
+    const maxHeight = () => Math.round(window.innerHeight * 0.94);
+
+    grip.addEventListener('pointerdown', (event) => {
+      if (!phone.matches) return;
+      const rect = sheet.getBoundingClientRect();
+      // Start from the height the sheet has now, so the first move does not jump.
+      drag = { id: event.pointerId, fromY: event.clientY, startH: rect.height };
+      sheet.style.maxHeight = 'none';
+      sheet.style.height = `${rect.height}px`;
+      try { grip.setPointerCapture(event.pointerId); } catch { /* capture is optional */ }
+      event.preventDefault();
+    });
+    grip.addEventListener('pointermove', (event) => {
+      if (!drag || event.pointerId !== drag.id) return;
+      // Dragging up grows the sheet (its top edge climbs); dragging down shrinks it.
+      const next = drag.startH + (drag.fromY - event.clientY);
+      sheet.style.height = `${Math.min(maxHeight(), Math.max(minHeight(), next))}px`;
+      event.preventDefault();
+    });
+    const end = (event) => {
+      if (!drag || (event && event.pointerId !== drag.id)) return;
+      drag = null;
+    };
+    grip.addEventListener('pointerup', end);
+    grip.addEventListener('pointercancel', end);
+
+    // A window grown to the desktop layout keeps the side column: drop the height
+    // the drag had fixed, so the panel goes back to filling the screen.
+    const clearOnDesktop = () => {
+      if (phone.matches) return;
+      sheet.style.height = '';
+      sheet.style.maxHeight = '';
+    };
+    window.addEventListener('resize', clearOnDesktop);
+  }
+
+  Object.assign(app.actions, {
+    capturePose,
+    applyState,
+    refresh,
+    loadHatModel,
+    patch: (scope, value) => applyState({ [scope]: value }, { scope }),
+    syncPanel: () => app.panel?.sync(),
+    imageCanvas: (options) => captureImage(options),
+    savePNG: (options = {}) => savePNG(options),
+    copyPNG: () => copyPNG(),
+    faceMapCanvas: (kind) => faceMapCanvas(kind),
+    saveFaceMap: (kind) => saveFaceMap(kind),
+    addGionStamp: (item) => addGionStamp(item),
+    saveSettings: () => exporter.downloadText(JSON.stringify(state, null, 2), `bluebey-settings-${exporter.timestamp()}.json`),
+    loadSettings: (text) => {
+      const data = JSON.parse(text);
+      applyState(data, { full: true, sync: true });
+      toast('設定を読み込みました');
+    },
+    toggleRecording: () => toggleRecording(recorder),
+    resetPose: () => {
+      rig.reset();
+      state.pose = { bones: {}, root: [0, 0, 0] };
+      applyState({}, { scope: 'pose', sync: true });
+      toast('ポーズをリセットしました');
+    },
+    setCameraPreset: (id) => setCameraPreset(id),
+    undo: () => restoreFromHistory('undo'),
+    redo: () => restoreFromHistory('redo'),
+    applyTheme: (id) => applyTheme(id),
+    rollGacha: (seed) => rollGacha(seed),
+    copyShareLink: () => copyShareLink(),
+    shareImage: (network) => shareImage(network),
+    toggleMouthFlap: () => toggleMouthFlap(),
+    storyThumbs: () => storyThumbs,
+    toggleCamera: (on) => toggleCamera(on),
+    addProp: (kind) => addProp(kind),
+    addStoryPanel: () => addStoryPanel(),
+    saveStory: (options) => saveStory(options),
+    loadFromHash: () => loadFromHash(),
+    toggleAnimation: () => {
+      state.anim.mode = (state.anim.mode ?? 'off') === 'off' ? 'idle' : 'off';
+      app.animator.reset();
+      panel.sync();
+      app.needsRender = true;
+    },
+  });
+
+  installKeys();
+  installTopbar();
+  window.addEventListener('focus', () => { app.needsRender = true; });
+  document.addEventListener('visibilitychange', () => {
+    if (!document.hidden) app.needsRender = true;
+  });
+
+  {
+    const buffer = renderer.getDrawingBufferSize(new THREE.Vector2());
+    console.log(
+      '[bluebey] ready'
+      + ` canvas=${canvas.clientWidth}x${canvas.clientHeight}`
+      + ` buffer=${Math.round(buffer.x)}x${Math.round(buffer.y)}`
+      + ` quality=${quality}`
+      + ` dpr=${window.devicePixelRatio}`
+      + ` model=${model.size.toArray().map((v) => v.toFixed(2)).join(',')}`
+      + ` camera=${view.camera.position.toArray().map((v) => v.toFixed(2)).join(',')}`,
+    );
+  }
+  window.addEventListener('resize', () => resizeViewport());
+
+  applyState({}, {});
+  await loadFromHash();
+  history.reset(state);
+  rig.select('master', { silent: true });
+  panel.sync();
+
+  let last = performance.now();
+  let faceSignature = '';
+  let slowFrames = 0;
+  let framesSinceFps = 0;
+  let fpsAt = performance.now();
+  const warmupUntil = performance.now() + 2500;
+
+  let snotSignature = null;
+  const loop = (now) => {
+    const rawDelta = now - last;
+    const dt = Math.min(0.05, rawDelta / 1000);
+    last = now;
+
+    animator.update(dt);
+
+    const posed = animator.pose(state.pose);
+    if (!rig.controls.dragging) rig.applyPose(posed);
+    container.position.set(posed.root[0] ?? 0, posed.root[1] ?? 0, posed.root[2] ?? 0);
+    updateKeyLight();
+    // The contact blob rides under the character, so it never lingers at the origin.
+    look.contact.position.x = container.position.x;
+    look.contact.position.z = container.position.z;
+
+    applyHat(state.face.hat?.kind ?? 'none');
+
+    if (snotBubble) {
+      const snot = state.face.eyes.snot;
+      const sig = `${snot.enabled}|${snot.size}|${snot.offsetX}|${snot.offsetY}|${snot.offsetZ}|${snot.color}`;
+      if (sig !== snotSignature) {
+        snotSignature = sig;
+        applySnotBubble(snotBubble, snot);
+        app.needsRender = true;
+      }
+      // The sleeping breath swells the bubble; `snotScale` is 1 otherwise.
+      if (snotBubble.visible) {
+        const scale = (snotBubble.userData.baseScale ?? 1) * (animator.snotScale ?? 1);
+        if (snotBubble.scale.x !== scale) {
+          snotBubble.scale.setScalar(scale);
+          app.needsRender = true;
+        }
+      }
+    }
+
+    const params = buildFaceParams();
+    const signature = JSON.stringify(params);
+    if (signature !== faceSignature) {
+      faceSignature = signature;
+      face.setParams(params, faceMode());
+      app.needsRender = true;
+    }
+    if (face.flush(now, 26)) app.needsRender = true;
+
+    view.controls.update(dt);
+
+    const animating = (state.anim.mode ?? 'off') !== 'off'
+      || state.anim.lookAround
+      || animator.blinkPhase >= 0
+      || recorder?.recorder?.state === 'recording';
+    // Always paint for the first couple of seconds, so a missed "needs render"
+    // can never leave an empty window staring back at the user.
+    if (app.needsRender || animating || view.controlsChanged || now < warmupUntil) {
+      app.needsRender = false;
+
+      // The screen-space outline needs one extra pass *before* the scene, so it
+      // wraps the normal render instead of following it. `outline.render` paints
+      // the scene itself (ink and all), so there must be no second render after
+      // it - that would clear the canvas and wipe the ink straight back off.
+      const wantsScreen = state.render.outlineMethod === 'screen' && state.render.outline !== false;
+      outline.render(() => renderer.render(scene, view.camera), {
+        enabled: wantsScreen,
+        camera: view.camera,
+        color: state.render.outlineColor,
+        // The label buffer is `quality`x the CSS size, so the radius has to carry
+        // that factor to keep the leaf/nose line a constant width in CSS pixels -
+        // otherwise it halves on a HiDPI display and jumps when the app lowers
+        // the quality on slow frames.
+        radius: outlineRadiusFor(view, state, quality),
+      });
+      view.controlsChanged = false;
+      framesSinceFps += 1;
+
+      // Hardware that cannot keep up gets a smaller drawing buffer instead of a
+      // slideshow (a very large buffer can look like "nothing ever appears").
+      if (rawDelta > 90) slowFrames += 1;
+      else slowFrames = 0;
+      if (slowFrames >= 8 && quality > QUALITY_FLOOR + 0.01 && autoReductions < 3) {
+        autoReductions += 1;
+        quality = Math.max(QUALITY_FLOOR, Math.round((quality - 0.25) * 100) / 100);
+        renderer.setPixelRatio(quality);
+        slowFrames = 0;
+        console.warn('[bluebey] frames are slow, lowering the drawing quality to', quality);
+      }
+    }
+
+    if (now - fpsAt > 1000) {
+      const fps = Math.round((framesSinceFps * 1000) / (now - fpsAt));
+      framesSinceFps = 0;
+      fpsAt = now;
+      const buffer = renderer.getDrawingBufferSize(new THREE.Vector2());
+      panel.setStats?.(
+        `描画 ${Math.round(buffer.x)}×${Math.round(buffer.y)}（画質 ${quality.toFixed(2)}×）／約 ${fps}fps`
+        + (autoReductions > 0 ? '　※重いので自動で軽くしました' : ''),
+      );
+    }
+
+    requestAnimationFrame(loop);
+  };
+  requestAnimationFrame(loop);
+
+  hideLoading();
+  app.ready = true;
+  window.__bluebeyReady = true;
+  window.__bluebey = {
+    app,
+    state,
+    set: (patch) => applyState(patch),
+    face: (patch) => applyState({ face: patch }),
+    pose: (patch) => applyState({ pose: patch }),
+    view: (patch) => applyState({ view: patch }),
+    render: (patch) => applyState({ render: patch }),
+    selectBone: (name) => rig.select(name),
+    preset: (id) => app.actions.applyFacePreset?.(id),
+    posePreset: (id) => app.actions.applyPosePreset?.(id),
+    png: (options) => captureImage(options),
+  };
+
+  /* ------------------------------------------------------------- internals */
+
+  /** The camera the outline pass should use (kept in step with the view rig). */
+  function outlineCamera() {
+    return view.camera;
+  }
+
+  /** The gizmo/frame loop re-reads the bones, so store them back into state. */
+  function capturePose() {
+    const pose = rig.getPose();
+    state.pose.bones = pose.bones;
+  }
+
+  function buildFaceParams() {
+    const eyes = state.face.eyes;
+    const openScale = animator.eyeOpen;
+    const drift = animator.look;
+
+    const eyeFor = (key) => {
+      const own = eyes[key];
+      // A shut artwork (1〜3線 / 3の目 / わらう) does not blink: its lid stays where
+      // the user put it, so the blink animation must not scale its `open`.
+      const shutArt = own.shape === 'three' || own.shape === 'arch'
+        || (typeof own.shape === 'string' && own.shape.startsWith('line'));
+      return {
+        // Which artwork this eye shows (ふつう / 1〜3線 / 3の目 / わらう / ハート).
+        // The renderer chooses the shut artwork from this, so it must travel.
+        shape: own.shape,
+        open: clamp01(own.open * (shutArt ? 1 : openScale)),
+        lookX: clamp(own.lookX + drift.x + aimOffset.x, -1, 1),
+        lookY: clamp(own.lookY + drift.y + aimOffset.y, -1, 1),
+        closed: own.closed,
+        closedLines: own.closedLines,
+        irisShape: own.irisShape,
+        threeFlip: own.threeFlip === true,
+        // 白目 / 光彩 are per eye and tri-state: `null` means "the usual" (open
+        // eyes show their white; the shared 光彩 switch decides the glint).
+        white: own.white ?? null,
+        highlight: own.highlight ?? null,
+        // 形の大きさ scales the drawn eye shape (the shut lines, the 3, the arch)
+        // and the heart.
+        shapeScale: own.shapeScale ?? 1,
+        // The tears are per eye, and only drawn when their own switch is on.
+        tearOn: own.tearOn === true,
+        tear: own.tear ?? 0.3,
+        tearY: own.tearY ?? 0,
+        // Where this eye sits, and where its teardrop hangs, can be nudged one
+        // side at a time (an asymmetric face).
+        eyeX: own.eyeX ?? 0,
+        tearX: own.tearX ?? 0,
+        tearTilt: own.tearTilt ?? 0,
+        // The lower lid can differ per eye (falling back to the shared value).
+        lowerLid: own.lowerLid ?? eyes.lowerLid ?? 0,
+        // The rest of the lids are per eye too now (null = use the shared value).
+        lidShape: own.lidShape ?? null,
+        lidWidth: own.lidWidth ?? null,
+        lidTilt: own.lidTilt ?? null,
+        lashes: own.lashes ?? null,
+        lashAngle: own.lashAngle ?? null,
+        lashPos: own.lashPos ?? null,
+      };
+    };
+
+    // Lip-sync and a look-at target both drive the mouth and the eyes without
+    // being part of the saved expression.
+    const mouth = { ...state.face.mouth };
+    if (flapMouth > 0.001) mouth.open = Math.max(mouth.open ?? 0, flapMouth);
+    // A talking mouth has no tongue sticking out. This follows the whole flap
+    // session rather than the opening: the envelope dips through zero between
+    // syllables, and keying it off the opening made the tongue flicker back into
+    // view on every closed frame.
+    if (mouthFlap.running) mouth.tongue = 0;
+
+    return {
+      eyes: {
+        // Everything shared by both eyes (colours, iris scale, brows, glasses,
+        // ...) has to travel with the per-eye values, because the face renderer
+        // reads them from here. Spreading `...eyes` is also what makes the face
+        // redraw when one of them (e.g. `glasses`) changes: the loop compares the
+        // JSON of this object against the last one it drew.
+        ...eyes,
+        left: eyeFor('left'),
+        right: eyeFor('right'),
+      },
+      mouth,
+    };
+  }
+
+  function faceMode() {
+    return isLineStyle(state.render.style) ? 'line' : 'paint';
+  }
+
+  function applyState(patch, { silent = false, full = false, scope = 'all', sync = false } = {}) {
+    if (full) {
+      const fresh = defaultState();
+      applyPatch(fresh, patch);
+      for (const key of Object.keys(fresh)) state[key] = fresh[key];
+    } else {
+      applyPatch(state, patch);
+    }
+    refresh(scope);
+    app.needsRender = true;
+    if (sync && !silent) app.panel?.sync();
+  }
+
+  /** Re-apply the parts of the state named by `scope` ('all' by default). */
+  function refresh(scope = 'all') {
+    app.needsRender = true;
+    // Where the eyes aim depends on where the character is, so it is recomputed
+    // whenever anything that moves could change.
+    if (scope === 'all' || scope === 'view' || scope === 'pose' || scope === 'face' || scope === 'lookAt') {
+      aimOffset = applyLookAt();
+    }
+    if (scope === 'all' || scope === 'render') {
+      styles.setStyle(state.render.style);
+      styles.setPaper(state.render.paper);
+      styles.setOutlineWidth(state.render.outlineWidth);
+      styles.setOutlineColor(state.render.outlineColor);
+      styles.setOutlineEnabled(state.render.outline);
+      // Which parts each outline method draws, and which hulls stand down for it
+      // (the leaves, plus the nose in a line drawing). See `applyOutlineMethod`.
+      applyOutlineMethod(state.render.style);
+
+      const line = isLineStyle(state.render.style);
+      // Shadows, reflections, body colours and mirroring are the look module's
+      // job, so that they are applied together and stay consistent.
+      look.apply(state.render, { line });
+      applyProps();
+      // 見えない壁: the one setting that changes what is drawn rather than how.
+      applyWalls();
+
+      const azimuth = state.render.lightAzimuth * DEG;
+      const elevation = state.render.lightElevation * DEG;
+      const distance = model.size.y * 3;
+      keyOffset.set(
+        Math.cos(elevation) * Math.sin(azimuth) * distance,
+        Math.sin(elevation) * distance,
+        Math.cos(elevation) * Math.cos(azimuth) * distance,
+      );
+      updateKeyLight();
+      key.intensity = state.render.lightIntensity;
+      ambient.intensity = state.render.ambient;
+      key.color.set(state.render.lightColor ?? '#ffffff');
+      ambient.color.set(state.render.ambientColor ?? '#ffffff');
+      renderer.toneMappingExposure = state.render.exposure;
+      applyBackground();
+      renderer.shadowMap.needsUpdate = true;
+    }
+
+    if (scope === 'all' || scope === 'view') {
+      applyBackdrop();
+      // The renderer's *clear* depends on the background mode as well: a preset
+      // photo, a loaded picture and the camera all show through a transparent
+      // clear (see `backgroundOf`). Without this, switching the background left
+      // the canvas clearing to the old opaque colour and hid the layer that had
+      // just been set up behind it - which is why the background only appeared
+      // after some other change (adding a prop) re-ran this.
+      applyBackground();
+    }
+    if (scope === 'all' || scope === 'view' || scope === 'caption' || scope === 'gion') redrawCaption();
+
+    // A style change flips the face between paint and ink (`faceMode`), so the
+    // 'render' scope has to rebuild the face too - otherwise the old drawing
+    // stays on the plates until something else (a blink) happens to dirty them.
+    if (scope === 'all' || scope === 'face' || scope === 'render') {
+      face.setSource('eyes', state.face.eyes.source);
+      face.setSource('mouth', state.face.mouth.source);
+      face.setParams(buildFaceParams(), faceMode());
+    }
+
+    if (scope === 'all' || scope === 'view') {
+      view.setProjection(state.view.projection);
+      view.apply(state.view, model.size);
+      view.setAutoRotate(state.view.autoRotate, state.view.autoRotateSpeed);
+    }
+
+    if (scope === 'all' || scope === 'pose') {
+      rig.applyPose(state.pose);
+      const root = state.pose.root ?? [0, 0, 0];
+      container.position.set(root[0] ?? 0, root[1] ?? 0, root[2] ?? 0);
+    }
+
+    // Every user-driven change funnels through here, so this is the one place
+    // that has to remember a step for undo. Undo itself sets `suspendHistory`.
+    if (!suspendHistory) history.push(state, SCOPE_LABELS[scope] ?? scope);
+  }
+
+  /** Paint the backdrop bitmap behind the model, for an export. */
+  function backdropSource() {
+    if (state.view.background === 'camera') {
+      const video = backdrop.el?.querySelector?.('video');
+      return video && video.readyState >= 2 ? video : null;
+    }
+    return backdropImage;
+  }
+
+  /**
+   * Freeze the current camera frame into a canvas.
+   *
+   * The AR preview is a live video, so by the time the PNG finishes encoding the
+   * feed has moved on and the saved picture no longer matches what was on screen
+   * when the shutter was pressed. Snapping the frame first keeps the two in step.
+   */
+  function snapshotBackdrop() {
+    if (state.view.background !== 'camera') return undefined;
+    const video = backdrop.el?.querySelector?.('video');
+    if (!video || video.readyState < 2 || !video.videoWidth || !video.videoHeight) return undefined;
+    const frame = document.createElement('canvas');
+    frame.width = video.videoWidth;
+    frame.height = video.videoHeight;
+    try {
+      frame.getContext('2d').drawImage(video, 0, 0);
+    } catch {
+      return undefined;
+    }
+    return frame;
+  }
+
+  /**
+   * Paint the backdrop *behind* the rendered model.
+   *
+   * WHY the scratch canvas and `destination-over`: `ctx` already holds the
+   * character on a transparent field (see `exporter.renderStill`), so drawing the
+   * photo straight onto it with the default `source-over` painted OVER the
+   * character and the export came out as a bare photograph. Building the scene in
+   * a scratch canvas first also lets the dark veil sit on top of the photo, which
+   * it could not do once the photo had been slipped underneath.
+   */
+  function drawBackdropInto(ctx, width, height, scale, sourceOverride) {
+    const source = sourceOverride ?? backdropSource();
+    if (!source) return false;
+    const style = backdropStyle(state.view);
+    if (!style.visible) return false;
+    const iw = source.naturalWidth || source.videoWidth || source.width || 0;
+    const ih = source.naturalHeight || source.videoHeight || source.height || 0;
+    if (!iw || !ih) return false;
+
+    const back = document.createElement('canvas');
+    back.width = width;
+    back.height = height;
+    const paint = back.getContext('2d');
+
+    // The same CSS the live layer uses: blur (bleeding past the edges), the fit,
+    // the mirror and the scale/offset transform. The offset is in the element's
+    // own (pre-scale) space, so it is multiplied by `scale` here, exactly as
+    // `transform: scale(s) translate(...)` does in CSS.
+    const bleed = style.blur > 0 ? style.blur * 2 : 0;
+    const w = width + bleed * 2;
+    const h = height + bleed * 2;
+    paint.save();
+    if (style.blur > 0) paint.filter = `blur(${style.blur}px)`;
+    paint.translate(
+      width / 2 + style.scale * style.offset.x * width,
+      height / 2 + style.scale * style.offset.y * height,
+    );
+    paint.scale(style.scale * (style.mirror ? -1 : 1), style.scale);
+    paint.translate(-width / 2, -height / 2);
+
+    if (style.backgroundSize === 'cover' || style.backgroundSize === 'contain') {
+      const fit = style.backgroundSize === 'cover'
+        ? Math.max(w / iw, h / ih)
+        : Math.min(w / iw, h / ih);
+      const dw = iw * fit;
+      const dh = ih * fit;
+      paint.drawImage(source, (w - dw) / 2 - bleed, (h - dh) / 2 - bleed, dw, dh);
+    } else if (style.backgroundSize === 'auto') {
+      // CSS repeats the picture at its own size and anchors the grid on the box
+      // centre (`background-position: center`). A canvas pattern instead repeats
+      // at the image's *intrinsic* size, which is `scale`x too small for a
+      // canvas that is `scale`x the CSS box - so the export tiled `scale`x more
+      // densely than the preview. Paint the tiles by hand at `scale`, on a grid
+      // centred on the box centre; the CTM above already carries the offset /
+      // backgroundScale / mirror, so the tiles must not re-apply them.
+      const tw = iw * scale;
+      const th = ih * scale;
+      // Cover only the user-space area the visible canvas needs (widened by the
+      // blur bleed), mapped back through the CTM. The CTM holds no rotation, so
+      // its `a`/`d` scale factors are enough to invert it for the bounds.
+      const matrix = paint.getTransform();
+      const [left, right] = [
+        (-bleed - matrix.e) / matrix.a,
+        (width + bleed - matrix.e) / matrix.a,
+      ].sort((a, b) => a - b);
+      const [top, bottom] = [
+        (-bleed - matrix.f) / matrix.d,
+        (height + bleed - matrix.f) / matrix.d,
+      ].sort((a, b) => a - b);
+      const cx = width / 2;
+      const cy = height / 2;
+      const firstCol = Math.floor((left - cx) / tw);
+      const lastCol = Math.ceil((right - cx) / tw);
+      const firstRow = Math.floor((top - cy) / th);
+      const lastRow = Math.ceil((bottom - cy) / th);
+      for (let row = firstRow; row <= lastRow; row++) {
+        for (let col = firstCol; col <= lastCol; col++) {
+          paint.drawImage(source, cx + col * tw - tw / 2, cy + row * th - th / 2, tw, th);
+        }
+      }
+    } else {
+      paint.drawImage(source, -bleed, -bleed, w, h);
+    }
+    paint.restore();
+
+    if (style.darken > 0) {
+      paint.fillStyle = `rgba(0, 0, 0, ${style.darken})`;
+      paint.fillRect(0, 0, width, height);
+    }
+
+    ctx.save();
+    ctx.globalCompositeOperation = 'destination-over';
+    ctx.drawImage(back, 0, 0);
+    ctx.restore();
+    return true;
+  }
+
+  /** Load (or clear) the backdrop bitmap that the exports composite. */
+  function applyBackdrop() {
+    const viewState = state.view;
+    backdrop.apply(viewState);
+    // The AR shutter rides the viewport, not the panel, so it shows/hides here.
+    if (app.arShutter) app.arShutter.hidden = viewState.background !== 'camera';
+    // The touch hint would sit under the shutter, so it steps aside in AR mode.
+    const touchHint = document.getElementById('viewport-hint-touch');
+    if (touchHint) touchHint.style.visibility = viewState.background === 'camera' ? 'hidden' : '';
+    const url = viewState.background === 'preset'
+      ? assetUrl(viewState.backgroundPreset, 'background')
+      : viewState.background === 'image'
+        ? viewState.backgroundImage
+        : viewState.background === 'effect'
+          ? backdrop.effectUrl(viewState.backgroundEffect)
+          : null;
+    if (!url) {
+      if (viewState.background !== 'camera') backdropImage = null;
+      return;
+    }
+    if (backdropImage?.dataset?.url === url) return;
+    const image = new Image();
+    image.dataset.url = url;
+    image.onload = () => { backdropImage = image; app.needsRender = true; };
+    image.onerror = () => { backdropImage = null; };
+    image.src = url;
+  }
+
+  function applyBackground() {
+    exporter.applyBackground(renderer, backgroundOf(state.view));
+  }
+
+  async function captureImage({ scale = 2, transparent = null, lines = null } = {}) {
+    // Freeze the live camera frame up front, so the saved picture holds the frame
+    // that was on screen when the button was pressed (see `snapshotBackdrop`).
+    const frozenBackdrop = snapshotBackdrop();
+    return withHiddenHelpers(async () => {
+      // The face texture is redrawn on a throttle (every 26 ms) while the render
+      // loop runs, so a capture taken straight after a change could still hold the
+      // *previous* expression. That is what made a comic panel occasionally show
+      // its neighbour's face: force the pending redraw through first.
+      face.flush(performance.now(), 0);
+      const wantsTransparent = transparent === true
+        || (transparent == null && state.view.background === 'transparent');
+      // A transparent export of a line drawing drops the paper as well. The whole
+      // point of transparent line art is to lay the lines over another picture,
+      // and a white silhouette would just hide it - so this renders the `outline`
+      // style, which draws the ink and leaves the body out. A shaded style keeps
+      // its body, because there it is an ordinary cut-out of the character.
+      const wantsLines = lines ?? (wantsTransparent && isLineStyle(state.render.style));
+      const drawn = wantsLines ? 'outline' : state.render.style;
+      const background = wantsTransparent ? { mode: 'transparent' } : backgroundOf(state.view);
+
+      const onScreen = styles.style;
+      if (drawn !== onScreen) styles.setStyle(drawn);
+      let result;
+      try {
+        result = await withOutlineFor({ style: drawn, method: wantsLines ? 'screen' : state.render.outlineMethod }, () => exporter.capturePNG({
+          renderer,
+          scene,
+          camera: view.camera,
+          outline,
+          outlineOptions: {
+            enabled: state.render.outlineMethod === 'screen' && state.render.outline !== false,
+            color: state.render.outlineColor,
+            // The export's label buffer is `scale`x the one on screen (renderStill
+            // sets it to the output size), so its radius grows with it to keep the
+            // line's weight relative to the model the same as in the preview.
+            radius: outlineRadiusFor(view, state, scale),
+              },
+          width: canvas.clientWidth || window.innerWidth,
+          height: canvas.clientHeight || window.innerHeight,
+          restore: () => { resizeViewport(); applyStateBackgroundAgain(); },
+        }, { scale, background }));
+      } finally {
+        if (drawn !== onScreen) styles.setStyle(onScreen);
+      }
+
+      // The backdrop, the 擬音 stamps and the bubbles live in DOM layers behind
+      // and above the WebGL canvas, so an export has to paint them in itself - in
+      // the same order as the screen: backdrop, then stamps, then bubbles.
+      const ctx = result.canvas.getContext('2d');
+      // The canvas' pixels-per-CSS-pixel: the same factor the 擬音 stamps and the
+      // bubbles are drawn at below. The backdrop's tiled fit needs it too, so its
+      // tiles come out the size the preview's CSS paints them.
+      const base = canvas.clientWidth || 1280;
+      const outputScale = result.width / Math.max(1, base);
+      let changed = false;
+      if (!wantsTransparent && drawBackdropInto(ctx, result.width, result.height, outputScale, frozenBackdrop)) changed = true;
+      const gionItems = state.gion?.items ?? [];
+      if (gionItems.length) {
+        for (const item of gionItems) await ensureGionSheet(item.sheet);
+        drawGion(ctx, gionItems, gionImages, {
+          width: result.width,
+          height: result.height,
+          scale: outputScale,
+        });
+        changed = true;
+      }
+      const bubbles = CAPTION_KEYS.map((key) => state[key]).filter((caption) => caption?.enabled);
+      if (bubbles.length) {
+        await ensureCaptionFont();
+        for (const caption of bubbles) {
+          drawCaption(ctx, caption, {
+            width: result.width,
+            height: result.height,
+            scale: outputScale,
+            font: captionFont,
+          });
+        }
+        changed = true;
+      }
+      if (changed) result.blob = await exporter.canvasToBlob(result.canvas);
+      return result;
+    });
+  }
+
+  function applyStateBackgroundAgain() {
+    exporter.applyBackground(renderer, backgroundOf(state.view));
+  }
+
+  async function savePNG({
+    scale = state.render.pngScale ?? 2,
+    transparent = state.render.pngTransparent ? true : undefined,
+  } = {}) {
+    const result = await captureImage({ scale, transparent });
+    exporter.downloadBlob(result.blob, `bluebey-${exporter.timestamp()}.png`);
+    toast(`PNGを書き出しました（${result.width}×${result.height}）`);
+  }
+
+  /**
+   * The canvas the 下地 export paints on, *without* downloading it. Kept separate
+   * from `saveFaceMap` so the in-page round-trip check can look at exactly the
+   * pixels the button writes.
+   */
+  function faceMapCanvas(kind) {
+    // A redraw is throttled while the render loop runs (see the loop's
+    // `face.flush(now, 26)`), so force any pending one through first or the map
+    // would show the expression before last.
+    face.flush(performance.now(), 0);
+    return kind === 'hair' ? exporter.buildHairMap(face) : exporter.buildFaceMap(face, kind);
+  }
+
+  async function saveFaceMap(kind) {
+    const canvas = faceMapCanvas(kind);
+    if (!canvas) {
+      toast('このモデルには髪のプレートがありません');
+      return;
+    }
+    const blob = await exporter.canvasToBlob(canvas);
+    const label = kind === 'eyes' ? '目' : kind === 'mouth' ? '口' : '髪';
+    exporter.downloadBlob(blob, `bluebey-face-${kind}-${exporter.timestamp()}.png`);
+    if (kind === 'hair') {
+      toast(`${label}の下地を書き出しました（${canvas.width}×${canvas.height}）。`
+        + 'この画像に前髪を描いて渡してください');
+      return;
+    }
+    toast(`${label}の下地を書き出しました（${canvas.width}×${canvas.height}）。`
+      + `この画像に描いて「${label}の画像を読み込む（PNG）」で読み込めます`);
+  }
+
+  /** Hide the gizmo and helpers so they never end up in an export. */
+  async function withHiddenHelpers(fn) {
+    const wasVisible = rig.helper.visible;
+    const guidesWereVisible = clip.guides.map((guide) => guide.visible);
+    rig.helper.visible = false;
+    clip.guides.forEach((guide) => { guide.visible = false; });
+    try {
+      return await fn();
+    } finally {
+      rig.setGizmoVisible(wasVisible);
+      clip.guides.forEach((guide, index) => { guide.visible = guidesWereVisible[index]; });
+      app.needsRender = true;
+    }
+  }
+
+  async function copyPNG() {
+    const result = await captureImage({ scale: 2 });
+    await exporter.copyCanvasToClipboard(result.canvas);
+    toast('画像をクリップボードにコピーしました');
+  }
+
+  /**
+   * Swap in another GLB.
+   *
+   * Kept as the implementation, but no longer wired to anything: the panel's
+   * 「GLBを差し替える」 button and the window-wide drop handler were removed on
+   * 2026-09-27. Putting the feature back is a button plus an action that calls
+   * this (the model returns on reload, so nothing is saved).
+   */
+  async function replaceModel(file) {
+    if (!file) return;
+    try {
+      const buffer = await exporter.readFileAsArrayBuffer(file);
+      const next = await loadModel(buffer);
+      styles.dispose();
+      rig.dispose();
+      container.remove(model.root);
+      scene.remove(container);
+      container.add(next.root);
+      scene.add(container);
+      app.model = next;
+      toast(`${file.name} を読み込みました（このモデルは再読込で元に戻ります）`);
+      app.needsRender = true;
+    } catch (error) {
+      console.error(error);
+      toast('このファイルは読み込めませんでした');
+    }
+  }
+
+  let recording = false;
+  async function toggleRecording(session) {
+    if (!session) {
+      toast('この環境では録画できません');
+      return;
+    }
+    if (!recording) {
+      if (!state.anim.idle) {
+        state.anim.idle = true;
+        panel.sync();
+      }
+      recording = startRecording(session);
+      toast('録画を開始しました（もう一度押すと停止）');
+    } else {
+      const blob = await stopRecording(session);
+      recording = false;
+      if (blob) {
+        exporter.downloadBlob(blob, `bluebey-animation-${exporter.timestamp()}.webm`);
+        toast('録画を保存しました（WebM）');
+      }
+    }
+    app.panel?.setRecording(recording);
+  }
+
+  function setCameraPreset(id) {
+    const presets = {
+      front: { azimuth: 0, polar: 82 },
+      threeQuarter: { azimuth: 34, polar: 78 },
+      side: { azimuth: 90, polar: 84 },
+      back: { azimuth: 180, polar: 82 },
+      top: { azimuth: 24, polar: 26 },
+    };
+    applyState({ view: presets[id] ?? presets.front }, { scope: 'view', sync: true });
+  }
+
+  function restoreFromHistory(kind) {
+    const snapshot = kind === 'undo' ? history.undo() : history.redo();
+    if (!snapshot) {
+      toast(kind === 'undo' ? 'これ以上戻れません' : 'やり直す操作がありません');
+      return;
+    }
+    suspendHistory = true;
+    try {
+      applyState(snapshot, { full: true, sync: true });
+    } finally {
+      suspendHistory = false;
+    }
+    toast(kind === 'undo' ? '1つ戻しました' : 'やり直しました');
+  }
+
+  function applyTheme(id) {
+    const theme = THEMES.find((item) => item.id === id);
+    if (!theme) return;
+    state.render.theme = id;
+    state.render.colors = { ...theme.colors };
+    applyState({}, { scope: 'render', sync: true });
+  }
+
+  /** おまかせ: a seeded random expression, so it can be reproduced and shared. */
+  function rollGacha(seedText) {
+    const seed = seedText || String(Math.floor(Math.random() * 1e9));
+    const roll = rollAll(seed);
+    suspendHistory = true;
+    try {
+      applyState(roll.patch, { scope: 'all' });
+    } finally {
+      suspendHistory = false;
+    }
+    history.push(state, 'おまかせ');
+    app.panel?.sync();
+    const shown = decodeSeed(seed) ?? String(seed);
+    toast(`おまかせ表情（シード ${shown}）`);
+    return shown;
+  }
+
+  async function copyShareLink() {
+    try {
+      const encoded = await encodeState(state);
+      const link = `${location.origin}${location.pathname}#s=${encoded}`;
+      await navigator.clipboard.writeText(link);
+      toast(`この見た目のリンクをコピーしました（${link.length}文字）`);
+    } catch (error) {
+      console.error(error);
+      toast('リンクを作れませんでした');
+    }
+  }
+
+  /** The tag every shared picture carries. */
+  const SHARE_TAG = '#ぶるべースタジオ';
+
+  /** The studio's own address - the landing page itself, not a state-reproducing
+   *  `#s=` link. Shared alongside the tag so a viewer can find the studio. */
+  function studioUrl() {
+    return `${location.origin}${location.pathname}`;
+  }
+
+  /**
+   * Share the current look as a *picture*, not as a link.
+   *
+   * Neither X nor Facebook's web dialog can attach an image, so the picture goes
+   * through the OS share sheet (`navigator.share` with files) wherever the browser
+   * has one - a single tap on a phone, and it can go to any app. Where there is no
+   * share sheet (most desktops), the PNG is put on the clipboard and the network's
+   * compose window is opened with the tag already typed, so the picture only has to
+   * be pasted. The tag *and* the studio link always travel together: the user asks
+   * for both to be kept when the picture is shared.
+   */
+  async function shareImage(network) {
+    let shot;
+    try {
+      shot = await captureImage({ scale: 1.5, transparent: false });
+    } catch (error) {
+      console.error(error);
+      toast('画像を作れませんでした');
+      return;
+    }
+    const file = new File([shot.blob], `bluebey-${exporter.timestamp()}.png`, { type: 'image/png' });
+
+    if (!network && navigator.canShare?.({ files: [file] })) {
+      try {
+        await navigator.share({ files: [file], text: SHARE_TAG, url: studioUrl() });
+        return;
+      } catch (error) {
+        if (error?.name === 'AbortError') return;
+        console.error(error);
+      }
+    }
+
+    // No share sheet (or a named network): put the picture on the clipboard and
+    // open that network's compose window with the tag and the studio link filled in.
+    let copied = false;
+    try {
+      await exporter.copyCanvasToClipboard(shot.canvas);
+      copied = true;
+    } catch (error) {
+      console.error(error);
+    }
+    const studio = studioUrl();
+    const target = network === 'facebook'
+      ? `https://www.facebook.com/sharer/sharer.php?u=${encodeURIComponent(studio)}`
+      : `https://twitter.com/intent/tweet?text=${encodeURIComponent(SHARE_TAG)}&url=${encodeURIComponent(studio)}`;
+    window.open(target, '_blank');
+    toast(copied
+      ? '画像をコピーしました。開いた画面に貼り付けて投稿してください（Ctrl+V）'
+      : '投稿画面を開きました（この環境では画像を自動で貼り付けできません）');
+  }
+
+  function toggleMouthFlap() {
+    if (mouthFlap.running) {
+      mouthFlap.stop();
+      toast('口パクを止めました');
+      return;
+    }
+    const seconds = mouthFlap.start(state.mouthFlap?.text ?? '', {
+      rate: state.mouthFlap?.rate ?? 1,
+    });
+    toast(`口パクを始めました（音は出ません・約${seconds.toFixed(1)}秒）`);
+  }
+
+  async function toggleCamera(on) {
+    if (!on) {
+      backdrop.stopCamera();
+      applyState({ view: { background: 'solid' } }, { scope: 'view', sync: true });
+      return;
+    }
+    const result = await backdrop.startCamera(state.view.cameraFacing ?? 'environment');
+    if (!result?.ok) {
+      toast(result?.reason ?? 'カメラを使えません');
+      return;
+    }
+    applyState({ view: { background: 'camera' } }, { scope: 'view', sync: true });
+    toast('カメラの映像を背景にしました（そのまま写真に撮れます）');
+  }
+
+  function addProp(kind) {
+    const def = PROP_DEFAULTS[kind] ?? { x: 1.6, y: 0, z: 0, rotX: 0, rotY: 0, rotZ: 0, scale: 1 };
+    state.props.items = [...(state.props.items ?? []), {
+      kind,
+      x: def.x ?? 0,
+      y: def.y ?? 0,
+      z: def.z ?? 0,
+      rotX: def.rotX ?? 0,
+      rotY: def.rotY ?? 0,
+      rotZ: def.rotZ ?? 0,
+      scale: def.scale ?? 1,
+    }];
+    applyState({}, { scope: 'render', sync: true });
+  }
+
+  async function addStoryPanel() {
+    const index = (state.story.panels?.length ?? 0) + 1;
+    // A counter, not the array length: removing a panel and then adding another
+    // used to hand out an id that was already taken, which confused the list.
+    storySeq += 1;
+    const id = `p${storySeq}`;
+    state.story.panels = [...(state.story.panels ?? []), {
+      id,
+      label: `コマ${index}`,
+      pose: JSON.parse(JSON.stringify(state.pose)),
+      face: JSON.parse(JSON.stringify(state.face)),
+      caption: JSON.parse(JSON.stringify(state.caption)),
+      caption2: JSON.parse(JSON.stringify(state.caption2)),
+      // The 擬音 stamps as well, so a panel comes back with everything it showed.
+      gion: JSON.parse(JSON.stringify(state.gion)),
+      // The camera as well as the pose: recalling a panel should put you back
+      // where you were looking when you recorded it.
+      view: JSON.parse(JSON.stringify(state.view)),
+    }];
+    applyState({}, { scope: 'caption', sync: true });
+
+    // Then photograph what was just recorded, so the list shows the shot rather
+    // than only its name.
+    try {
+      const shot = await captureImage({ scale: 0.35, transparent: false });
+      storyThumbs.set(id, {
+        version: (storyThumbs.get(id)?.version ?? 0) + 1,
+        url: thumbDataUrl(shot.canvas),
+      });
+      toast(`コマ${index}を追加しました`);
+    } catch (error) {
+      console.warn('[bluebey] story thumbnail failed', error);
+      toast(`コマ${index}を追加しました（プレビューは作れませんでした）`);
+    }
+    app.panel?.sync();
+  }
+
+  /** Shrink a capture down to a thumbnail data URL for the panel list. */
+  function thumbDataUrl(source, width = 168) {
+    const scale = width / Math.max(1, source.width);
+    const canvas = document.createElement('canvas');
+    canvas.width = Math.max(1, Math.round(source.width * scale));
+    canvas.height = Math.max(1, Math.round(source.height * scale));
+    const ctx = canvas.getContext('2d');
+    ctx.imageSmoothingQuality = 'high';
+    ctx.drawImage(source, 0, 0, canvas.width, canvas.height);
+    return canvas.toDataURL('image/png');
+  }
+
+  /**
+   * A number in the corner of a comic panel, for sheets whose reading order is
+   * not obvious. Drawn with the vendored caption font, so the sheet does not
+   * depend on whatever fonts the viewer happens to have.
+   */
+  function drawPanelNumber(ctx, number, x, y, cellW, cellH) {
+    const radius = Math.max(18, Math.min(64, Math.min(cellW, cellH) * 0.075));
+    const margin = radius * 0.7;
+    const cx = x + margin + radius;
+    const cy = y + margin + radius;
+    ctx.save();
+    ctx.beginPath();
+    ctx.arc(cx, cy, radius, 0, Math.PI * 2);
+    ctx.fillStyle = 'rgba(255, 255, 255, 0.92)';
+    ctx.fill();
+    ctx.lineWidth = Math.max(2, radius * 0.14);
+    ctx.strokeStyle = '#3f2b52';
+    ctx.stroke();
+    ctx.fillStyle = '#3f2b52';
+    ctx.textAlign = 'center';
+    ctx.textBaseline = 'middle';
+    ctx.font = `bold ${Math.round(radius * 1.25)}px ${captionFontStack()}`;
+    ctx.fillText(String(number), cx, cy + radius * 0.04);
+    ctx.restore();
+  }
+
+  /**
+   * まんが: replay the panels, capture each one, and hand back a framed sheet.
+   */
+  async function saveStory({ scale = 2 } = {}) {
+    const panels = state.story?.panels ?? [];
+    if (!panels.length) {
+      toast('コマがありません。「今の状態をコマに追加」で作ってください');
+      return;
+    }
+    const saved = JSON.parse(JSON.stringify(state));
+    const captures = [];
+    suspendHistory = true;
+    try {
+      for (const panel of panels) {
+        const pose = panel.pose ?? {};
+        state.pose = { bones: pose.bones ?? {}, root: pose.root ?? [0, 0, 0] };
+        state.face = JSON.parse(JSON.stringify(saved.face));
+        applyPatch(state.face, panel.face ?? {});
+        state.caption = { ...saved.caption, ...(panel.caption ?? {}) };
+        state.caption2 = { ...saved.caption2, ...(panel.caption2 ?? {}) };
+        // The camera the panel was framed with, so a comic keeps its angles.
+        if (panel.view) state.view = JSON.parse(JSON.stringify(panel.view));
+        refresh('all');
+        // Two frames: one to apply the pose, one to draw it.
+        await new Promise((resolve) => requestAnimationFrame(() => requestAnimationFrame(resolve)));
+        captures.push(await captureImage({ scale, transparent: false }));
+      }
+    } finally {
+      suspendHistory = false;
+      applyState(saved, { full: true, sync: true });
+    }
+
+    const columns = Math.max(1, Math.min(4, state.story?.columns ?? 2));
+    const gap = state.story?.gap ?? 12;
+    const padding = state.story?.padding ?? 20;
+    const rows = Math.ceil(captures.length / columns);
+    const cellW = Math.max(...captures.map((item) => item.width));
+    const cellH = Math.max(...captures.map((item) => item.height));
+
+    // Once the sheet is more than one column wide, the reading order stops being
+    // obvious from the layout alone, so the panels are numbered. A single column
+    // (or two panels) reads top to bottom without help.
+    const numbered = columns >= 2 && captures.length >= 3;
+    if (numbered) await ensureCaptionFont();
+
+    // A manga page reads by its panel frames. They are drawn inside each cell so
+    // the artwork keeps its full size, and are thick enough to survive the 1×
+    // export.
+    const frame = Math.max(3, Math.round(Math.min(cellW, cellH) * 0.012));
+
+    const sheet = document.createElement('canvas');
+    sheet.width = padding * 2 + cellW * columns + gap * (columns - 1);
+    sheet.height = padding * 2 + cellH * rows + gap * (rows - 1);
+    const ctx = sheet.getContext('2d');
+    ctx.fillStyle = state.story?.sheetBackground ?? '#ffffff';
+    ctx.fillRect(0, 0, sheet.width, sheet.height);
+    captures.forEach((capture, index) => {
+      const x = padding + (index % columns) * (cellW + gap);
+      const y = padding + Math.floor(index / columns) * (cellH + gap);
+      ctx.drawImage(capture.canvas, x, y, cellW, cellH);
+      ctx.lineWidth = frame;
+      ctx.strokeStyle = '#2a1e33';
+      ctx.strokeRect(x + frame / 2, y + frame / 2, cellW - frame, cellH - frame);
+      if (numbered) drawPanelNumber(ctx, index + 1, x, y, cellW, cellH);
+    });
+
+    const stamp = exporter.timestamp();
+    exporter.downloadBlob(await exporter.canvasToBlob(sheet), `bluebey-comic-${stamp}.png`);
+    toast(`まんがを書き出しました（${captures.length}コマ）`);
+  }
+
+  /** A shared link restores the whole look; otherwise this is the start point. */
+  async function loadFromHash() {
+    // 共有ボタンは `?s=`、ボタン「この見た目のリンクをコピー」は `#s=` を使う。
+    // Facebook は `#` 以降を落とすので、共有はクエリにする必要がある - どちらでも読む。
+    const query = new URLSearchParams(location.search).get('s');
+    const hash = location.hash.replace(/^#/, '');
+    const encoded = query ?? new URLSearchParams(hash).get('s');
+    if (!encoded) return false;
+    const decoded = await decodeState(encoded);
+    if (!decoded?.state) return false;
+    suspendHistory = true;
+    try {
+      applyState(decoded.state, { full: true, sync: true });
+    } finally {
+      suspendHistory = false;
+    }
+    return true;
+  }
+
+  function installTopbar() {
+    // The panel itself is visible from the start, but each of its sections starts
+    // collapsed (see `section()` in src/ui.js): the panel reads as a short list of
+    // headings, and you open the one you want instead of scrolling past twelve.
+    resizeViewport();
+
+    // AR (カメラ＝AR) get a camera-app style shutter over the viewport, so a photo
+    // is one tap away. It is a DOM layer over the canvas, so it never lands in the
+    // export; its visibility follows the background mode (see `applyBackdrop`).
+    const shutter = document.createElement('button');
+    shutter.type = 'button';
+    shutter.id = 'ar-shutter';
+    shutter.className = 'ar-shutter';
+    shutter.title = '写真を撮る';
+    shutter.setAttribute('aria-label', '写真を撮る');
+    shutter.hidden = true;
+    shutter.addEventListener('click', () => { savePNG(); });
+    stage.append(shutter);
+    app.arShutter = shutter;
+
+    document.getElementById('btn-shot')?.addEventListener('click', () => app.actions.savePNG());
+    document.getElementById('btn-copy')?.addEventListener('click', () => app.actions.copyPNG());
+    document.getElementById('btn-help')?.addEventListener('click', toggleHelp);
+    document.getElementById('btn-panel')?.addEventListener('click', togglePanel);
+    document.getElementById('help-close')?.addEventListener('click', () => { toggleHelp(false); });
+    const help = document.getElementById('help');
+    help?.addEventListener('click', (event) => { if (event.target === help) toggleHelp(false); });
+
+    // --- the mouse on the character ----------------------------------------
+    // Dragging the model slides it across the view; holding Shift and dragging
+    // turns it around. Bone posing lives in the panel, so the mouse never grabs a
+    // bone (see `Rig` in src/rig.js). On a small screen the panel is a sheet over
+    // the view, so a tap on the view first just dismisses the sheet.
+    const dragRay = new THREE.Raycaster();
+    const dragNdc = new THREE.Vector2();
+    const dragHitAt = (event) => {
+      const rect = canvas.getBoundingClientRect();
+      dragNdc.set(
+        ((event.clientX - rect.left) / Math.max(1, rect.width)) * 2 - 1,
+        -((event.clientY - rect.top) / Math.max(1, rect.height)) * 2 + 1,
+      );
+      dragRay.setFromCamera(dragNdc, view.camera);
+      const targets = [model.parts.eyeMesh, model.parts.mouthMesh, ...model.parts.body]
+        .filter((mesh) => mesh?.visible);
+      return dragRay.intersectObjects(targets, false)[0] ?? null;
+    };
+    // World units per screen pixel, so the model keeps up with the cursor.
+    const dragScale = () => {
+      const camera = view.camera;
+      const height = Math.max(1, canvas.clientHeight);
+      if (camera.isOrthographicCamera) return (camera.top - camera.bottom) / camera.zoom / height;
+      const distance = camera.position.distanceTo(view.controls.target);
+      return (2 * distance * Math.tan((camera.fov * Math.PI) / 360)) / height;
+    };
+    let modelDrag = null;
+
+    canvas.addEventListener('pointerdown', (event) => {
+      if (event.button !== 0 || event.target !== canvas) return;
+      if (narrowScreenLayout.matches && !document.body.classList.contains('panel-hidden')) {
+        togglePanel();
+        return;
+      }
+      if (!dragHitAt(event)) return;
+      const root = state.pose.root ?? [0, 0, 0];
+      const master = state.pose.bones?.master ?? [0, 0, 0];
+      modelDrag = {
+        pointerId: event.pointerId,
+        spin: event.shiftKey,
+        x: event.clientX,
+        y: event.clientY,
+        root: [...root],
+        masterY: master[1] ?? 0,
+      };
+      view.controls.enabled = false;
+      stage.style.cursor = 'grabbing';
+      try { canvas.setPointerCapture(event.pointerId); } catch { /* ignore */ }
+    });
+
+    canvas.addEventListener('pointermove', (event) => {
+      if (!modelDrag || event.pointerId !== modelDrag.pointerId) return;
+      const dx = event.clientX - modelDrag.x;
+      const dy = event.clientY - modelDrag.y;
+      if (modelDrag.spin) {
+        const bones = { ...(state.pose.bones ?? {}) };
+        const master = bones.master ?? [0, 0, 0];
+        bones.master = [master[0], round2(modelDrag.masterY + dx * 0.7), master[2]];
+        state.pose.bones = bones;
+      } else {
+        const k = dragScale();
+        state.pose.root = [
+          round2(modelDrag.root[0] + dx * k),
+          round2(modelDrag.root[1] - dy * k),
+          modelDrag.root[2],
+        ];
+      }
+      app.panel?.onRigChanged?.();
+      app.needsRender = true;
+    });
+
+    const endModelDrag = (event) => {
+      if (!modelDrag || (event && event.pointerId !== modelDrag.pointerId)) return;
+      modelDrag = null;
+      view.controls.enabled = true;
+      stage.style.cursor = '';
+      if (event) { try { canvas.releasePointerCapture(event.pointerId); } catch { /* ignore */ } }
+      capturePose();
+      app.panel?.onRigChanged?.();
+      app.needsRender = true;
+    };
+    canvas.addEventListener('pointerup', endModelDrag);
+    canvas.addEventListener('pointercancel', endModelDrag);
+  }
+
+  function togglePanel() {
+    document.body.classList.toggle('panel-hidden');
+    resizeViewport();
+  }
+
+  // The panel is a side column on a wide screen and a bottom sheet on a small
+  // one, where it starts closed so the 3D view fills the screen (see the
+  // small-screen rules in src/style.css). When the window is resized across that
+  // breakpoint, match the layout the other side expects instead of leaving the
+  // panel stuck in the previous state.
+  const narrowScreenLayout = window.matchMedia('(max-width: 768px), (max-height: 500px) and (pointer: coarse)');
+  narrowScreenLayout.addEventListener('change', (event) => {
+    document.body.classList.toggle('panel-hidden', event.matches);
+    resizeViewport();
+  });
+
+  function installKeys() {
+    window.addEventListener('keydown', (event) => {
+      // Undo/redo are the only shortcuts that need a modifier, so they are dealt
+      // with before the "no modifiers" rule below.
+      if ((event.ctrlKey || event.metaKey) && !event.altKey) {
+        const combo = event.key.toLowerCase();
+        if (combo === 'z' || combo === 'y') {
+          event.preventDefault();
+          restoreFromHistory(combo === 'y' || event.shiftKey ? 'redo' : 'undo');
+          return;
+        }
+      }
+      if (event.metaKey || event.ctrlKey || event.altKey) return;
+      const target = event.target;
+      if (target instanceof HTMLInputElement || target instanceof HTMLSelectElement) return;
+      switch (event.key) {
+        case 's': case 'S': savePNG(); break;
+        case 'c': case 'C': copyPNG(); break;
+        case 'r': case 'R': app.actions.resetPose(); break;
+        case 'g': case 'G': {
+          const visible = !rig.helper.visible;
+          rig.setGizmoVisible(visible);
+          toast(visible ? 'ギズモを表示' : 'ギズモを隠しました');
+          app.needsRender = true;
+          break;
+        }
+        case ' ': event.preventDefault(); app.actions.toggleAnimation(); break;
+        case '?': toggleHelp(); break;
+        case 'Tab': event.preventDefault(); togglePanel(); break;
+        case '1': setCameraPreset('front'); break;
+        case '2': setCameraPreset('threeQuarter'); break;
+        case '3': setCameraPreset('side'); break;
+        case '4': setCameraPreset('back'); break;
+        case '5': setCameraPreset('top'); break;
+        default: break;
+      }
+    });
+  }
+
+}
+
+function toggleHelp(force) {
+  const help = document.getElementById('help');
+  if (help) help.hidden = force === undefined ? !help.hidden : !force;
+}
+
+/** A persistent message for problems that need the user to do something. */
+function showNotice(text) {
+  let node = document.getElementById('notice');
+  if (!node) {
+    node = document.createElement('div');
+    node.id = 'notice';
+    document.body.append(node);
+  }
+  node.textContent = text;
+  node.hidden = false;
+}
+
+/* ------------------------------------------------------------------- camera */
+
+/**
+ * Orbiting camera with both a perspective and an orthographic projection.
+ * Only one of them is enabled at a time; switching copies the framing across so
+ * the model does not jump.
+ */
+class ViewRig {
+  constructor(domElement, onChange) {
+    this.domElement = domElement;
+    this.onChange = onChange;
+    this.projection = 'persp';
+    this.controlsChanged = false;
+
+    this.cameras = {
+      // A tight near/far keeps depth precision high, which matters because the
+      // outline hull is only a few hundredths of a unit away from the surface.
+      // The far plane has to sit well past the furthest the camera may go, or
+      // zooming all the way out slices the character on it - which is exactly
+      // what the old `far = 120` did, because the zoom limit was also 120.
+      persp: new THREE.PerspectiveCamera(30, 1, 1, 2000),
+      ortho: new THREE.OrthographicCamera(-1, 1, 1, -1, -2000, 2000),
+    };
+    this.controlsMap = {};
+    for (const [key, camera] of Object.entries(this.cameras)) {
+      const controls = new OrbitControls(camera, domElement);
+      controls.enableDamping = false;
+      controls.enablePan = true;
+      controls.minDistance = 1;
+      controls.maxDistance = 400;
+      controls.addEventListener('change', () => {
+        this.controlsChanged = true;
+        this.onChange?.();
+      });
+      this.controlsMap[key] = controls;
+    }
+    this.controlsMap.ortho.enabled = false;
+    this.camera = this.cameras.persp;
+    this.controls = this.controlsMap.persp;
+    this.target = new THREE.Vector3(0, 1, 0);
+    this.autoDistance = 12;
+    // `orthoHeight` is where the orthographic camera is *now* - `apply()` writes
+    // the camera's actual height back into it - so the framing `frame()` chose
+    // needs its own field to measure the zoom against (see `zoomFactor`).
+    this.orthoHeight = 6;
+    this.autoOrthoHeight = 6;
+    this.resize();
+  }
+
+  /** Fit the model into view. */
+  frame(size, targetY) {
+    this.target.set(0, targetY, 0);
+    for (const controls of Object.values(this.controlsMap)) controls.target.copy(this.target);
+    const fov = this.cameras.persp.fov * Math.PI / 180;
+    const heightDistance = size.y / (2 * Math.tan(fov / 2));
+    const widthDistance = size.x / (2 * Math.tan(fov / 2) * Math.max(0.4, this.aspect));
+    this.autoDistance = Math.max(heightDistance, widthDistance) * 1.45;
+    this.orthoHeight = size.y * 1.5;
+    this.autoOrthoHeight = this.orthoHeight;
+    for (const camera of Object.values(this.cameras)) {
+      camera.position.set(0, targetY + this.autoDistance * 0.12, this.autoDistance);
+    }
+    this.resize();
+  }
+
+  get aspect() {
+    const width = this.domElement.clientWidth || 1;
+    const height = this.domElement.clientHeight || 1;
+    return width / height;
+  }
+
+  resize() {
+    const width = this.domElement.clientWidth || window.innerWidth;
+    const height = this.domElement.clientHeight || window.innerHeight;
+    const aspect = width / height;
+    this.cameras.persp.aspect = aspect;
+    this.cameras.persp.updateProjectionMatrix();
+    this.setOrthoHeight(this.orthoHeight);
+  }
+
+  setOrthoHeight(height) {
+    const ortho = this.cameras.ortho;
+    const halfHeight = height / 2;
+    const halfWidth = halfHeight * this.aspect;
+    ortho.left = -halfWidth;
+    ortho.right = halfWidth;
+    ortho.top = halfHeight;
+    ortho.bottom = -halfHeight;
+    ortho.zoom = 1;
+    ortho.updateProjectionMatrix();
+  }
+
+  /**
+   * How large the character looks now, relative to the auto-fit framing.
+   *
+   * 1 means "the default framing", which is where the panel's 大きさ sliders sit
+   * and so where the screen-space outline's pixel width is written against. The
+   * two projections scale differently - a perspective camera shows 1/distance as
+   * much per world unit, an orthographic one 1/height, and OrbitControls zooms
+   * the latter with `camera.zoom` rather than by moving it - so each is
+   * normalised by its own auto-fit value.
+   */
+  zoomFactor() {
+    if (this.projection === 'ortho') {
+      const ortho = this.cameras.ortho;
+      const height = (ortho.top - ortho.bottom) / Math.max(ortho.zoom, 0.0001);
+      return this.autoOrthoHeight > 0 && height > 0 ? this.autoOrthoHeight / height : 1;
+    }
+    const distance = this.camera.position.distanceTo(this.controls?.target ?? this.target);
+    return this.autoDistance > 0 && distance > 0 ? this.autoDistance / distance : 1;
+  }
+
+  setProjection(kind) {
+    if (kind === this.projection) return;
+    const previous = this.camera;
+    const next = kind === 'ortho' ? this.cameras.ortho : this.cameras.persp;
+    next.position.copy(previous.position);
+    next.quaternion.copy(previous.quaternion);
+    this.controlsMap[this.projection].enabled = false;
+    this.projection = kind;
+    this.camera = next;
+    this.controls = this.controlsMap[kind];
+    this.controls.enabled = true;
+    this.controls.target.copy(this.target);
+    this.resize();
+    this.onCameraChange?.(next, this.controls);
+    this.controlsChanged = true;
+    this.onChange?.();
+  }
+
+  /** Apply the serialisable view state (azimuth/polar/zoom). */
+  apply(viewState, size) {
+    const distance = viewState.distance > 0 ? viewState.distance : this.autoDistance;
+    const targetY = viewState.targetY || this.target.y;
+    this.target.set(viewState.targetX || 0, targetY, viewState.targetZ || 0);
+    const spherical = new THREE.Spherical(
+      distance,
+      clamp(viewState.polar, 1, 179) * Math.PI / 180,
+      viewState.azimuth * Math.PI / 180,
+    );
+    const offset = new THREE.Vector3().setFromSpherical(spherical);
+    for (const camera of Object.values(this.cameras)) {
+      camera.position.copy(this.target).add(offset);
+      camera.lookAt(this.target);
+    }
+    if (viewState.orthoHeight > 0) this.setOrthoHeight(viewState.orthoHeight);
+    else if (size) this.setOrthoHeight(this.orthoHeight);
+    for (const controls of Object.values(this.controlsMap)) {
+      controls.target.copy(this.target);
+      controls.update();
+    }
+    this.orthoHeight = this.cameras.ortho.top - this.cameras.ortho.bottom;
+    this.cameras.persp.updateProjectionMatrix();
+    this.cameras.ortho.updateProjectionMatrix();
+  }
+
+  /**
+   * Write where the camera actually is back into the serialisable view state.
+   *
+   * WHY: the mouse orbit and pan move the `controls`, not the state. Without this
+   * the state kept the last camera the *panel* set, so any refresh of the view
+   * snapped the camera back to it - which is what made picking a background jump
+   * the camera. Keeping the two in step also means a shared link carries the
+   * camera you framed.
+   */
+  captureInto(viewState) {
+    if (!viewState || typeof viewState !== 'object') return;
+    const target = this.controls?.target ?? this.target;
+    const spherical = new THREE.Spherical().setFromVector3(
+      new THREE.Vector3().subVectors(this.camera.position, target),
+    );
+    viewState.azimuth = Math.round(spherical.theta * RAD_TO_DEG * 100) / 100;
+    viewState.polar = Math.round(spherical.phi * RAD_TO_DEG * 100) / 100;
+    viewState.distance = Math.round(spherical.radius * 1000) / 1000;
+    viewState.targetX = Math.round(target.x * 1000) / 1000;
+    viewState.targetY = Math.round(target.y * 1000) / 1000;
+    viewState.targetZ = Math.round(target.z * 1000) / 1000;
+    if (this.projection === 'ortho') {
+      const ortho = this.cameras.ortho;
+      viewState.orthoHeight = Math.round((ortho.top - ortho.bottom) * 1000) / 1000;
+    }
+  }
+
+  setAutoRotate(enabled, speed) {
+    for (const controls of Object.values(this.controlsMap)) {
+      controls.autoRotate = enabled;
+      controls.autoRotateSpeed = speed;
+    }
+  }
+}
+
+/**
+ * How far the screen-space outline may follow the zoom (see `outlineRadiusFor`).
+ *
+ * A clamp keeps a very deep zoom from turning the line into either a smear or
+ * nothing at all, which is what the linear factor would do at the extremes of
+ * the 大きさ slider.
+ */
+const OUTLINE_ZOOM_MIN = 0.25;
+const OUTLINE_ZOOM_MAX = 3;
+
+/**
+ * The screen-space outline's radius, in pixels of the label buffer.
+ *
+ * The body's lines are inverted hulls, i.e. an offset in *world* units, so they
+ * thicken as the character grows on screen and thin as it shrinks. This pass
+ * works in pixels instead, so left alone its lines - the waist leaves and the
+ * nose - keep one width at every zoom, and the leaf skirt reads as far too heavy
+ * beside the body the moment you pull back. Scaling the radius by the rig's
+ * `zoomFactor` makes the two behave the same way, with `outlinePixels` staying
+ * the width at the auto-fit framing (so the slider keeps its meaning).
+ *
+ * `extra` carries whatever else changes the buffer's pixels-per-CSS-pixel: the
+ * preview's buffer is `quality`x the CSS size (and `quality` itself follows the
+ * display density and the automatic reductions), and an export renders the label
+ * buffer at `scale`x. Passing that factor keeps the line a constant *CSS* width,
+ * so `outlinePixels` reads as CSS pixels whichever buffer it lands in.
+ */
+function outlineRadiusFor(view, state, extra = 1) {
+  const zoom = clamp(view.zoomFactor(), OUTLINE_ZOOM_MIN, OUTLINE_ZOOM_MAX);
+  return (state.render.outlinePixels ?? 2) * zoom * extra;
+}
+
+/* ----------------------------------------------------------------- scenery */
+
+function makeEnvironment(renderer) {
+  const canvas = document.createElement('canvas');
+  canvas.width = 64;
+  canvas.height = 32;
+  const ctx = canvas.getContext('2d');
+  const gradient = ctx.createLinearGradient(0, 0, 0, 32);
+  gradient.addColorStop(0, '#ffffff');
+  gradient.addColorStop(0.5, '#ece7fa');
+  gradient.addColorStop(1, '#b9aade');
+  ctx.fillStyle = gradient;
+  ctx.fillRect(0, 0, 64, 32);
+  const texture = new THREE.CanvasTexture(canvas);
+  texture.mapping = THREE.EquirectangularReflectionMapping;
+  texture.colorSpace = THREE.SRGBColorSpace;
+  const generator = new THREE.PMREMGenerator(renderer);
+  const environment = generator.fromEquirectangular(texture).texture;
+  texture.dispose();
+  generator.dispose();
+  return environment;
+}
+
+/** Round to 2 decimals: the precision the panel's prop sliders show. */
+const round2 = (value) => Math.round(value * 100) / 100;
+
+const clampNumber = (value, lo, hi) => Math.min(hi, Math.max(lo, value));
+
+export function backgroundOf(viewState) {
+  if (viewState.background === 'transparent') return { mode: 'transparent' };
+  // A preset, a loaded photo, a drawn effect line or the phone's camera all
+  // live in a DOM layer behind the WebGL canvas, so the renderer is left clear
+  // and the exports paint the bitmap into the picture themselves (see
+  // drawBackdropInto in main.js).
+  if (viewState.background === 'preset' || viewState.background === 'image'
+    || viewState.background === 'effect' || viewState.background === 'camera') {
+    return { mode: 'transparent' };
+  }
+  return { mode: 'solid', color: viewState.backgroundColor };
+}
+
+/* ------------------------------------------------------------------- errors */
+
+function setLoadingProgress(ratio) {
+  const node = document.getElementById('loading-text');
+  if (node) node.textContent = `ぶるべーを読み込んでいます… ${Math.round(ratio * 100)}%`;
+}
+
+function hideLoading() {
+  const node = document.getElementById('loading');
+  if (node) node.classList.add('done');
+}
+
+function showLoadError(error) {
+  const text = document.getElementById('loading-text');
+  const detail = document.getElementById('loading-error');
+  if (text) text.textContent = '読み込みに失敗しました。';
+  if (detail) {
+    detail.textContent = `${error?.message ?? error}\n\n`
+      + 'このページはローカルサーバー経由で開く必要があります。\n'
+      + 'bluebey-studio フォルダで  python serve.py  を実行し、\n'
+      + '表示された http://127.0.0.1:8000/ をブラウザで開いてください。';
+  }
+}
+
+const clamp = (value, lo, hi) => Math.min(hi, Math.max(lo, value));
+const clamp01 = (value) => clamp(value, 0, 1);
-- 
cgit v1.3.1