@solidrt/3d 0.0.56 → 0.0.58

package/src/orbit.ts

@@ -1,98 +1,90 @@
-// An orbit camera for a scene: azimuth/elevation/distance around a target,
-// with drag-to-rotate, pinch- and wheel-to-zoom, optional auto-orbit, and
-// clamps - the standard interactive-viewer camera, extracted so apps stop
-// rebuilding it.
+// An orbit camera for a scene: azimuth/elevation/distance around a target
+// - the standard interactive-viewer camera, extracted so apps stop
+// rebuilding it. The control is pure (ARCHITECTURE.md: controls through
+// an abstraction, never direct event handling): it consumes three axes,
+// `rotate` (vec2), `zoom` (axis) and `pan` (vec2), in device-free units,
+// and exposes pose verbs for scripted moves. What drives the axes is the
+// app's business - an input map with a pointer feed, a pad, a debug
+// command - and the `orbitBindings` preset in input.ts is the standard
+// wiring (drag rotates, pinch and wheel zoom, two fingers pan).
 //
-// Rotation is viewport-relative when the control knows its viewport
-// (OrbitCameraOptions.viewport): dragging one viewport height sweeps one
-// full turn, Three's OrbitControls convention, so the same drag feels the
-// same on a phone and a 4k window. Without a viewport it falls back to a
-// fixed angle per pixel.
-//
-// Input goes through core's merged transform recognizer (createTransform), so
-// the drag participates in the gesture arena: on a viewport embedded in a
-// scrollable layout, an orbit drag and the ancestor scroller's pan arbitrate
-// instead of both acting, and rotation only starts once the drag crosses the
-// recognizer's slop. One finger (or a mouse drag) rotates; two-finger pinch
-// zooms (pair rotation is ignored); wheel zooms directly - no recognizer
-// needed for a discrete wheel step. What two-finger translation means
-// depends on `viewport`: with it, two fingers pan - the scene slides under
-// the fingers 1:1 at the target's depth, the touch convention everywhere
-// (three.js DOLLY_PAN, Sketchfab, touch CAD) - which also keeps an
-// imperfect pinch from smearing rotation into the zoom. Without it there
-// is no pixel-to-world mapping, so focal translation falls back to
-// rotating regardless of finger count.
+// Rotation is element-relative: one element height of drag travel (the
+// feed's unit) sweeps DRAG_TURNS full turns on either axis, Three's
+// OrbitControls convention, so the same drag feels the same on a phone
+// and a 4k window. A rate on `rotate` (a stick) turns at ROTATE_RATE
+// turns per second at full deflection. `zoom` is in octaves: a delta of
+// 1 halves the distance, a rate of 1 halves it per second. `pan` slides
+// the target so the scene tracks the fingers 1:1 at the target's depth,
+// the touch convention everywhere (three.js DOLLY_PAN, Sketchfab, touch
+// CAD), mapped through the target camera's fov - which is why the driven
+// target must report its camera (a Scene or a View does).
 //
 // Zoom aims at the target by default. With `zoomAnchor` the app maps the
-// pinch focal / wheel cursor to a world point, and the zoom scales the pose
-// about that point instead - the spot under the fingers stays under the
-// fingers, with the target sliding toward it. Only the app can own that
-// mapping: screen-to-ray needs the projection (fov, aspect), which lives
-// in the app's camera, not here.
-//
-// Anchored zoom leaves the target wherever the zoom carried it - possibly a
-// point in empty air near the eye, and a drag orbiting THAT swings the scene
-// wildly around nothing. `rotateAnchor` is the countermeasure: at gesture
-// start the app names the world point rotation should pivot about (what the
-// camera is actually looking at), and the pivot is re-seated there. The
-// point is projected onto the view axis first, so re-seating moves only the
-// target's depth - the eye and the picture do not change, only what a drag
-// swings around.
+// gesture's focal point (a fraction of the element) to a world point,
+// and the zoom scales the pose about that point instead - the spot under
+// the fingers stays under the fingers, with the target sliding toward it.
+// Only the app can own that mapping: focal-to-ray needs the projection,
+// which lives in the app's scene (scene.unproject over the scene's size).
+// A pinch keeps ONE anchor for its whole gesture, taken at its first
+// delta: the two fingers' events interleave, so the measured span
+// oscillates around its true value even while the fingers rest, and
+// per-delta re-anchoring turns that noise into a visible crawl; about a
+// fixed point the pairs cancel exactly. The wheel, unbracketed, anchors
+// per notch: its notches are discrete, there is no noise to cancel, and
+// each notch re-aiming at the cursor is the point.
 //
-// Every pointer coordinate here is in the input element's own frame: the
-// recognizer measures drag deltas in the element's parent frame and the
-// pinch focal in its local frame, and the wheel uses localX/localY, so a
-// viewport under a designSize fit or any scaled ancestor behaves the same
-// as one filling the window. `viewport().height` is therefore the
-// element's laid-out height, which the OrbitCamera component passes.
+// Anchored zoom leaves the target wherever the zoom carried it - possibly
+// a point in empty air near the eye, and a drag orbiting THAT swings the
+// scene wildly around nothing. `rotateAnchor` is the countermeasure: when
+// a rotate gesture begins, the app names the world point rotation should
+// pivot about (what the camera is actually looking at), and setPivot
+// re-seats the target there, projected onto the view axis first, so only
+// the target's depth moves - the eye and the picture do not change.
 //
-// Pose is plain mutable state advanced by update(dt) from the app's own
-// onFrame; the control registers no frame loop of its own, so whether
-// anything animates stays the app's decision (a paused orbit with no other
-// animation costs nothing). Only `orbiting` is a signal - slow UI state
-// that HUDs read - while the pose moves at frame rate and bypasses
-// reactivity: the package's structure-vs-motion split. update() pushes the
-// pose to the scene camera only when it actually changed (one setCamera:
-// the scene's own shared write carries uViewProj and uCamPos), and reports
-// that, so per-frame dependents (reprojecting HUD overlays via
-// scene.project) can follow the camera without recomputing every frame.
+// Pose is plain mutable state; a nudge or a verb pushes it to the target
+// at once (the next paint carries the new camera, no frame loop needed
+// for a drag), and update(dt) integrates the rates and the auto-orbit
+// from the app's own onFrame - the control registers no frame loop of its
+// own. Only `orbiting` and `active` are reactive: slow UI state a HUD
+// reads, and the frame-loop gate. update() returns whether the pose
+// changed since the previous update (pushes included), so per-frame
+// dependents (reprojecting HUD overlays via scene.project) can follow the
+// camera without recomputing every frame.
 //
-// Options are read where they apply, not copied out: the clamps, the rates
-// (auto-orbit, rotate, zoom, pan), `viewport` and the anchor callbacks are
-// re-read from the options object on every input or update, so a caller
-// may change a field (or hand in an object of getters, which is what
-// `<OrbitCamera>` does with its props) and the next gesture or frame sees
-// it. Only the initial pose is copied at creation; later pose changes go
-// through set().
+// Options are read where they apply, not copied out: the clamps, the
+// rates and the anchor callbacks are re-read from the options object on
+// every input or update, so a caller may change a field (or hand in an
+// object of getters, which is what `<OrbitCamera>` does with its props)
+// and the next gesture or frame sees it. Only the initial pose is copied
+// at creation; later pose changes go through set() and the verbs.
 
-import { createSignal } from "@solidjs/signals"
-import { createTransform } from "@solidrt/core"
-import type { PointerEvent } from "@solidrt/core"
-import type { CameraUpdate } from "./camera.ts"
+import { createMemo, createSignal, untrack } from "@solidjs/signals"
+import { createAxes } from "@solidrt/core/input"
+import type { Axes, Vec2 } from "@solidrt/core/input"
+import type { CameraState, CameraUpdate } from "./camera.ts"
 import type { Vec3 } from "./math.ts"
 
-// Baseline sensitivities at rotateSpeed/zoomSpeed 1. With a `viewport` the
-// rotation is viewport-relative: dragging one viewport height sweeps
-// DRAG_TURNS full turns on either axis (Three's OrbitControls convention),
-// so the feel survives any window size or form factor. Without one it
-// falls back to fixed radians per dragged pixel - the relative rule frozen
-// at a ~785 px viewport. Wheel zoom is an exponent per wheel-delta unit.
+// One element height of drag sweeps this many full turns on either axis
+// (Three's OrbitControls convention).
 const DRAG_TURNS = 1
-const DRAG_AZIMUTH = 0.008
-const DRAG_ELEVATION = 0.006
-const WHEEL_ZOOM = 0.0015
+// Rate axes at full deflection, per second: turns of rotation, octaves of
+// zoom (1 = the distance halves each second), element heights of pan.
+const ROTATE_RATE = 0.5
+const ZOOM_RATE = 1
+const PAN_RATE = 1
 // Default clamps: the distance floor keeps the eye off the target; the
 // elevation limits (radians) stop just short of the poles.
 const MIN_DISTANCE = 0.01
 const ELEVATION_LIMIT = 1.55
 
-/** What an orbit camera drives: a Scene, or one of its Views. */
-export type OrbitTarget = { setCamera(update: CameraUpdate): void }
+/** What an orbit camera drives: a Scene, or one of its Views - anything
+ * with their setCamera and camera() (the fov maps pan travel to world). */
+export type OrbitTarget = { setCamera(update: CameraUpdate): void; camera(): Pick<CameraState, "fov"> }
 
 /** The pose fields (target, azimuth, elevation, distance) are initial
  * values, copied at creation and changed through set() afterwards. Every
  * other field is live: read from this object where it applies, so a
- * change takes effect on the next gesture or update. */
+ * change takes effect on the next input or update. */
 export type OrbitCameraOptions = {
   /** Initial point the camera orbits and looks at (default origin). */
   target?: Vec3
@@ -107,38 +99,27 @@ export type OrbitCameraOptions = {
   minElevation?: number
   maxElevation?: number
   /** Auto-orbit rate in radians/second (default 0: none). Runs while
-   * `orbiting()` and no drag/pinch is in progress; `orbiting` starts on
-   * when the initial rate is positive and toggles with set({ orbiting }).
-   * A new rate applies from the next update. */
+   * `orbiting()` and no gesture is in progress; `orbiting` starts on when
+   * the initial rate is positive and toggles with set({ orbiting }). A new
+   * rate applies from the next update. */
   orbitSpeed?: number
-  /** Multipliers over the built-in drag and zoom (wheel + pinch) sensitivities. */
+  /** Multipliers over the built-in rotate, zoom and pan sensitivities
+   * (drags and rates alike). */
   rotateSpeed?: number
   zoomSpeed?: number
-  /** Multiplier over the two-finger pan (1 = the scene tracks the fingers
-   * exactly at the target's depth). */
   panSpeed?: number
-  /** The viewport the drag lives in: the input element's own height (its
-   * laid-out size, the frame the recognizer's deltas arrive in), and the
-   * camera's vertical fov in degrees; return null while neither is known
-   * yet. Providing it makes rotation viewport-relative (a drag across the
-   * viewport height is one full turn, whatever the window size) and
-   * enables two-finger pan; without it rotation is a fixed angle per pixel
-   * and two-finger translation rotates. */
-  viewport?: () => { height: number; fov: number } | null
   /** Constrain where a pan may put the target - return the target to use.
    * The typical use: keep the pivot within a few radii of the subject so
    * panning cannot strand the camera. Zoom and rotation do not consult it. */
   clampTarget?: (target: Vec3) => Vec3
-  /** Map a point in the input element's own pixels (the frame
-   * localX/localY report in) to the world point a zoom there should keep
-   * pinned. Called once per pinch gesture (at its first span change - the
-   * anchor then holds for the whole gesture, see the jitter note at
-   * `pinchAnchor`) and once per wheel event, with the pose the zoom is
-   * about to apply to. Return null to zoom toward the target as usual
-   * (also the default). */
-  zoomAnchor?: (x: number, y: number, view: { eye: Vec3; target: Vec3 }) => Vec3 | null
-  /** The world point rotation should pivot about - called when a drag or
-   * pinch starts. It is projected onto the view axis and becomes the new
+  /** Map a zoom gesture's focal point - a fraction of the driven element,
+   * [0..1, 0..1] - to the world point the zoom should keep pinned, with
+   * the pose the zoom is about to apply to. Called once per pinch gesture
+   * (the anchor holds for the whole gesture) and once per wheel notch.
+   * Return null to zoom toward the target as usual (also the default). */
+  zoomAnchor?: (focal: Vec2, view: { eye: Vec3; target: Vec3 }) => Vec3 | null
+  /** The world point rotation should pivot about - called when a rotate
+   * gesture begins. It is projected onto the view axis and becomes the new
    * target, preserving the picture exactly (only the pivot's depth moves),
    * so a drag after an anchored zoom orbits the scene under the camera
    * instead of wherever the zoom left the target. Points at or behind the
@@ -154,48 +135,65 @@ export type OrbitPose = {
   orbiting?: boolean
 }
 
+export type OrbitAxes = { rotate: "vec2"; zoom: "axis"; pan: "vec2" }
+
 export type OrbitCamera = {
   /** Camera position for the current pose (a fresh array per call). */
   eye(): Vec3
   /** Pose snapshot - the shape debug commands return and set() takes. */
-  pose(): { azimuth: number; elevation: number; distance: number }
-  /** Merge a pose in (clamps apply); reaches the scene at the next
-   * update(). Also the auto-orbit switch: set({ orbiting: false }). */
+  pose(): { azimuth: number; elevation: number; distance: number; target: Vec3 }
+  /** Merge a pose in (clamps apply) and push it. Also the auto-orbit
+   * switch: set({ orbiting: false }). */
   set(pose: OrbitPose): void
   /** Whether the auto-orbit is running. Reactive (signal-backed), so HUD
    * text can read it. */
   orbiting(): boolean
-  /** Whether update(dt) still has work: the auto-orbit on with a non-zero
-   * `orbitSpeed`. The frame-loop gate, reactive like orbiting() - the
-   * same predicate the 2d camera and the first-person camera expose. */
+  /** Whether update(dt) has work: the auto-orbit on with a non-zero
+   * `orbitSpeed`, or any axis rate non-zero. The frame-loop gate,
+   * reactive - the same predicate every camera control exposes. */
   active(): boolean
-  /** Advance the auto-orbit and push any pose change to the driven camera.
-   * Call from onFrame with the frame's dt in seconds; returns whether the
-   * pose changed. */
+  /** Integrate the axis rates and the auto-orbit over dt seconds and push
+   * any pose change. Call from onFrame; returns whether the pose changed
+   * since the previous update (nudges and verbs included). */
   update(dt: number): boolean
-  /** Spread onto the element that receives input:
-   * `<window {...orbit.handlers} />`. */
-  handlers: {
-    onPointerDown(e: PointerEvent): void
-    onPointerMove(e: PointerEvent): void
-    onPointerUp(e: PointerEvent): void
-    /** Position (the element's own pixels) is optional so a bare
-     * `{ deltaY }` still zooms; without it a zoomAnchor cannot apply and
-     * the zoom falls back to the target. */
-    onWheel(e: { deltaY: number; localX?: number; localY?: number }): void
-  }
+  /** The input abstraction: `rotate` (vec2, element heights of drag /
+   * turns per second), `zoom` (axis, octaves, positive in) and `pan`
+   * (vec2, element heights). An input map drives it by name
+   * (InputMap.drive); an app may add sources and nudge directly. */
+  axes: Axes<OrbitAxes>
+  /** Turn by radians (azimuth positive counterclockwise seen from above,
+   * elevation positive upward; clamps apply) and push. */
+  rotateBy(azimuth: number, elevation: number): void
+  /** Scale the distance by 1/factor (factor > 1 zooms in, as the 2d
+   * camera's zoomAt) about `anchor`, a world point that stays where it
+   * projects (default: toward the target), and push. */
+  zoomBy(factor: number, anchor?: Vec3 | null): void
+  /** Slide the target (and the eye with it) by world units along the
+   * camera's right and up, through `clampTarget`, and push. */
+  panBy(right: number, up: number): void
+  /** Re-seat the pivot at a world point without moving the picture: the
+   * point is projected onto the view axis and becomes the target at that
+   * depth. Points at or behind the eye are ignored. */
+  setPivot(point: Vec3): void
 }
 
 let clampNum = (v: number, lo: number, hi: number) => Math.min(hi, Math.max(lo, v))
 
+function finite(what: string, v: number): void {
+  if (!Number.isFinite(v)) throw new Error(`createOrbitCamera: ${what} must be a finite number, got ${v}`)
+}
+
 /**
  * Create an orbit camera driving `camera`'s position and target, where
- * `camera` is a Scene or one of its Views (fov, near, and far stay yours via
- * its setCamera). The initial pose applies immediately. In a component
- * tree, prefer the `<OrbitCamera>` component: it wires all of this -
- * scene, input, viewport, frame loop - through the Scene context.
+ * `camera` is a Scene or one of its Views (fov, near, and far stay yours
+ * via its setCamera). The initial pose applies immediately. In a component
+ * tree, prefer the `<OrbitCamera>` component: it drives the enclosing
+ * scene or view through context and takes an input map as a prop.
  */
 export function createOrbitCamera(camera: OrbitTarget, options: OrbitCameraOptions = {}): OrbitCamera {
+  if (!camera || typeof camera.setCamera !== "function" || typeof camera.camera !== "function") {
+    throw new Error("createOrbitCamera: the target needs setCamera() and camera() (a Scene or a View)")
+  }
   let target: Vec3 = options.target ? [options.target[0], options.target[1], options.target[2]] : [0, 0, 0]
   let azimuth = options.azimuth ?? 0
   let elevation = options.elevation ?? 0
@@ -208,35 +206,20 @@ export function createOrbitCamera(camera: OrbitTarget, options: OrbitCameraOptio
   let orbitSpeed = () => options.orbitSpeed ?? 0
   let rotateSpeed = () => options.rotateSpeed ?? 1
   let zoomSpeed = () => options.zoomSpeed ?? 1
+  let panSpeed = () => options.panSpeed ?? 1
 
   // The auto-orbit switch: starts on when the initial rate is positive.
-  let [orbiting, setOrbiting] = createSignal(orbitSpeed() > 0)
-  let interacting = false
-  let dirty = false
+  // Plain state is the truth (update() reads it in the same block set()
+  // wrote it); the signal notifies HUDs and the active() gate.
+  let running = orbitSpeed() > 0
+  let [orbiting, setOrbiting] = createSignal(running, { ownedWrite: true })
+  // Whether the pose changed since the last update() (see the header).
+  let changed = false
 
   let clampPose = () => {
     elevation = clampNum(elevation, minElevation(), maxElevation())
     distance = clampNum(distance, minDistance(), maxDistance())
   }
-  // Slide eye and target together along the camera's right/up so the scene
-  // tracks the fingers: dragged pixels map to world units through the
-  // frustum height at the target's depth. Screen +y is down, so the up-axis
-  // term is added (fingers down -> camera up -> scene follows down).
-  let pan = (dx: number, dy: number, vp: { height: number; fov: number }) => {
-    let wpp = ((2 * Math.tan((vp.fov * Math.PI) / 360) * distance) / vp.height) * (options.panSpeed ?? 1)
-    let sa = Math.sin(azimuth)
-    let ca = Math.cos(azimuth)
-    let se = Math.sin(elevation)
-    let ce = Math.cos(elevation)
-    // right = (ca, 0, -sa), up = (-sa*se, ce, -ca*se) for this pose - the
-    // camera basis written out for eye() = target + distance * (ce*sa, se, ce*ca).
-    let next: Vec3 = [
-      target[0] - dx * wpp * ca - dy * wpp * sa * se,
-      target[1] + dy * wpp * ce,
-      target[2] + dx * wpp * sa - dy * wpp * ca * se,
-    ]
-    target = options.clampTarget ? options.clampTarget(next) : next
-  }
   let eye = (): Vec3 => {
     let ce = Math.cos(elevation)
     return [
@@ -245,8 +228,26 @@ export function createOrbitCamera(camera: OrbitTarget, options: OrbitCameraOptio
       target[2] + distance * ce * Math.cos(azimuth),
     ]
   }
-  let apply = () => camera.setCamera({ position: eye(), target })
+  let view = () => ({ eye: eye(), target: [target[0], target[1], target[2]] as Vec3 })
+  let push = () => {
+    changed = true
+    camera.setCamera({ position: eye(), target })
+  }
+  // World units per element height at the target's depth: the frustum's
+  // height there, from the target camera's vertical fov.
+  let worldPerHeight = () => 2 * Math.tan((camera.camera().fov * Math.PI) / 360) * distance
 
+  // Slide eye and target together along the camera's right/up. The
+  // basis for this pose: right = (ca, 0, -sa), up = (-sa*se, ce, -ca*se),
+  // written out from eye() = target + distance * (ce*sa, se, ce*ca).
+  let slide = (right: number, up: number) => {
+    let sa = Math.sin(azimuth)
+    let ca = Math.cos(azimuth)
+    let se = Math.sin(elevation)
+    let ce = Math.cos(elevation)
+    let next: Vec3 = [target[0] + right * ca - up * sa * se, target[1] + up * ce, target[2] - right * sa - up * ca * se]
+    target = options.clampTarget ? options.clampTarget(next) : next
+  }
   // Zoom by `ratio` (new distance over old, before clamping) about a world
   // anchor (null zooms toward the target). Scaling eye and target about the
   // anchor by the distance ratio keeps it projecting to the same pixel, so
@@ -256,122 +257,152 @@ export function createOrbitCamera(camera: OrbitTarget, options: OrbitCameraOptio
   let zoomAbout = (ratio: number, anchor: Vec3 | null | undefined) => {
     let prev = distance
     // Bounds widen to the current distance so a pose already outside them
-    // (a rotateAnchor re-seat may land anywhere) zooms back toward range
+    // (a setPivot re-seat may land anywhere) zooms back toward range
     // instead of snap-jumping into it.
     distance = clampNum(distance * ratio, Math.min(minDistance(), prev), Math.max(maxDistance(), prev))
     if (!anchor) return
     let s = distance / prev
-    target = [
-      anchor[0] + (target[0] - anchor[0]) * s,
-      anchor[1] + (target[1] - anchor[1]) * s,
-      anchor[2] + (target[2] - anchor[2]) * s,
-    ]
+    target = [anchor[0] + (target[0] - anchor[0]) * s, anchor[1] + (target[1] - anchor[1]) * s, anchor[2] + (target[2] - anchor[2]) * s]
+  }
+  let setPivot = (point: Vec3) => {
+    let e = eye()
+    let fx = (target[0] - e[0]) / distance
+    let fy = (target[1] - e[1]) / distance
+    let fz = (target[2] - e[2]) / distance
+    let depth = (point[0] - e[0]) * fx + (point[1] - e[1]) * fy + (point[2] - e[2]) * fz
+    if (!(depth > 0)) return
+    // Deliberately unclamped: the picture is unchanged, this only decides
+    // what the next gesture pivots about. The zoom clamps widen to the
+    // current distance, so a pivot outside [min, max] cannot make the next
+    // zoom snap-jump either.
+    distance = depth
+    target = [e[0] + fx * depth, e[1] + fy * depth, e[2] + fz * depth]
+  }
+  let rotate = (dAzimuth: number, dElevation: number) => {
+    azimuth += dAzimuth
+    elevation = clampNum(elevation + dElevation, minElevation(), maxElevation())
   }
-  let anchorAt = (x?: number, y?: number) =>
-    x !== undefined && y !== undefined && options.zoomAnchor
-      ? options.zoomAnchor(x, y, { eye: eye(), target: [target[0], target[1], target[2]] })
-      : null
 
-  // The pinch keeps ONE anchor for its whole gesture, taken at the first
-  // span change. Per-event re-derivation looks equivalent but jitters in
-  // practice: the two fingers' events interleave, so the measured span
-  // oscillates around its true value even while the fingers rest, and every
-  // event zooms slightly in or back out. About a fixed point those pairs
-  // cancel exactly (the ratios telescope); about a fresh anchor each time -
-  // a new raycast against a pose the previous event just moved, or the
-  // app's hit/fallback choice flipping between events - each pair leaves a
-  // residual target slide and the model visibly crawls under resting
-  // fingers. The wheel stays per-event: its notches are discrete, there is
-  // no noise to cancel, and each notch re-aiming at the cursor is the point.
+  // A pinch holds one anchor for its whole gesture (see the header); the
+  // wheel, unbracketed, anchors per notch.
   let pinchAnchor: Vec3 | null = null
   let pinchSeen = false
+  let anchorAt = (focal: Vec2 | undefined): Vec3 | null => (focal && options.zoomAnchor ? options.zoomAnchor(focal, view()) : null)
 
-  clampPose()
-  apply()
-
-  // One finger rotates. With two down, the span change zooms and the focal
-  // translation pans (or, with no viewport to map pixels through, keeps
-  // rotating). Pair rotation has no orbit meaning and is ignored.
-  let transform = createTransform({
-    onTransformStart: () => {
-      interacting = true
-      let anchor = options.rotateAnchor?.({ eye: eye(), target: [target[0], target[1], target[2]] })
-      if (anchor) {
-        let e = eye()
-        let fx = (target[0] - e[0]) / distance
-        let fy = (target[1] - e[1]) / distance
-        let fz = (target[2] - e[2]) / distance
-        let depth = (anchor[0] - e[0]) * fx + (anchor[1] - e[1]) * fy + (anchor[2] - e[2]) * fz
-        if (depth > 0) {
-          // Deliberately unclamped: the pose is unchanged, this only decides
-          // what the gesture pivots about. The zoom clamps below widen to
-          // the current distance, so a pivot outside [min, max] cannot make
-          // the next zoom snap-jump either.
-          distance = depth
-          target = [e[0] + fx * depth, e[1] + fy * depth, e[2] + fz * depth]
+  let axes = createAxes<OrbitAxes>(
+    { rotate: "vec2", zoom: "axis", pan: "vec2" },
+    {
+      onBegin: name => {
+        if (name === "rotate") {
+          let anchor = options.rotateAnchor?.(view())
+          if (anchor) setPivot(anchor)
         }
-      }
-    },
-    onTransformMove: (t) => {
-      let vp = options.viewport?.() ?? null
-      if (t.pointers >= 2 && vp !== null) {
-        pan(t.dx, t.dy, vp)
-      } else {
-        // Viewport-relative when the height is known, per-pixel otherwise.
-        let rel = vp !== null ? (DRAG_TURNS * 2 * Math.PI) / vp.height : null
-        azimuth -= t.dx * (rel ?? DRAG_AZIMUTH) * rotateSpeed()
-        elevation = clampNum(elevation + t.dy * (rel ?? DRAG_ELEVATION) * rotateSpeed(), minElevation(), maxElevation())
-      }
-      if (t.scale !== 1) {
-        // Fingers spreading (scale > 1) zooms in: the distance shrinks by the
-        // span ratio, exponent-weighted by the zoomSpeed multiplier, about
-        // the gesture's anchor when the app provides one.
-        if (!pinchSeen) {
-          pinchSeen = true
-          pinchAnchor = anchorAt(t.x, t.y)
+        if (name === "zoom") {
+          pinchSeen = false
+          pinchAnchor = null
         }
-        zoomAbout(1 / Math.pow(t.scale, zoomSpeed()), pinchAnchor)
-      }
-      dirty = true
-    },
-    onTransformEnd: () => {
-      interacting = false
-      pinchSeen = false
-      pinchAnchor = null
+      },
+      onNudge: (name, delta, focal) => {
+        if (name === "rotate") {
+          let d = delta as Vec2
+          let rel = DRAG_TURNS * 2 * Math.PI * rotateSpeed()
+          rotate(-d[0] * rel, d[1] * rel)
+        } else if (name === "zoom") {
+          let anchor: Vec3 | null
+          if (axes.inGesture("zoom")) {
+            if (!pinchSeen) {
+              pinchSeen = true
+              pinchAnchor = anchorAt(focal)
+            }
+            anchor = pinchAnchor
+          } else {
+            anchor = anchorAt(focal)
+          }
+          zoomAbout(Math.pow(2, -(delta as number) * zoomSpeed()), anchor)
+        } else {
+          let d = delta as Vec2
+          let w = worldPerHeight() * panSpeed()
+          slide(-d[0] * w, d[1] * w)
+        }
+        push()
+      },
     },
-  })
+  )
+  let interacting = () => axes.inGesture("rotate") || axes.inGesture("zoom") || axes.inGesture("pan")
+  let active = createMemo(() => (orbiting() && orbitSpeed() !== 0) || axes.active())
+
+  clampPose()
+  push()
+  changed = false
 
   return {
     eye,
-    pose: () => ({ azimuth, elevation, distance }),
+    pose: () => ({ azimuth, elevation, distance, target: [target[0], target[1], target[2]] }),
     orbiting,
-    active: () => orbiting() && orbitSpeed() !== 0,
+    active,
+    axes,
     set(pose) {
       if (pose.azimuth !== undefined) azimuth = pose.azimuth
       if (pose.elevation !== undefined) elevation = pose.elevation
       if (pose.distance !== undefined) distance = pose.distance
       if (pose.target) target = [pose.target[0], pose.target[1], pose.target[2]]
       clampPose()
-      if (pose.orbiting !== undefined) setOrbiting(pose.orbiting)
-      dirty = true
+      if (pose.orbiting !== undefined) {
+        running = pose.orbiting
+        setOrbiting(running)
+      }
+      push()
     },
     update(dt) {
+      finite("update dt", dt)
+      let moved = false
       let rate = orbitSpeed()
-      if (rate !== 0 && !interacting && orbiting()) {
+      if (rate !== 0 && !interacting() && running) {
         azimuth += dt * rate
-        dirty = true
+        moved = true
+      }
+      let [rx, ry] = untrack(() => axes.rate("rotate"))
+      if (rx !== 0 || ry !== 0) {
+        let step = ROTATE_RATE * 2 * Math.PI * rotateSpeed() * dt
+        rotate(-rx * step, ry * step)
+        moved = true
+      }
+      let z = untrack(() => axes.rate("zoom"))
+      if (z !== 0) {
+        zoomAbout(Math.pow(2, -z * ZOOM_RATE * zoomSpeed() * dt), null)
+        moved = true
       }
-      if (!dirty) return false
-      dirty = false
-      apply()
-      return true
+      let [px, py] = untrack(() => axes.rate("pan"))
+      if (px !== 0 || py !== 0) {
+        let w = worldPerHeight() * PAN_RATE * panSpeed() * dt
+        slide(-px * w, py * w)
+        moved = true
+      }
+      if (moved) push()
+      let result = changed
+      changed = false
+      return result
     },
-    handlers: {
-      ...transform.handlers,
-      onWheel(e) {
-        zoomAbout(Math.exp(e.deltaY * WHEEL_ZOOM * zoomSpeed()), anchorAt(e.localX, e.localY))
-        dirty = true
-      },
+    rotateBy(dAzimuth, dElevation) {
+      finite("rotateBy azimuth", dAzimuth)
+      finite("rotateBy elevation", dElevation)
+      rotate(dAzimuth, dElevation)
+      push()
+    },
+    zoomBy(factor, anchor) {
+      if (!(Number.isFinite(factor) && factor > 0)) throw new Error(`createOrbitCamera: zoomBy factor must be positive, got ${factor}`)
+      zoomAbout(1 / factor, anchor)
+      push()
+    },
+    panBy(right, up) {
+      finite("panBy right", right)
+      finite("panBy up", up)
+      slide(right, up)
+      push()
+    },
+    setPivot(point) {
+      setPivot(point)
+      push()
     },
   }
 }