Summary

Replaces the stateful, burst-based device classifier in the camera controller's wheel handler with a modifier-key-only model, mirroring the approach already shipped in supersplat-viewer. The previous classifier tried to distinguish a physical mouse wheel from a trackpad two-finger swipe using per-event heuristics (wheelDelta % 120, deltaMode, fractional/diagonal deltas), but hi-res mice emit smooth deltas that are indistinguishable from a trackpad. On the reporter's Logitech G502 X (Windows 11), a bare wheel was misclassified as a trackpad ~95% of the time and fell through to the orbit branch instead of zooming. The same misfire was reported on Windows Edge.

Root cause

A bare wheel event and a bare trackpad swipe carry no reliable signal that tells them apart — every delta-based heuristic misfires on some hardware or in momentum tails. The old code's else fallback was orbit(deltaX, deltaY), so any wheel that failed the "is this a mouse?" check orbited.

Approach

The one trackpad signal the platform reliably provides is the macOS/Magic-Mouse pinch: a wheel event with ctrlKey set while the physical Ctrl key is not held (synthetic Ctrl). So we stop classifying the device entirely and drive everything from modifier keys, treating a bare scroll as zoom regardless of input device.

To tell a synthetic-Ctrl pinch from a user physically holding Ctrl while scrolling, we track the physical Ctrl key with keydown/keyup listeners on window (cleared on blur and on destroy). These are registered directly rather than through the wrap helper, so Ctrl tracking does not fire camera.controller on every keystroke.

Input mapping: old vs new

Orbit mode

Fly mode

The Shift-remap workaround (reading deltaX when macOS maps a Shift+wheel to deltaY === 0) is preserved.

Trade-off

A bare macOS trackpad two-finger swipe now zooms instead of orbiting; trackpad orbit moves to Ctrl + swipe. This is unavoidable — a bare wheel and a bare swipe are physically indistinguishable, so fixing mouse-wheel zoom necessarily changes bare-swipe behavior. This matches the resolution already adopted in supersplat-viewer and the tension noted by the maintainer in the issue thread.

Test plan

Verified at runtime by dispatching synthetic wheel events against the live editor canvas and reading camera state before/after:

• Bare wheel → distance changes, azim/elevation unchanged (zoom)
• Synthetic-Ctrl pinch (ctrlKey, no physical Ctrl) → zoom
• Physical Ctrl held + scroll → azim/elevation change (orbit)
• Shift + scroll → focal point moves (pan)
• Ctrl keyup resets tracking → a subsequent ctrlKey wheel is treated as pinch/zoom again
• tsc --noEmit and eslint clean, no console errors