feat(notch): trackpad gestures, Space-swipe collapse, and motion/appearance controls

[Mrjamedd]

Adds trackpad gesture control to the island, collapses it when the user swipes to another desktop, and adds two appearance/motion settings. Rebased on main at v1.0.32; swift test is green (957 tests, 0 failures).

Gestures on the island

Scroll gestures over the island, interpreted by a small pure state machine (NotchGestureInterpreter) with a 24pt activation threshold and a 1.2× axis-dominance rule so a diagonal drift doesn't fire the wrong action:

• swipe down — open
• swipe up — close
• swipe horizontally — move through the session filters (all / status / cli)

Horizontal direction is invertible via a setting. Gesture delivery uses both a local and a global monitor, because a nonactivating panel doesn't receive scroll events when macOS routes them to the app underneath the physical notch.

Collapse when switching desktops

An expanded island used to ride through the entire Spaces animation and snap shut at the end, because NSWorkspace.activeSpaceDidChangeNotification only fires once macOS has finished the transition.

Detecting the swipe as it happens turned out to need a mechanism worth spelling out, since two obvious candidates don't work:

• NSEvent global monitors deliver scroll events to a background app but strip the NSTouch payload — measured 156 scroll events with zero touches and no gesture events at all.
• NSEvent.trackSwipeEvent only operates on .scrollWheel events with precise deltas, i.e. the two-finger page-swipe gesture. A three/four-finger Spaces swipe never arrives as a scroll event, so there is no gestureAmount to read.
• The responder-chain touchesMoved(with:) path only fires while the panel is the key window — never the case when the user is working in another app.

This uses MultitouchSupport, the private framework, which does deliver normalized finger positions to a background app with no permission prompt. Flagging that explicitly for review, since it's a private-API dependency:

• resolved through dlopen/dlsym rather than linked
• MTTouch's 96-byte stride is validated before any field is read
• if either check fails, isAvailable goes false and the existing activeSpaceDidChange observer remains as the fallback — no crash, just the old late-collapse behaviour
• devices are re-armed on wake, since MultitouchSupport stops delivering frames across sleep

Scoping keeps it from firing on unrelated gestures:

• armed only while a collapsible surface is actually on screen, so an idle island does no per-frame work
• the threshold crossing collapses only when the pointer is over the island's rendered content — reusing the same NotchGestureHitbox and visible-content geometry the in-view gesture monitor already computes. Without this, a global gesture would collapse the island on every desktop switch anywhere on screen.
• required finger count is read from the system trackpad preference (TrackpadThreeFingerHorizSwipeGesture / TrackpadFourFingerHorizSwipeGesture), so it matches whichever gesture the user actually has bound, and stays silent if they've turned it off

Verified against real hardware: 2006 contact frames, 4 simultaneous fingers, 4/4 swipes detected, one crossing per gesture.

Contrast edge tinted by mascot

Opt-in, off by default, so the white edge stays the shipped look and turning it off restores it exactly. Each mascot's signature color is lifted 62% toward white before it reaches the stroke, keeping a hairline highlight that carries a hue rather than a colored outline. Geometry and opacities are untouched — toggling changes hue and nothing else. The edge follows whichever mascot the bar is showing and crossfades over 0.35s when that changes.

Signature colors moved out of a private table in MascotsPage into MascotPalette, so the settings swatches and the edge tint can't drift, keyed to match MascotView's routing including aliases (cursor-cli, qoderwork, google-antigravity, omp).

Open / close speed

A 0.5×–2.0× slider scaling the spring response for the island's open and close animations. It scales the response rather than swapping in a fixed duration, so the motion stays interruptible at every setting, and responses are read live so the slider applies without a relaunch. It sits outside the hover-to-open branch deliberately: it governs the close animation too, which runs regardless of what opened the island.

Default change

openOnHover now ships off. Brushing the notch on the way to the menu bar shouldn't expand the island; click and swipe remain the deliberate ways in. showContrastEdge stays on. Flagging this as the one behaviour change for existing users who never touched the setting.

Tests

swift test — 957 tests, 0 failures. New coverage for the gesture interpreter, action policy, hitbox, Spaces-swipe detector (direction, single-fire-per-gesture, jitter rejection, finger-count filtering, rebaselining after a lift), mascot palette (fallbacks, aliases, blend bounds, plus guards that every mascot stays pale enough to read as a highlight and saturated ones keep a distinguishable hue), and the animation-speed math.

One drive-by fix: 810711a flipped the horizontal direction mapping and updated one of its two tests. testGestureEmitsOnlyOnceUntilEnded kept the pre-flip expectation, so the suite had two tests asserting opposite results for identical input. Corrected in d0d0d7a.

Notes for review

• The private-framework dependency is the main thing to weigh. It's isolated to MultitouchDevice.swift behind an availability check, and the feature degrades to the previous behaviour if it ever stops resolving — but it is private API, and it would block Mac App Store distribution if that's ever a goal.
• MascotPalette needed a grok entry after rebasing onto the new Grok mascot; it uses white to match GrokView. Worth a glance that the color is right.

[wxtsky]

Ran the full swift test suite against this branch for you, since your notes say your Mac can't resolve XCTest. Rebased onto current main first (see conflict note below), then built and ran everything.

Result: 919 tests, 2 skipped, 2 failures — both in this PR's own test file, and they are a genuine contradiction rather than an environment problem.

NotchGestureInterpreterTests.swift:22: testGestureEmitsOnlyOnceUntilEnded :
  XCTAssertEqual failed: ("Optional(NotchGestureAction.navigatePrevious)") is not equal to ("Optional(NotchGestureAction.navigateNext)")
NotchGestureInterpreterTests.swift:25: (same)

Two of your tests disagree about the same input:

// testPhysicalDirectionsMapToNaturalNotchActions — matches the implementation
XCTAssertEqual(action(x: -30), .navigatePrevious)

// testGestureEmitsOnlyOnceUntilEnded — expects the opposite for the same delta
XCTAssertEqual(interpreter.consume(sample(x: -30)), .navigateNext)

Both helpers feed physicalDeltaX, so the inputs really are identical. The implementation returns .navigatePrevious, which agrees with the first test and with your PR description ("physical swipe left navigates right; physical swipe right navigates left"), so the second test looks like a stale expectation from before the direction was settled. Suggested fix — it also keeps the test's actual subject (emit-once-per-gesture) intact:

-        XCTAssertEqual(interpreter.consume(sample(x: -30)), .navigateNext)
+        XCTAssertEqual(interpreter.consume(sample(x: -30)), .navigatePrevious)
         XCTAssertNil(interpreter.consume(sample(x: -30)))
         XCTAssertNil(interpreter.consume(sample(ended: true)))
-        XCTAssertEqual(interpreter.consume(sample(x: -30, began: true)), .navigateNext)
+        XCTAssertEqual(interpreter.consume(sample(x: -30, began: true)), .navigatePrevious)

Everything else passes, including all of NotchGestureHitboxTests and the settings/localization/hover coverage.

Rebase note: main moved a fair bit today. One conflict, in PanelWindowController.swift: sessionObservationTask is gone — the 500 ms observation poll was replaced by isSessionObservationArmed + re-arming from onChange (#299, it was a permanent wakeup for no benefit). Your activeSpaceTransitionTask sits alongside it unchanged; resolving is just keeping both declarations. Nothing else in your branch conflicted, and it builds clean after that.

No rush — leaving it as a draft since it's yours to finish. The gesture work reads well; the hitbox tests in particular are nice.