feat(scroll): per-device inverted scrolling

[AprilNEA]

Summary

Adds an opt-in, per-device "Invert scroll direction" toggle. A user can keep
macOS natural scrolling on the trackpad while running a traditional (reversed)
wheel on the mouse — which the native setting can't do.

Closes #126.

Why the native option isn't enough

macOS scroll direction is a single global default (com.apple.swipescrolldirection);
the Mouse and Trackpad panes both reflect that same value. So "natural trackpad

• reversed mouse" is impossible natively — exactly the case [Feature]: Inverted Scrolling Support #126 describes, and the
  reason tools like Scroll Reverser / LinearMouse exist. The toggle here inverts
  relative to whatever the OS produces, so it composes correctly with the global
  setting.

Design — layered so each crate keeps its own job

• openlogi-core (data model): a per-device invert_scroll flag on DeviceConfig
  with accessors; omitted from config.toml when false.
• openlogi-hook (pure OS mechanism, no policy): MouseEvent::Scroll gains
  is_continuous (a trackpad/Magic-Mouse gesture vs. a discrete wheel detent), and
  a new EventDisposition::ReplaceScroll that re-emits the wheel with transformed
  deltas. Each platform's hook owns its own re-emission:
  • macOS mutates the CGEvent in place and scales its smooth-scroll companion
    fields (fixed-point + pixel deltas), so pixel-precise apps invert too — not just
    line-based ones.
  • Linux re-injects the flipped tick via uinput.
  • Windows re-injects via SendInput; the hook already drops LLMHF_INJECTED
    events, so the synthetic tick can't loop back.
• openlogi-agent-core (policy): the orchestrator publishes the active device's
  flag into a shared AtomicBool; the hook runtime inverts only a discrete wheel
  and passes continuous trackpad scrolling through untouched.
• openlogi-gui: a "Scrolling" card in the Pointer tab with an On/Off toggle
  (pure config — no hardware read). New strings translated across all 20 locales.

Known limitation (documented in code)

The macOS CGEventTap reads the merged HID stream and can't attribute a scroll to a
physical device, so the flag applies to the selected device — the same
active-device compromise as thumbwheel_sensitivity. The trackpad-vs-wheel split is
done via is_continuous, which is what makes #126's actual scenario work. On Linux
(thread-per-device evdev) true per-device is structurally possible later.

Verification

• cargo clippy --workspace --all-targets -- -D warnings — clean (macOS)
• cargo clippy --target x86_64-pc-windows-gnu -p openlogi-hook -p openlogi-inject -p openlogi-agent-core --all-targets -- -D warnings — clean (Windows code cross-linted)
• cargo fmt --all --check — clean
• Tests pass: openlogi-core, openlogi-hook, openlogi-agent-core (incl. wire_format
  — no IPC protocol bump, the flag travels via config + ReloadConfig), and
  openlogi-gui (incl. the i18n ordered-key parity test)

[greptile-apps]

Greptile Summary

This PR adds a per-device "Invert scroll direction" toggle that lets users reverse a mouse's discrete scroll wheel while leaving trackpad natural-scrolling untouched — working around the macOS single-global-setting limitation described in issue #126. The implementation is layered cleanly across four crates: config model, OS hooks, orchestrator, and GUI.

• openlogi-core: DeviceConfig.invert_scroll bool, skipped in TOML when false; round-trip and omission tests cover it.
• openlogi-hook: MouseEvent::Scroll gains is_continuous; new EventDisposition::ReplaceScroll drives platform-specific re-emission — macOS mutates all three CGEvent delta fields in-place, Linux re-injects via uinput, Windows drops the original and re-injects via SendInput (LLMHF_INJECTED prevents loop-back).
• openlogi-agent-core/gui: Arc<AtomicBool> shared between orchestrator and hook runtime carries the active-device flag; the GUI adds a "Scrolling" card with an On/Off pill, and all 20 locale files are updated.

Confidence Score: 5/5

Safe to merge — all three platform scroll-inversion paths are implemented and guarded correctly, with no loop-back or data-loss risk.

The Windows injection path (previously unimplemented) is now complete with LLMHF_INJECTED loop-back prevention. macOS mutates all three CGEvent delta fields so pixel-precise apps invert correctly. Linux reinjects through uinput with proper hi-res scaling. Config round-trips cleanly and the AtomicBool handoff from orchestrator to hook runtime follows the established thumbwheel_sensitivity pattern. No correctness issues were found across the changed paths.

No files require special attention.

Important Files Changed

Filename	Overview
crates/openlogi-hook/src/lib.rs	Adds is_continuous to MouseEvent::Scroll and ReplaceScroll variant to EventDisposition; Eq is no longer derived because f32 fields are not Eq — this is a breaking change to the public API surface noted in a previous review comment.
crates/openlogi-hook/src/macos.rs	Adds rewrite_scroll/rewrite_scroll_axis that mutates all three CGEvent delta fields (line, fixed-point, pixel) in place; correctly scales by new_line/old_line and skips the division when old_line == 0.
crates/openlogi-hook/src/windows.rs	Adds inject_scroll via SendInput for ReplaceScroll; correctly suppresses original event and lets LLMHF_INJECTED filter prevent loop-back. The i32→u32 bit-cast for mouseData is idiomatic for Win32 signed-delta-in-DWORD fields.
crates/openlogi-hook/src/linux.rs	reinject_scroll correctly inverts the inverse of translate for all four axis codes (REL_WHEEL, REL_WHEEL_HI_RES, REL_HWHEEL, REL_HWHEEL_HI_RES), re-scaling hi-res values by HIRES_UNITS_PER_TICK as the forward direction does.
crates/openlogi-agent-core/src/hook_runtime.rs	Adds invert_scroll AtomicBool parameter; correctly gates inversion on !is_continuous && delta_y != 0.0, matching the discrete-wheel-only, vertical-only spec.
crates/openlogi-agent-core/src/orchestrator.rs	Seeds invert_scroll AtomicBool in SharedRuntime and refreshes it on rebuild; uses Relaxed ordering, consistent with the existing thumbwheel_sensitivity pattern.
crates/openlogi-core/src/config.rs	Adds invert_scroll bool to DeviceConfig with skip_serializing_if = is_false; round-trip and omission tests verify correct TOML behaviour.
crates/openlogi-gui/src/app.rs	Adds scrolling_card with invert_scroll_toggle to the Pointer tab; reads state via AppState::current_invert_scroll and commits via commit_invert_scroll, following the existing immediate-mode pattern.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[OS delivers scroll event] --> B{is_continuous?}
    B -- yes / trackpad --> C[PassThrough — no change]
    B -- no / mouse wheel --> D{invert_scroll AtomicBool?}
    D -- false --> C
    D -- true --> E{delta_y != 0?}
    E -- no / horizontal only --> C
    E -- yes --> F[EventDisposition::ReplaceScroll\ndelta_x unchanged, delta_y negated]
    F --> G{Platform}
    G -- macOS --> H[rewrite_scroll: mutate CGEvent\nline + fixed-point + pixel fields\nCallbackResult::Keep]
    G -- Linux --> I[reinject_scroll: replace evdev\nevent value via uinput]
    G -- Windows --> J[inject_scroll via SendInput\noriginal suppressed\nLLMHF_INJECTED prevents loop-back]
    H --> K[Inverted event reaches app]
    I --> K
    J --> K

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
flowchart TD
    A[OS delivers scroll event] --> B{is_continuous?}
    B -- yes / trackpad --> C[PassThrough — no change]
    B -- no / mouse wheel --> D{invert_scroll AtomicBool?}
    D -- false --> C
    D -- true --> E{delta_y != 0?}
    E -- no / horizontal only --> C
    E -- yes --> F[EventDisposition::ReplaceScroll\ndelta_x unchanged, delta_y negated]
    F --> G{Platform}
    G -- macOS --> H[rewrite_scroll: mutate CGEvent\nline + fixed-point + pixel fields\nCallbackResult::Keep]
    G -- Linux --> I[reinject_scroll: replace evdev\nevent value via uinput]
    G -- Windows --> J[inject_scroll via SendInput\noriginal suppressed\nLLMHF_INJECTED prevents loop-back]
    H --> K[Inverted event reaches app]
    I --> K
    J --> K

Loading

_{Reviews (2): Last reviewed commit: "feat(scroll): implement Windows wheel in..." | Re-trigger Greptile}

[AprilNEA]

Addressed the P1 in fdd916e — Windows now actually inverts instead of silently no-op'ing.

translate_event already drops LLMHF_INJECTED events, so re-injecting the flipped tick via SendInput can't loop back through the WH_MOUSE_LL hook. The dispatch now suppresses the original wheel and re-injects a transformed one, making Windows symmetric with Linux (uinput) and macOS (in-place CGEvent mutation). Verified with an x86_64-pc-windows-gnu cross-lint (-D warnings) plus a unit test on the mouseData reconstruction (one notch → 120, inverted → −120, exact negation).